MACHINE

LEARNING
WITH PYTHON
Theory and Applications
 MACHINE
LEARNING
WITH PYTHON
Theory and Applications

G. R. Liu
University of Cincinnati, USA

World Scientific
NEW JERSEY • LONDON • SINGAPORE • BEIJING • SHANGHAI • HONG KONG • TAIPEI • CHENNAI • TOKYO
Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE

Library of Congress Cataloging-in-Publication Data

Names: Liu, G. R. (Gui-Rong), author.
Title: Machine learning with Python : theory and applications / G.R. Liu, University of Cincinnati, USA.
Description: Singapore ; Hackensack, NJ : World Scientific Publishing Co. Pte. Ltd., [2023] |
Includes bibliographical references and index.
Identifiers: LCCN 2022001048 | ISBN 9789811254178 (hardcover) |
ISBN 9789811254185 (ebook for institutions) | ISBN 9789811254192 (ebook for individuals)
Subjects: LCSH: Machine learning. | Python (Computer program language)
Classification: LCC Q325.5 .L58 2023 | DDC 006.3/1--dc23/eng20220328
LC record available at https://lccn.loc.gov/2022001048

British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library.

Copyright © 2023 by World Scientific Publishing Co. Pte. Ltd.

All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or
mechanical, including photocopying, recording or any information storage and retrieval system now known or to
be invented, without written permission from the publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center,
Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from
the publisher.

For any available supplementary material, please visit

https://www.worldscientific.com/worldscibooks/10.1142/12774#t=suppl

Desk Editors: Jayanthi Muthuswamy/Steven Patt

Typeset by Stallion Press

Email: enquiries@stallionpress.com

Printed in Singapore
 About the Author

G. R. Liu received his Ph.D from Tohoku University,

Japan, in 1991. He was a post-doctoral fellow at
Northwestern University, USA, from 1991–1993. He
was a Professor at the National University of Singa-
pore until 2010. He is currently a Professor at the
University of Cincinnati, USA. He is the founder of the
Association for Computational Mechanics (Singapore)
(SACM) and served as the President of SACM until
2010. He served as the President of the Asia-Pacific
Association for Computational Mechanics (APACM)
(2010–2013) and an Executive Council Member of the International Asso-
ciation for Computational Mechanics (IACM) (2005–2010; 2020–2026). He
has authored a large number of journal papers and books including two
bestsellers — Mesh Free Method: Moving Beyond the Finite Element Method
and Smoothed Particle Hydrodynamics: A Meshfree Particle Methods. He is
the Editor-in-Chief of the International Journal of Computational Methods
and served as an Associate Editor for IPSE and MANO. He is the recipient
of numerous awards, including the Singapore Defence Technology Prize,
NUS Outstanding University Researcher Award and Best Teacher Award,
APACM Computational Mechanics Awards, JSME Computational Mechan-
ics Awards, ASME Ted Belytschko Applied Mechanics Award, Zienkiewicz
Medal from APACM, the AJCM Computational Mechanics Award, and the
Humboldt Research Award. He has been listed as a world top 1% most
influential scientist (Highly Cited Researchers) by Thomson Reuters in 2014–
2016, 2018, and 2019. ISI citations by others: ∼22000; His journal and
book credentials include the following. ISI H-index: ∼85; Google Scholar
H-Index: 110.

v
MACHINE LEARNING
WITH PYTHON
 Contents

About the Author v

1 Introduction 1
1.1 Naturally Learned Ability for Problem Solving . . . . . . . 1
1.2 Physics-Law-based Models . . . . . . . . . . . . . . . . . . 1
1.3 Machine Learning Models, Data-based . . . . . . . . . . . 3
1.4 General Steps for Training Machine Learning Models . . . 4
1.5 Some Mathematical Concepts, Variables, and Spaces . . . 5
1.5.1 Toy examples . . . . . . . . . . . . . . . . . . . . . 5
1.5.2 Feature space . . . . . . . . . . . . . . . . . . . . . 6
1.5.3 Affine space . . . . . . . . . . . . . . . . . . . . . 7
1.5.4 Label space . . . . . . . . . . . . . . . . . . . . . . 8
1.5.5 Hypothesis space . . . . . . . . . . . . . . . . . . . 9
1.5.6 Definition of a typical machine learning model,
a mathematical view . . . . . . . . . . . . . . . . . 10
1.6 Requirements for Creating Machine Learning Models . . . 11
1.7 Types of Data . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.8 Relation Between Physics-Law-based and Data-based
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.9 This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.10 Who May Read This Book . . . . . . . . . . . . . . . . . . 14
1.11 Codes Used in This Book . . . . . . . . . . . . . . . . . . . 14
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

vii
viii Machine Learning with Python: Theory and Applications

2 Basics of Python 19
2.1 An Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Briefing on Python . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Variable Types . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.1 Numbers . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.2 Underscore placeholder . . . . . . . . . . . . . . . 28
2.3.3 Strings . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.4 Conversion between types of variables . . . . . . . 36
2.3.5 Variable formatting . . . . . . . . . . . . . . . . . 38
2.4 Arithmetic Operators . . . . . . . . . . . . . . . . . . . . . 39
2.4.1 Addition, subtraction, multiplication, division,
and power . . . . . . . . . . . . . . . . . . . . . . 39
2.4.2 Built-in functions . . . . . . . . . . . . . . . . . . 40
2.5 Boolean Values and Operators . . . . . . . . . . . . . . . . 41
2.6 Lists: A diversified variable type container . . . . . . . . . 42
2.6.1 List creation, appending, concatenation,
and updating . . . . . . . . . . . . . . . . . . . . . 42
2.6.2 Element-wise addition of lists . . . . . . . . . . . . 44
2.6.3 Slicing strings and lists . . . . . . . . . . . . . . . 46
2.6.4 Underscore placeholders for lists . . . . . . . . . . 49
2.6.5 Nested list (lists in lists in lists) . . . . . . . . . . 49
2.7 Tuples: Value preserved . . . . . . . . . . . . . . . . . . . . 50
2.8 Dictionaries: Indexable via keys . . . . . . . . . . . . . . . 51
2.8.1 Assigning data to a dictionary . . . . . . . . . . . 51
2.8.2 Iterating over a dictionary . . . . . . . . . . . . . 52
2.8.3 Removing a value . . . . . . . . . . . . . . . . . . 53
2.8.4 Merging two dictionaries . . . . . . . . . . . . . . 54
2.9 Numpy Arrays: Handy for scientific computation . . . . . . 55
2.9.1 Lists vs. Numpy arrays . . . . . . . . . . . . . . . 55
2.9.2 Structure of a numpy array . . . . . . . . . . . . . 55
2.9.3 Axis of a numpy array . . . . . . . . . . . . . . . . 60
2.9.4 Element-wise computations . . . . . . . . . . . . . 61
2.9.5 Handy ways to generate multi-dimensional
arrays . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.9.6 Use of external package: MXNet . . . . . . . . . . 63
2.9.7 In-place operations . . . . . . . . . . . . . . . . . 66
2.9.8 Slicing from a multi-dimensional array . . . . . . . 67
2.9.9 Broadcasting . . . . . . . . . . . . . . . . . . . . . 67
 Contents ix

2.9.10 Converting between MXNet NDArray

and NumPy . . . . . . . . . . . . . . . . . . . . . 70
2.9.11 Subsetting in Numpy . . . . . . . . . . . . . . . . 71
2.9.12 Numpy and universal functions (ufunc) . . . . . . 71
2.9.13 Numpy array and vector/matrix . . . . . . . . . . 72
2.10 Sets: No Duplication . . . . . . . . . . . . . . . . . . . . . 75
2.10.1 Intersection of two sets . . . . . . . . . . . . . . . 75
2.10.2 Difference of two sets . . . . . . . . . . . . . . . . 75
2.11 List Comprehensions . . . . . . . . . . . . . . . . . . . . . 76
2.12 Conditions, “if” Statements, “for” and “while” Loops . . . 77
2.12.1 Comparison operators . . . . . . . . . . . . . . . . 77
2.12.2 The “in” operator . . . . . . . . . . . . . . . . . . 78
2.12.3 The “is” operator . . . . . . . . . . . . . . . . . . 78
2.12.4 The ‘not’ operator . . . . . . . . . . . . . . . . . . 80
2.12.5 The “if” statements . . . . . . . . . . . . . . . . . 80
2.12.6 The “for” loops . . . . . . . . . . . . . . . . . . . 81
2.12.7 The “while” loops . . . . . . . . . . . . . . . . . . 82
2.12.8 Ternary conditionals . . . . . . . . . . . . . . . . . 84
2.13 Functions (Methods) . . . . . . . . . . . . . . . . . . . . . 84
2.13.1 Block structure for function definition . . . . . . . 84
2.13.2 Function with arguments . . . . . . . . . . . . . . 84
2.13.3 Lambda functions (Anonymous functions) . . . . 86
2.14 Classes and Objects . . . . . . . . . . . . . . . . . . . . . . 86
2.14.1 A simplest class . . . . . . . . . . . . . . . . . . . 86
2.14.2 A class for scientific computation . . . . . . . . . 89
2.14.3 Subclass (class inheritance) . . . . . . . . . . . . . 90
2.15 Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
2.16 Generation of Plots . . . . . . . . . . . . . . . . . . . . . . 92
2.17 Code Performance Assessment . . . . . . . . . . . . . . . . 93
2.18 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

3 Basic Mathematical Computations 95

3.1 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.1.1 Scalar numbers . . . . . . . . . . . . . . . . . . . . 96
3.1.2 Vectors . . . . . . . . . . . . . . . . . . . . . . . . 96
3.1.3 Matrices . . . . . . . . . . . . . . . . . . . . . . . 98
3.1.4 Tensors . . . . . . . . . . . . . . . . . . . . . . . . 100
x Machine Learning with Python: Theory and Applications

3.1.5 Sum and mean of a tensor . . . . . . . . . . . . . 101

3.1.6 Dot-product of two vectors . . . . . . . . . . . . . 102
3.1.7 Outer product of two vectors . . . . . . . . . . . . 105
3.1.8 Matrix-vector product . . . . . . . . . . . . . . . . 106
3.1.9 Matrix-matrix multiplication . . . . . . . . . . . . 106
3.1.10 Norms . . . . . . . . . . . . . . . . . . . . . . . . 108
3.1.11 Solving algebraic system equations . . . . . . . . . 109
3.1.12 Matrix inversion . . . . . . . . . . . . . . . . . . . 111
3.1.13 Eigenvalue decomposition of a matrix . . . . . . . 113
3.1.14 Condition number of a matrix . . . . . . . . . . . 116
3.1.15 Rank of a matrix . . . . . . . . . . . . . . . . . . 118
3.2 Rotation Matrix . . . . . . . . . . . . . . . . . . . . . . . . 119
3.3 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.3.1 1-D piecewise linear interpolation using
numpy.interp . . . . . . . . . . . . . . . . . . . . . 121
3.3.2 1-D least-squares solution approximation . . . . . 122
3.3.3 1-D interpolation using interp1d . . . . . . . . . . 124
3.3.4 2-D spline representation
using bisplrep . . . . . . . . . . . . . . . . . . . . 124
3.3.5 Radial basis functions for smoothing and
interpolation . . . . . . . . . . . . . . . . . . . . . 126
3.4 Singular Value Decomposition . . . . . . . . . . . . . . . . 129
3.4.1 SVD formulation . . . . . . . . . . . . . . . . . . . 129
3.4.2 Algorithms for SVD . . . . . . . . . . . . . . . . . 130
3.4.3 Numerical examples . . . . . . . . . . . . . . . . . 131
3.4.4 SVD for data compression . . . . . . . . . . . . . 133
3.5 Principal Component Analysis . . . . . . . . . . . . . . . . 135
3.5.1 PCA formulation . . . . . . . . . . . . . . . . . . 135
3.5.2 Numerical examples . . . . . . . . . . . . . . . . . 137
3.6 Numerical Root Finding . . . . . . . . . . . . . . . . . . . 143
3.7 Numerical Integration . . . . . . . . . . . . . . . . . . . . . 145
3.7.1 Trapezoid rule . . . . . . . . . . . . . . . . . . . . 145
3.7.2 Gauss integration . . . . . . . . . . . . . . . . . . 147
3.8 Initial data treatment . . . . . . . . . . . . . . . . . . . . . 148
3.8.1 Min-max scaling . . . . . . . . . . . . . . . . . . . 149
3.8.2 “One-hot” encoding . . . . . . . . . . . . . . . . . 152
3.8.3 Standard scaling . . . . . . . . . . . . . . . . . . . 153
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
 Contents xi

4 Statistics and Probability-based Learning Model 157

4.1 Analysis of Probability of an Event . . . . . . . . . . . . . 158
4.1.1 Random sampling, controlled random
sampling . . . . . . . . . . . . . . . . . . . . . . . 158
4.1.2 Probability . . . . . . . . . . . . . . . . . . . . . . 160
4.2 Random Distributions . . . . . . . . . . . . . . . . . . . . . 164
4.2.1 Uniform distribution . . . . . . . . . . . . . . . . . 165
4.2.2 Normal distribution (Gaussian distribution) . . . 165
4.3 Entropy of Probability . . . . . . . . . . . . . . . . . . . . 167
4.3.1 Example 1: Probability and its entropy . . . . . . 169
4.3.2 Example 2: Variation of entropy . . . . . . . . . . 170
4.3.3 Example 3: Entropy for events with a variable
that takes different numbers of values of uniform
distribution . . . . . . . . . . . . . . . . . . . . . . 172
4.4 Cross-Entropy: Predicated and True Probability . . . . . . 173
4.4.1 Example 1: Cross-entropy of a quality
prediction . . . . . . . . . . . . . . . . . . . . . . . 174
4.4.2 Example 2: Cross-entropy of a poor
prediction . . . . . . . . . . . . . . . . . . . . . . . 175
4.5 KL-Divergence . . . . . . . . . . . . . . . . . . . . . . . . . 175
4.5.1 Example 1: KL-divergence of a distribution
of quality prediction . . . . . . . . . . . . . . . . . 176
4.5.2 Example 2: KL-divergence of a poorly
predicted distribution . . . . . . . . . . . . . . . . 176
4.6 Binary Cross-Entropy . . . . . . . . . . . . . . . . . . . . . 177
4.6.1 Example 1: Binary cross-entropy for a distribution
of quality prediction . . . . . . . . . . . . . . . . . 178
4.6.2 Example 2: Binary cross-entropy for a poorly
predicted distribution . . . . . . . . . . . . . . . . 178
4.6.3 Example 3: Binary cross-entropy for more uniform
true distribution: A quality prediction . . . . . . . 179
4.6.4 Example 4: Binary cross-entropy for more uniform
true distribution: A poor prediction . . . . . . . . 180
4.7 Bayesian Statistics . . . . . . . . . . . . . . . . . . . . . . . 180
4.8 Naive Bayes Classification: Statistics-based Learning . . . 181
4.8.1 Formulation . . . . . . . . . . . . . . . . . . . . . 181
4.8.2 Case study: Handwritten digits recognition . . . . 181
4.8.3 Algorithm for the Naive Bayes classification . . . 182
4.8.4 Testing the Naive Bayes model . . . . . . . . . . . 185
4.8.5 Discussion . . . . . . . . . . . . . . . . . . . . . . 187
xii Machine Learning with Python: Theory and Applications

5 Prediction Function and Universal Prediction Theory 189

5.1 Linear Prediction Function and Affine Transformation . . . 190
5.1.1 Linear prediction function: A basic
hypothesis . . . . . . . . . . . . . . . . . . . . . . 191
5.1.2 Predictability for constants, the role
of the bias . . . . . . . . . . . . . . . . . . . . . . 192
5.1.3 Predictability for linear functions:
The role of the weights . . . . . . . . . . . . . . . 192
5.1.4 Prediction of linear functions: A machine
learning procedure . . . . . . . . . . . . . . . . . . 193
5.1.5 Affine transformation . . . . . . . . . . . . . . . . 194
5.2 Affine Transformation Unit (ATU), A Simplest
Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
5.3 Typical Data Structures . . . . . . . . . . . . . . . . . . . 198
5.4 Demonstration Examples of Affine Transformation . . . . . 199
5.4.1 An edge, a rectangle under affine
transformation . . . . . . . . . . . . . . . . . . . . 202
5.4.2 A circle under affine transformation . . . . . . . . 204
5.4.3 A spiral under affine transformation . . . . . . . . 205
5.4.4 Fern leaf under affine transformation . . . . . . . 205
5.4.5 On linear prediction function with affine
transformation . . . . . . . . . . . . . . . . . . . . 206
5.4.6 Affine transformation wrapped with activation
function . . . . . . . . . . . . . . . . . . . . . . . . 206
5.5 Parameter Encoding and the Essential Mechanism
of Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
5.5.1 The x to ŵ encoding, a data-parameter
converter unit . . . . . . . . . . . . . . . . . . . . 210
5.5.2 Uniqueness of the encoding . . . . . . . . . . . . . 211
5.5.3 Uniqueness of the encoding: Not affected
by activation function . . . . . . . . . . . . . . . . 212
5.6 The Gradient of the Prediction Function . . . . . . . . . . 213
5.7 Affine Transformation Array (ATA) . . . . . . . . . . . . . 213
5.8 Predictability of High-Order Functions of a Deepnet . . . . 214
5.8.1 A role of activation functions . . . . . . . . . . . . 214
5.8.2 Formation of a deepnet by chaining ATA . . . . . 215
5.8.3 Example: A 1 → 1 → 1 network . . . . . . . . . . 217
5.9 Universal Prediction Theory . . . . . . . . . . . . . . . . . 218
 Contents xiii

5.10 Nonlinear Affine Transformations . . . . . . . . . . . . . . 219

5.11 Feature Functions in Physics-Law-based Models . . . . . . 220
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

6 The Perceptron and SVM 223

6.1 Linearly Separable Classification Problems . . . . . . . . . 224
6.2 A Python Code for the Perceptron . . . . . . . . . . . . . 226
6.3 The Perceptron Convergence Theorem . . . . . . . . . . . 233
6.4 Support Vector Machine . . . . . . . . . . . . . . . . . . . 237
6.4.1 Problem statement . . . . . . . . . . . . . . . . . 237
6.4.2 Formulation of objective function
and constraints . . . . . . . . . . . . . . . . . . . . 238
6.4.3 Modified objective function with constraints:
Multipliers method . . . . . . . . . . . . . . . . . 242
6.4.4 Converting to a standard quadratic programming
problem . . . . . . . . . . . . . . . . . . . . . . . . 245
6.4.5 Prediction in SVM . . . . . . . . . . . . . . . . . . 249
6.4.6 Example: A Python code for SVM . . . . . . . . . 250
6.4.7 Confusion matrix . . . . . . . . . . . . . . . . . . 254
6.4.8 Example: A Sickit-learn class for SVM . . . . . . 254
6.4.9 SVM for datasets not separable with
hyperplanes . . . . . . . . . . . . . . . . . . . . . 256
6.4.10 Kernel trick . . . . . . . . . . . . . . . . . . . . . 257
6.4.11 Example: SVM classification with curves . . . . . 258
6.4.12 Multiclass classification via SVM . . . . . . . . . . 260
6.4.13 Example: Use of SVM classifiers for
iris dataset . . . . . . . . . . . . . . . . . . . . . . 260
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

7 Activation Functions and Universal

Approximation Theory 265
7.1 Sigmoid Function (σ(z)) . . . . . . . . . . . . . . . . . . . 266
7.2 Sigmoid Function of an Affine Transformation
Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
7.3 Neural-Pulse-Unite (NPU) . . . . . . . . . . . . . . . . . . 269
7.4 Universal Approximation Theorem . . . . . . . . . . . . . 274
7.4.1 Function approximation using NPUs . . . . . . . . 274
7.4.2 Function approximations using neuron
basis functions . . . . . . . . . . . . . . . . . . . . 275
xiv Machine Learning with Python: Theory and Applications

7.4.3 Remarks . . . . . . . . . . . . . . . . . . . . . . . 281

7.5 Hyperbolic Tangent Function (tanh) . . . . . . . . . . . . . 282
7.6 Relu Functions . . . . . . . . . . . . . . . . . . . . . . . . . 283
7.7 Softplus Function . . . . . . . . . . . . . . . . . . . . . . . 286
7.8 Conditions for activation functions . . . . . . . . . . . . . . 288
7.9 Novel activation functions . . . . . . . . . . . . . . . . . . 288
7.9.1 Rational activation function . . . . . . . . . . . . 288
7.9.2 Power function . . . . . . . . . . . . . . . . . . . . 292
7.9.3 Power-linear function . . . . . . . . . . . . . . . . 294
7.9.4 Power-quadratic function . . . . . . . . . . . . . . 297
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
8 Automatic Differentiation and Autograd 303
8.1 General Issues on Optimization and Minimization . . . . . 303
8.2 Analytic Differentiation . . . . . . . . . . . . . . . . . . . . 304
8.3 Numerical Differentiation . . . . . . . . . . . . . . . . . . . 305
8.4 Automatic Differentiation . . . . . . . . . . . . . . . . . . . 305
8.4.1 The concept of automatic or algorithmic
differentiation . . . . . . . . . . . . . . . . . . . . 305
8.4.2 Differentiation of a function with respect
to a vector and matrix . . . . . . . . . . . . . . . 306
8.5 Autograd Implemented in Numpy . . . . . . . . . . . . . . 308
8.6 Autograd Implemented in the MXNet . . . . . . . . . . . . 310
8.6.1 Gradients of scalar functions with simple
variable . . . . . . . . . . . . . . . . . . . . . . . . 311
8.6.2 Gradients of scalar functions in high
dimensions . . . . . . . . . . . . . . . . . . . . . . 313
8.6.3 Gradients of scalar functions with quadratic
variables in high dimensions . . . . . . . . . . . . 318
8.6.4 Gradient of scalar function with a matrix of
variables in high dimensions . . . . . . . . . . . . 319
8.6.5 Head gradient . . . . . . . . . . . . . . . . . . . . 320
8.7 Gradients for Functions with Conditions . . . . . . . . . . 322
8.8 Example: Gradients of an L2 Loss Function for
a Single Neuron . . . . . . . . . . . . . . . . . . . . . . . . 323
8.9 Examples: Differences Between Analytical, Autograd,
and Numerical Differentiation . . . . . . . . . . . . . . . . 327
8.10 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
 Contents xv

9 Solution Existence Theory and

Optimization Techniques 331
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 331
9.2 Analytic Optimization Methods: Ideal Cases . . . . . . . . 332
9.2.1 Least square formulation . . . . . . . . . . . . . . 332
9.2.2 L2 loss function . . . . . . . . . . . . . . . . . . . 333
9.2.3 Normal equation . . . . . . . . . . . . . . . . . . . 334
9.2.4 Solution existence analysis . . . . . . . . . . . . . 334
9.2.5 Solution existence theory . . . . . . . . . . . . . . 336
9.2.6 Effects of parallel data-points . . . . . . . . . . . . 337
9.2.7 Predictability of the solution against
the label . . . . . . . . . . . . . . . . . . . . . . . 337
9.3 Considerations in Optimization for Complex
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
9.3.1 Local minima . . . . . . . . . . . . . . . . . . . . 339
9.3.2 Saddle points . . . . . . . . . . . . . . . . . . . . . 340
9.3.3 Convex functions . . . . . . . . . . . . . . . . . . 343
9.4 Gradient Descent (GD) Method for Optimization . . . . . 344
9.4.1 Gradient descent in one dimension . . . . . . . . . 345
9.4.2 Remarks . . . . . . . . . . . . . . . . . . . . . . . 346
9.4.3 Gradient descent in hyper-dimensions . . . . . . . 347
9.4.4 Property of a convex function . . . . . . . . . . . 348
9.4.5 The convergence theorem for the Gradient
Decent algorithm . . . . . . . . . . . . . . . . . . 349
9.4.6 Setting or the learning rates . . . . . . . . . . . . 351
9.5 Stochastic Gradient Descent . . . . . . . . . . . . . . . . . 353
9.5.1 Numerical experiment . . . . . . . . . . . . . . . . 354
9.6 Gradient Descent with Momentum . . . . . . . . . . . . . 363
9.6.1 The most critical problem with GD methods . . . 363
9.6.2 Formulation . . . . . . . . . . . . . . . . . . . . . 365
9.6.3 Numerical experiment . . . . . . . . . . . . . . . . 368
9.7 Nesterov Accelerated Gradient . . . . . . . . . . . . . . . . 370
9.7.1 Formulation . . . . . . . . . . . . . . . . . . . . . 370
9.8 AdaGrad Gradient Algorithm . . . . . . . . . . . . . . . . 371
9.8.1 Formulation . . . . . . . . . . . . . . . . . . . . . 371
9.8.2 Numerical experiment . . . . . . . . . . . . . . . . 372
9.9 RMSProp Gradient Algorithm . . . . . . . . . . . . . . . . 374
9.9.1 Formulation . . . . . . . . . . . . . . . . . . . . . 375
9.9.2 Numerical experiment . . . . . . . . . . . . . . . . 375
xvi Machine Learning with Python: Theory and Applications

9.10 AdaDelta Gradient Algorithm . . . . . . . . . . . . . . . . 378

9.10.1 The idea . . . . . . . . . . . . . . . . . . . . . . . 378
9.10.2 Numerical experiment . . . . . . . . . . . . . . . . 378
9.11 Adam Gradient Algorithm . . . . . . . . . . . . . . . . . . 381
9.11.1 Formulation . . . . . . . . . . . . . . . . . . . . . 381
9.11.2 Numerical experiment . . . . . . . . . . . . . . . . 382
9.12 A Case Study: Compare Minimization Techniques
Used in MLPClassifier . . . . . . . . . . . . . . . . . . . . 385
9.13 Other Algorithms . . . . . . . . . . . . . . . . . . . . . . . 386
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387

10 Loss Functions for Regression 389

10.1 Formulations for Linear Regression . . . . . . . . . . . . . 390
10.1.1 Mathematical model . . . . . . . . . . . . . . . . . 390
10.1.2 Neural network configuration . . . . . . . . . . . . 390
10.1.3 The xw formulation . . . . . . . . . . . . . . . . . 391
10.2 Loss Functions for Linear Regression . . . . . . . . . . . . 391
10.2.1 Mean squared error loss or L2 loss function . . . . 392
10.2.2 Absolute error loss or L1 loss function . . . . . . . 393
10.2.3 Huber loss function . . . . . . . . . . . . . . . . . 394
10.2.4 Log-cosh loss function . . . . . . . . . . . . . . . . 394
10.2.5 Comparison between these loss functions . . . . . 395
10.2.6 Python codes for these loss functions . . . . . . . 396
10.3 Python Codes for Regression . . . . . . . . . . . . . . . . . 398
10.3.1 Linear regression using high-order polynomial
and other feature functions . . . . . . . . . . . . . 401
10.3.2 Linear regression using Gaussian basis
functions . . . . . . . . . . . . . . . . . . . . . . . 404
10.4 Neural Network Model for Linear Regressions
with Big Datasets . . . . . . . . . . . . . . . . . . . . . . . 406
10.4.1 Setting up neural network models . . . . . . . . . 406
10.4.2 Create data iterators . . . . . . . . . . . . . . . . 409
10.4.3 Training parameters . . . . . . . . . . . . . . . . . 411
10.4.4 Define the neural network . . . . . . . . . . . . . . 412
10.4.5 Define the loss function . . . . . . . . . . . . . . . 412
10.4.6 Use of optimizer . . . . . . . . . . . . . . . . . . . 412
10.4.7 Execute the training . . . . . . . . . . . . . . . . . 412
10.4.8 Examining training progress . . . . . . . . . . . . 413
 Contents xvii

10.5 Neural Network Model for Nonlinear Regression . . . . . . 415

10.5.1 Train models on the Boston housing price
dataset . . . . . . . . . . . . . . . . . . . . . . . . 416
10.5.2 Plotting partial dependence for two features . . . 416
10.5.3 Plot curves on top of each other . . . . . . . . . . 418
10.6 On Nonlinear Regressions . . . . . . . . . . . . . . . . . . . 418
10.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
11 Loss Functions and Models for Classification 421
11.1 Prediction Functions . . . . . . . . . . . . . . . . . . . . . 421
11.1.1 Linear function . . . . . . . . . . . . . . . . . . . . 422
11.1.2 Logistic prediction function . . . . . . . . . . . . . 422
11.1.3 The tanh prediction function . . . . . . . . . . . . 423
11.2 Loss Functions for Classification Problems . . . . . . . . . 423
11.2.1 The margin concept . . . . . . . . . . . . . . . . . 423
11.2.2 0–1 loss . . . . . . . . . . . . . . . . . . . . . . . . 424
11.2.3 Hinge loss . . . . . . . . . . . . . . . . . . . . . . 425
11.2.4 Logistic loss . . . . . . . . . . . . . . . . . . . . . 426
11.2.5 Exponential loss . . . . . . . . . . . . . . . . . . . 427
11.2.6 Square loss . . . . . . . . . . . . . . . . . . . . . . 427
11.2.7 Binary cross-entropy loss . . . . . . . . . . . . . . 429
11.2.8 Remarks . . . . . . . . . . . . . . . . . . . . . . . 432
11.3 A Simple Neural Network for Classification . . . . . . . . . 432
11.4 Example of Binary Classification Using Neural
Network with mxnet . . . . . . . . . . . . . . . . . . . . . 433
11.4.1 Dataset for binary classification . . . . . . . . . . 433
11.4.2 Define loss functions . . . . . . . . . . . . . . . . . 435
11.4.3 Plot the convergence curve of the
loss function . . . . . . . . . . . . . . . . . . . . . 437
11.4.4 Computing the accuracy of the
trained model . . . . . . . . . . . . . . . . . . . . 437
11.5 Example of Binary Classification Using Sklearn . . . . . . 438
11.6 Regression with Decision Tree, AdaBoost,
and Gradient Boosting . . . . . . . . . . . . . . . . . . . . 443
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
12 Multiclass Classification 445
12.1 Softmax Activation Neural Networks for
k-Classifications . . . . . . . . . . . . . . . . . . . . . . . . 445
xviii Machine Learning with Python: Theory and Applications

12.2 Cross-Entropy Loss Function for k-Classifications . . . . . 447

12.3 Case Study 1: Handwritten Digit Classification
with 1-Layer NN . . . . . . . . . . . . . . . . . . . . . . . . 448
12.3.1 Set contexts according to computer hardware . . . 448
12.3.2 Loading the MNIST dataset . . . . . . . . . . . . 448
12.3.3 Set model parameters . . . . . . . . . . . . . . . . 451
12.3.4 Multiclass logistic regression . . . . . . . . . . . . 451
12.3.5 Defining a neural network model . . . . . . . . . . 452
12.3.6 Defining the cross-entropy loss function . . . . . . 452
12.3.7 Optimization method . . . . . . . . . . . . . . . . 453
12.3.8 Accuracy evaluation . . . . . . . . . . . . . . . . . 453
12.3.9 Initiation of the model and training execution . . 453
12.3.10 Prediction with the trained model . . . . . . . . . 455
12.4 Case Study 2: Handwritten Digit Classification with
Sklearn Random Forest Multi-Classifier . . . . . . . . . . . 456
12.5 Case Study 3: Comparison of Random Forest,
Extra-Forest, and Gradient Boosting for
Multi-Classifier . . . . . . . . . . . . . . . . . . . . . . . . 460
12.6 Multi-Classification via TensorFlow . . . . . . . . . . . . . 464
12.7 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
13 Multilayer Perceptron (MLP) for Regression
and Classification 467
13.1 The General Architecture and Formulations of MLP . . . . 467
13.1.1 The general architecture . . . . . . . . . . . . . . 467
13.1.2 The xw+b formulation . . . . . . . . . . . . . . . 469
13.1.3 The xw formulation, use of affine transformation
weight matrix . . . . . . . . . . . . . . . . . . . . 471
13.1.4 MLP configuration with affine transformation
weight matrix . . . . . . . . . . . . . . . . . . . . 473
13.1.5 Space evolution process in MLP . . . . . . . . . . 474
13.2 Neurons-Samples Theory . . . . . . . . . . . . . . . . . . . 474
13.2.1 Affine spaces and the training parameters
used in an MLP . . . . . . . . . . . . . . . . . . . 475
13.2.2 Neurons-Samples Theory for MLPs . . . . . . . . 476
13.3 Nonlinear Activation Functions for the Hidden Layers . . . 478
13.4 General Rule for Estimating Learning Parameters
in an MLP . . . . . . . . . . . . . . . . . . . . . . . . . . . 478
 Contents xix

13.5 Key Techniques for MLP and Its Capability . . . . . . . . 479

13.6 A Case Study on Handwritten Digits Using MXNet . . . . 481
13.6.1 Import necessary libraries and load data . . . . . 481
13.6.2 Set neural network model parameters . . . . . . . 482
13.6.3 Softmax cross entropy loss function . . . . . . . . 482
13.6.4 Define a neural network model . . . . . . . . . . . 483
13.6.5 Optimization method . . . . . . . . . . . . . . . . 484
13.6.6 Model accuracy evaluation . . . . . . . . . . . . . 484
13.6.7 Training the neural network and timing
the training . . . . . . . . . . . . . . . . . . . . . . 484
13.6.8 Prediction with the model trained . . . . . . . . . 486
13.6.9 Remarks . . . . . . . . . . . . . . . . . . . . . . . 487
13.7 Visualization of MLP Weights Using Sklearn . . . . . . . . 488
13.7.1 Import necessary Sklearn module . . . . . . . . . 488
13.7.2 Load MNIST dataset . . . . . . . . . . . . . . . . 488
13.7.3 Set an MLP model . . . . . . . . . . . . . . . . . . 489
13.7.4 Training the MLP model and time the
training . . . . . . . . . . . . . . . . . . . . . . . . 489
13.7.5 Performance analysis . . . . . . . . . . . . . . . . 489
13.7.6 Viewing the weight matrix as images . . . . . . . 490
13.8 MLP for Nonlinear Regression . . . . . . . . . . . . . . . . 490
13.8.1 California housing data and preprocessing . . . . . 492
13.8.2 Configure, train, and test the MLP . . . . . . . . 493
13.8.3 Compute and plot the partial dependence . . . . . 494
13.8.4 Comparison studies on different regressors . . . . 495
13.8.5 Gradient boosting regressor . . . . . . . . . . . . . 495
13.8.6 Decision tree regressor . . . . . . . . . . . . . . . . 498
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
14 Overfitting and Regularization 501
14.1 Why Regularization . . . . . . . . . . . . . . . . . . . . . . 501
14.2 Tikhonov Regularization . . . . . . . . . . . . . . . . . . . 504
14.2.1 Demonstration examples: One data-point . . . . . 508
14.2.2 Demonstration examples: Two data-points . . . . 517
14.2.3 Demonstration examples: Three data-points . . . 521
14.2.4 Summary of the case studies . . . . . . . . . . . . 525
14.3 A Case Study on Regularization Effects using MXNet . . . 526
14.3.1 Load the MNIST dataset . . . . . . . . . . . . . . 527
14.3.2 Define a neural network model . . . . . . . . . . . 527
xx Machine Learning with Python: Theory and Applications

14.3.3 Define loss function and optimizer . . . . . . . . . 527

14.3.4 Define a function to evaluate the accuracy . . . . 528
14.3.5 Define a utility function plotting
convergence curve . . . . . . . . . . . . . . . . . . 528
14.3.6 Train the neural network model . . . . . . . . . . 529
14.3.7 Evaluation of the trained model: A typical case
of overfitting . . . . . . . . . . . . . . . . . . . . . 531
14.3.8 Application of L2 regularization . . . . . . . . . . 531
14.3.9 Re-initializing the parameters . . . . . . . . . . . 531
14.3.10 Training the L2-regularized neural
network model . . . . . . . . . . . . . . . . . . . . 531
14.3.11 Effect of the L2 regularization . . . . . . . . . . . 533
14.4 A Case Study on Regularization Parameters
Using Sklearn . . . . . . . . . . . . . . . . . . . . . . . . . 534
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 538
15 Convolutional Neural Network (CNN)
for Classification and Object Detection 539
15.1 Filter and Convolution . . . . . . . . . . . . . . . . . . . . 539
15.2 Affine Transformation Unit in CNNs . . . . . . . . . . . . 542
15.3 Pooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
15.4 Up Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . 545
15.5 Configuration of a Typical CNN . . . . . . . . . . . . . . . 545
15.6 Some Landmark CNNs . . . . . . . . . . . . . . . . . . . . 546
15.6.1 LeNet-5 . . . . . . . . . . . . . . . . . . . . . . . . 547
15.6.2 AlexNet . . . . . . . . . . . . . . . . . . . . . . . . 548
15.6.3 VGG-16 . . . . . . . . . . . . . . . . . . . . . . . . 549
15.6.4 ResNet . . . . . . . . . . . . . . . . . . . . . . . . 549
15.6.5 Inception . . . . . . . . . . . . . . . . . . . . . . . 551
15.6.6 YOLO: A CONV net for object detection . . . . . 551
15.7 An Example of Convolutional Neural Network . . . . . . . 552
15.7.1 Import TensorFlow . . . . . . . . . . . . . . . . . 553
15.7.2 Download and preparation of a CIFAR10
dataset . . . . . . . . . . . . . . . . . . . . . . . . 553
15.7.3 Verification of the data . . . . . . . . . . . . . . . 553
15.7.4 Creation of Conv2D layers . . . . . . . . . . . . . 554
15.7.5 Add Dense layers to the Conv2D layers . . . . . . 556
15.7.6 Compile and train the CNN model . . . . . . . . . 557
15.7.7 Evaluation of the trained CNN model . . . . . . . 557
 Contents xxi

15.8 Applications of YOLO for Object Detection . . . . . . . . 558

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562
16 Recurrent Neural Network (RNN) and Sequence
Feature Models 563
16.1 A Typical Structure of LSTMs . . . . . . . . . . . . . . . . 564
16.2 Formulation of LSTMs . . . . . . . . . . . . . . . . . . . . 565
16.2.1 General formulation . . . . . . . . . . . . . . . . . 565
16.2.2 LSTM layer and standard neural layer . . . . . . . 566
16.2.3 Reduced LSTM . . . . . . . . . . . . . . . . . . . 566
16.3 Peephole LSTM . . . . . . . . . . . . . . . . . . . . . . . . 567
16.4 Gated Recurrent Units (GRUs) . . . . . . . . . . . . . . . 568
16.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
16.5.1 A simple reduced LSTM with a standard NN layer
for regression . . . . . . . . . . . . . . . . . . . . . 569
16.5.2 LSTM class in tensorflow.keras . . . . . . . . . . . 574
16.5.3 Using LSTM for handwritten digit recognition . . 575
16.5.4 Using LSTM for predicting dynamics of
moving vectors . . . . . . . . . . . . . . . . . . . . 578
16.6 Examples of LSTM for Speech Recognition . . . . . . . . . 584
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584
17 Unsupervised Learning Techniques 585
17.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . 585
17.2 K-means for Clustering . . . . . . . . . . . . . . . . . . . . 585
17.2.1 Initialization of means . . . . . . . . . . . . . . . . 586
17.2.2 Assignment of data-points to clusters . . . . . . . 587
17.2.3 Update of means . . . . . . . . . . . . . . . . . . . 588
17.2.4 Example 1: Case studies on comparison of
initiation methods for K-means clustering . . . . 590
17.2.5 Example 2: K-means clustering on the
handwritten digit dataset . . . . . . . . . . . . . . 601
17.3 Mean-Shift for Clustering Without Pre-Specifying k . . . . 605
17.4 Autoencoders . . . . . . . . . . . . . . . . . . . . . . . . . 609
17.4.1 Basic structure of autoencoders . . . . . . . . . . 610
17.4.2 Example 1: Image compression and denoising . . . 611
17.4.3 Example 2: Image segmentation . . . . . . . . . . 611
17.5 Autoencoder vs. PCA . . . . . . . . . . . . . . . . . . . . . 615
17.6 Variational Autoencoder (VAE) . . . . . . . . . . . . . . . 617
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
xxii Machine Learning with Python: Theory and Applications

18 Reinforcement Learning (RL) 625

18.1 Basic Underlying Concept . . . . . . . . . . . . . . . . . . 625
18.1.1 Problem statement . . . . . . . . . . . . . . . . . 625
18.1.2 Applications in sciences, engineering,
and business . . . . . . . . . . . . . . . . . . . . . 626
18.1.3 Reinforcement learning approach . . . . . . . . . . 627
18.1.4 Actions in discrete time: Solution strategy . . . . 628
18.2 Markov Decision Process . . . . . . . . . . . . . . . . . . . 629
18.3 Policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 630
18.4 Value Functions . . . . . . . . . . . . . . . . . . . . . . . . 630
18.5 Bellman Equation . . . . . . . . . . . . . . . . . . . . . . . 631
18.6 Q-learning Algorithm . . . . . . . . . . . . . . . . . . . . . 633
18.6.1 Example 1: A robot explores a room with
unknown obstacles with Q-learning algorithm . . . 633
18.6.2 OpenAI Gym . . . . . . . . . . . . . . . . . . . . . 635
18.6.3 Define utility functions . . . . . . . . . . . . . . . 636
18.6.4 A simple Q-learning algorithm . . . . . . . . . . . 636
18.6.5 Hyper-parameters and convergence . . . . . . . . 640
18.7 Q-Network Learning . . . . . . . . . . . . . . . . . . . . . . 641
18.7.1 Example 2: A robot explores a room with
unknown obstacles with Q-Network . . . . . . . . 641
18.7.2 Building TensorFlow graph . . . . . . . . . . . . . 642
18.7.3 Results from the Q-Network . . . . . . . . . . . . 644
18.8 Policy gradient methods . . . . . . . . . . . . . . . . . . . 646
18.8.1 PPO with NN policy . . . . . . . . . . . . . . . . 646
18.8.2 Strategy used in policy gradient methods
and PPO . . . . . . . . . . . . . . . . . . . . . . . 647
18.8.3 Ratio policy . . . . . . . . . . . . . . . . . . . . . 649
18.8.4 PPO: Controlling a pole staying upright . . . . . . 650
18.8.5 Save and reload the learned model . . . . . . . . . 654
18.8.6 Evaluate and view the trained model . . . . . . . 654
18.8.7 PPO: Self-driving car . . . . . . . . . . . . . . . . 657
18.8.8 View samples of the racing car before training . . 658
18.8.9 Train the racing car using the CNN policy . . . . 659
18.8.10 Evaluate and view the learned model . . . . . . . 660
18.9 Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662
Index 663
 Chapter 1

Introduction

1.1 Naturally Learned Ability for Problem Solving

We are constantly dealing with all kinds of problems every day, and would
like to solve these problems for timely decisions and actions. We may notice
that for many of the daily-life problems, our decisions are often made spon-
taneously, swiftly without much consciousness. This is because we have been
constantly learning to solve such problems in the past since we were born,
and therefore the solutions have already been encoded in the neuron cells in
our brain. When facing similar problems, our decision is spontaneous.
For many complicated problems, especially in science and engineering,
one would need to think harder and even conduct extensive research and
study on the related issues before we can provide a solution. What if we want
to give spontaneous reliable solutions to these types of problems as well?
Some scientists and engineers may be able to do this for some problems, but
not many. Those scientists are intensively trained or educated in specially
designed courses for dealing with complicated problems.
What if a normal layman would also like to be able solve these challenging
types of problems? One way is to go through a special learning process.
The alternative may be through machine learning, to develop a special
computer model with a mechanism that can be trained to extract features
from experience or data to provide a reliable and instantaneous solution for
a type of problem.

1.2 Physics-Law-based Models

Problems in science and engineering are usually much more difficult to solve.
This is because we humans can only experience or observe the phenomena

1
2 Machine Learning with Python: Theory and Applications

associated with the problem. However, many phenomena are not easily
observable and have very complicated underlying logic. Scientists have been
trying to unveil the underlying logic by developing some theories (or laws or
principles) that can help to best describe these phenomena. These theories
are then formulated in the form of algebraic, differential, or integral system
equations that govern the key variables involved in the phenomena. The
next step is then to find a method that can solve these equations for these
variables varying in space and with time. The final step is to find a way
to validate the theory by observation and/or experiments to measure the
values of these variables. The validated theory is used to build models to
solve problems that exhibit the same phenomena. This type of model is
called physics-law-based model.
The above-mentioned process is essentially what humans on earth have
been doing in trying to understand nature, and we have made tremendous
progress so far. In this process, we have established a huge number of areas
of studies, physics, mathematics, biology, etc., which are now referred to as
sciences.
Understanding nature is only a part of the story. Humans want to
invent and build new things. A good understanding of various phenomena
enables us to do so, and we have practically built everything around us,
buildings, bridges, airplanes, space stations, cars, ships, computers, cell
phones, internet, communication systems, and energy systems. Such a list is
endless. In this process, we humans established a huge number of areas of
development, which we are now referred to as engineering.
Understanding biology helped us to discover medicines, treatments for
illnesses of humans and animals, treatments for plants and the environment,
as well as proper measures and policies dealing with the relationships
between humans, animals, plants, and environments. In this process, we
humans established a huge number of areas of studies, including medicine,
agriculture, and ecology.
In the relentless quest by humans in history, countless theories, laws,
techniques, methods, etc., have been developed in various areas of science,
engineering, and biology. For example, in the study of a small area of compu-
tational mechanics for designing structural systems, we have developed the
finite element method (FEM) [1], smoothed finite element method (S-FEM)
[2], meshfree methods [3, 4], inverse techniques [5], etc., just to name a few
that the author has been working on. It is not possible and necessary to list
all of these kinds of methods and techniques. Our discussion here is just to
provide an overall view of how a problem can be solved based on physics laws.
 Introduction 3

Note that there are many problems in nature, engineering, and society
for which it is difficult to describe and find proper physics laws to accurately
and effectively solve them. Alternative means are thus needed.

1.3 Machine Learning Models, Data-based

There is a large class of complicated problems (in science, engineering,

biology, and daily-life) that do not yet have known governing physics laws, or
the solutions to the governing laws’ equations are too expensive to obtain. For
this type of problem, on the other hand, we often have some data obtained
and accumulated through observations or measurements or historic records.
When the data are sufficiently large and of good quality, it is possible to
develop computer models to learn from these data. Such a model can then be
used to find a solution for this type of problem. This kind of computer model
is defined as a data-based model or machine learning model in this book.
Different types of effective artificial Neural Networks (NNs) with various
configurations have been developed and widely used for practical problems
in sciences and engineering, including multilayer perceptron (MLP) [6–9],
Convolutional Neural Networks (CNNs) [10–14], and Recurrent Neural
Networks (RNNs) [15–17]. TrumpetNets [8] and TubeNets [9, 18–20] were
also recently proposed by the author for creating two-way deepnets using
physics-law-based models as trainers, such as the FEM [1] and S-FEM [2].
The unique feature of TrumpetNets and TubeNets is their effectiveness for
both forward and inverse problems [5]. It has a unique net architecture.
Most importantly, solutions to inverse problems can be analytically derived
in explicit formulae for the first time. This implies that when a data-based
model is built properly, one can find solutions very efficiently.
Machine learning is essentially to mimic the natural learning process
occurring in biological brains that can have a huge number of neurons. In
terms of usage of data, we may have three major categories:

1. Supervised Learning, using data with true labels (teachers).

2. Unsupervised Learning, using data without labels.
3. Reinforcement Learning, using a predefined environment.

In terms of problems to solve, there are the following:

1. Binary classification problems, answer in probability to yes or no.

2. k-classification problems, answer in probabilities to k classes.
3. k-clustering problems, answer in k clusters of data-points.
4 Machine Learning with Python: Theory and Applications

4. Regression (linear or nonlinear), answer in predictions of continuous

functions.
5. Feature extraction, answer in key features in the dataset.
6. Abnormality detection, answer in abnormal data.
7. Inverse analysis, answer in prediction on features from known responses.

In terms of learning methodology or algorithms, we have the following:

1. Linear and logistic regression, supervised.

2. Decision Tree, supervised.
3. Support Vector Machine (SVM), supervised.
4. Naive Bayes, supervised.
5. Multi-Layer Perceptron (MLP) or artificial Neural Networks (NNs),
supervised.
6. k-Nearest Neighbors (kNN), supervised.
7. Random Forest, supervised.
8. Gradient Boosting types of algorithms, supervised.
9. Principal Components Analysis (PCA), unsupervised.
10. K-means, Mean-Shift, unsupervised.
11. Autoencoders, unsupervised.
12. Markov Decision Process, reinforcement Learning.

This book will cover most of these algorithms, but our focus will be
more on neural network-based models because rigorous theory and predictive
models can be established.
Machine learning is a very active area of research and development. New
models, including the so-called cognitive machine learning models, are being
studied. There are also techniques for manipulating various ML models. This
book, however, will not cover those topics.

1.4 General Steps for Training Machine Learning Models

General steps for training machine learning models are summarized as

follows:

1. Obtaining the dataset for the problem, by your own means of data
generation, or imported from other existing sources, or computer
syntheses.
2. Clean up the dataset if there are objectively known defaults in it.
3. Determine the type of hypothesis for the model.
 Introduction 5

4. Develop or import proper module for the needed algorithm for the
problem. The learning ability (number of the learning parameters) of
the model and the size of the dataset shall be properly balanced, if
possible. Otherwise, consider the use of regularization techniques.
5. Randomly initialize the learning parameters, or import some known pre-
trained learning parameter.
6. Perform the training with proper optimization techniques and monitor-
ing measures.
7. Test the trained model using an independent test dataset. This can also
be done during the training.
8. Deploy the trained and tested model to the same type of problems, where
the training and testing datasets are collected/generated.

1.5 Some Mathematical Concepts, Variables, and Spaces

We shall define variables and spaces often used in this book for ease of
discussion. We first state that this book deals with only real numbers, unless
specified when geometrically closed operations are required. Let us introduce
two toy examples.

1.5.1 Toy examples

Toy Example-1, Regression: Assume we are to build a machine learning
model to predict the quality of fruits. Based on its three features, size,
weight, and roundness (that can easily observe and measure), we aim to
establish a machine learning regression model to predict the values of two
characteristics, sweetness and vitamin-C content (that are difficult to
quantify nondestructively), for any given fruit. To build such a model, we
make 8,000 measurements to randomly selected fruits from the market and
create a dataset with 8,000 paired data-points. Each data-point records
the values of these three features and pairs with the values of these two
characteristics. The values of these two characteristics are called labels
(ground truth) to the data-point. The dataset is called labeled dataset that
can be used systematically to train a machine learning model.
Toy Example-2, Classification: Assume we are to build a machine learning
model to classify the type of fruits based on its three features (size, weight,
and roundness). In this case, we want a machine to predict whether any
given frait is an apple or orange, so that it can be packaged separately
in an automatic manner. To achieve this, we make 8,000 measurements to
6 Machine Learning with Python: Theory and Applications

randomly selected fruits of these two types from the market, and create a
dataset with 8,000 paired data-points. Each data-point records the values
of these three features and pairs with two labels (ground truth) of yes-or-no
for apple or yes-or-no for orange. The dataset is also called labeled dataset
for model training.
With an understanding of these two typical types of examples, it should
be easy to extend this to many other types of problems for which a machine
learning model can be effective.

1.5.2 Feature space

Feature space Xp : Machine learning uses datasets that contain observed or
measured p variables of real numbers in R, often called features. In our two
toy examples, p = 3. We may define a p-dimensional feature space Xp which is
a vector space (https://en.wikipedia.org/wiki/Vector space) over real num-
bers in R with inner product defined. A vector in Xp for an arbitrary point
(x1 , x2 , . . . , xp ) is written as
x = [x1 , x2 , . . . , xp ], x ∈ Xp (1.1)
The origin of Xp is at x = [0, 0, . . . , 0] following the standard for all vector
spaces. Note that we use italic for scalar variables, bold face for all vectors
and matrices, and blackboard bold for spaces (or sets or of that nature),
and this convention is followed throughout this book. Also, we define, in
general, all vectors in row vectors by default, as we usually do in Python
programming. A column vector is treated as special case of 2D array (matrix)
with only one column.
It is clear that the feature space Xp is a special (with vector operations
defined) case of the real space Rp . Thus, Xp ∈ Rp .
Also, xi (i = 1, 2, . . . , p) is called linear basis functions (not to be confused
with the basis vectors), because a linear combination of xi gives a new x that
is still in Xp . A two-dimensional (2D) feature space X2 is the black plane
x1 − x2 shown in Fig. 1.1.
An observed data-point xi with p features is a discrete point in the space,
and the corresponding vector xi is expressed as
xi = [xi1 , xi2 , . . . , xip ], xi ∈ Xp , ∀ i = 1, 2, . . . , m (1.2)
where m is the number of measurements or observations or data-points in
the dataset. It is also often referred as number of samples in a dataset. For
these two toy examples, m = 8,000. For the example shown in Fig. 1.1, these
4 blue vectors are for four data-points in space X2 , and m = 4.
 Introduction 7

Figure 1.1: Data-points in a 2D feature space X2 with blue vectors: xi = [xi1 , xi2 ]; and
2
the same data-points in the augmented feature space X , called affine space, with red
vectors: xi = [1, xi1 , xi2 ]; i = 1, 2, 3, 4.

These data-points xi (i = 1, 2, . . . , m) can be stacked to form a dataset

noted as X ∈ Xp . This is for convenience in formulation. We do not form
such a matrix in computation because it is usually very large for big datasets
with large m.

1.5.3 Affine space

p
Affine space X : It is an augmented feature space. It is the red plane
shown in Fig. 1.1. It has a “complete” linear bases (or basis functions):
x = [1, x1 , x2 , . . . , xp ] (1.3)
By complete linear bases, we mean all bases up to the 1st order of all the
variables including the 0th order. The 0th order basis is the constant basis 1
that provides the augmentation. Affine space is not a vector space, because
p p
0∈ / X and (xi + xj ) ∈
/ X where i, j=1 or 2 or 3 or 4 in Fig. 1.1. This special
and fundamentally useful space always has a constant 1 as a component, and
thus it does not have an origin by definition. Operation that occurs on an
affine space and still stays in an affine space is called affine transformation.
It is the most essential operation in major machine learning models, and the
fundamental reason for such models being predictive.
An observed data-point with p features can also be presented as an
p
augmented discrete point in the X space and can be expressed by
p
xi = [1, xi1 , xi2 , . . . , xip ], xi ∈ X , ∀ i = 1, 2, . . . , m (1.4)
8 Machine Learning with Python: Theory and Applications

p
A X space can be created by first spanning Xp by one dimension to Xp+1
via introduction of a new variable x0 as

[x0 , x1 , x2 , . . . , xp ] (1.5)

and then set x0 = 1. These 4 red vectors shown in Fig. 1.1 live in an affine
2
space X .
p
Note that the affine space X is neither Xp+1 nor Xp , and is quite
p
special. A vector in a X is in Xp+1 , but the tip of the vector is confined
in “hyperplane” of x0 = 1. For convenience of discussion in this book, we
say that an affine space has a pseudo-dimension that is p + 1. Its true
dimension is p, but it is a hyperplane in a Xp+1 space.
In terms of function approximation, the linear bases given in Eq. (1.3)
can be used to construct any arbitrary linear function in the feature
space. A proper linear combination of these complete linear bases is still
in the affine space. Such a combination can be used to perform an affine
transformation, which will be discussed in detail in Chapter 5.
These data-points xi (i = 1, 2, . . . , m) are stacked to form an augmented
p
dataset X ∈ X , which is the well-known moment matrix in function
approximation theory [1–4]. Again, this is for convenience in formulation.
We may not form such a matrix in computation.

1.5.4 Label space

Label space Yk : Consider a labeled dataset for a supervised machine
learning model creation. We shall introduce variables (y1 , y2 , . . . , yk ) of real
numbers in R. For toy example-1, k = 2. We may define a label space Yk over
real numbers. It is a vector space. A vector in space Yk for can be written as

y = [y1 , y2 , . . . , yk ], y ∈ Yk ∈ Rk (1.6)

A label in a dataset is paired with a data-point. The label for data-point xi

which is denoted as yi can be expressed as

yi = [yi1 , yi2 , . . . , yik ], yi ∈ Yk , ∀i = 1, 2, . . . , m (1.7)

For the toy example-1, yij (i = 1, 2, . . . , 8000; j = 1, 2) are 8,000 real numbers
in 2D space Y2 . For the toy example-2, each label, yi1 or yi2 , has a value of
0 or 1 (or −1 or 1), but the labels can still be viewed living in Y2 .
These labels yi (i = 1, 2, . . . , m) can be stacked to form a label set Y ∈ Yk ,
although we may not really do so in computation.
 Introduction 9

Typically, affine transformations end at the output layer in a neural

network and produces a vector in a label space, so that a loss function can
be constructed there for “terminal control”.

1.5.5 Hypothesis space

The learning parameters ŵ in a machine learning model are continuous vari-
ables that live in a hypothesis space noted as WP over the real numbers.
Learning parameters are also called training or trainable parameters. We
use these terms interchangeably. The learning parameters include weights
and biases in each and all the layer. The hat above w implying that it
is a collection of all weights and biases, so that we have single notation
in a vector for all learning parameters. Its dimension P depends on type
of hypothesis used including the configuration of neural networks or ML
models. These parameters always work with feature vectors, resulting in
intermediate feature vectors in a new feature space or in a label space,
thorough a properly designed architecture.
These parameters need to be updated which involves vector operations.
To ensure convergence, we would need the vector of all learning parameters
obey important vector properties, such as inner products, norms and the
Cauchy-Schwartz inequality, etc. We will do such proofs multiple times in
this book. Therefore, we require WP be a vector space, so that each update
to the current learning parameters results new parameters that are still in
the same vector space, until they converge.
Note that the learning parameters, in general, are in matrix form or
column vectors (that can be viewed as a special case of matrix). In a typical
machine learning model, there could be multiple matrices of different sizes.
These matrices form affine transformation matrices that operates on
features on affine spaces. A component in a “vector” of the hypothesis space
can be in fact a matrix in general, and thus it is not easy to comprehend
intuitively. The easiest (and valid) way is to “flatten” all the matrix and
then “concatenate” them together to form a tall vector, and then treat it as
a usual vector. We do this kind of flattening and concatenation all the time
in Python. Such a flattened tall vector ŵ in the hypothesis space WP can
be written generally as,

ŵ = [W 0 , W 1 , . . . , W P ] ∈ WP (1.8)

We will discuss in later chapters the details about WP for various models
including estimation of the dimension P .
10 Machine Learning with Python: Theory and Applications

1.5.6 Definition of a typical machine learning model,

a mathematical view
Finally, we can define mathematically ML models for prediction as a mapping
operator:
p
M(ŵ ∈ WP ; X ∈ X , Y ∈ Yk ) : Xp → Yk (1.9)

It reads that the ML model M uses a given dataset X with Y to train its
learning parameters ŵ, and produces a map (or giant functions) that makes
a prediction in the label space for any point in the feature space.
The ML model shown in Eq. (1.9) is in fact a data-parameter
converter: it converts a given dataset to learning parameters during training
and then converts the parameters back in making a prediction for a given set
of feature variables. It can also be mathematically viewed as a giant function
with k components in the feature space Xp and controlled (parameterized)
by the training parameters in WP . When the parameters are tuned, one gets
a set of k giant functions over the feature space.
On the other hand, this set of k giant functions can also be viewed as
continuous (differentiable) functions of these parameters for any given data-
point in the dataset, which can be used to form a loss function that is also
differentiable. Such a loss function can be the error between these k giant
functions and the corresponding k labels given in the dataset. It can be
viewed as a functional of prediction functions that in turn are functions of
ŵ in the vector space WP . The training is to minimize such a loss function
for all the data-points in the dataset, by updating the training parameters
to become minimizers. This overview picture will be made explicitly into
a formula in later chapters. The success factors for building a quality ML
model include: (1) type of hypothesis, (2) number of learning parameters
in WP , (3) quality (representativeness to the underlaying problem to be
modeled, including correctness, size, data-point distribution over the features
space, and noise level) of the dataset in Xp , and (4) techniques to find the
minimizer of learning parameters to best produce the label in the dataset.
We will discuss this in detail in later chapters for different machine learning
models.
Concepts on spaces are helpful in our later analysis of the predictive
properties of machine learning models. Readers may find difficulty in
comprehending these concepts at this stage, and thus are advised to just have
some rough idea for now and to revisit this section when reading relevant
chapters. Readers may jump to Section 13.1.5 and take a look at Eq. (13.13)
there just for a quick glance on how the spaces evolve in a deepnet.
 Introduction 11

Note also that there are ML models for discontinuous feature variables,
and the learning parameters may not need to be continuous. Such methods
are often developed based on proper intuitive rules and techniques, and
we will discuss some of those. The concepts on spaces may not be directly
applicable but can often help.

1.6 Requirements for Creating Machine Learning Models

To train a machine learning model, one would need the following:

1. A dataset, which may be obtained via observations, experiments, and

physics-law-based models. The dataset is usually divided (in a random
manner) into two mutually independent subsets, training dataset and
testing dataset, typically at a rate of 75:25. The independence of the test-
ing dataset is critical, because ML models are determined largely by the
training dataset, and hence their reliability depends on objective testing.
2. Labels with the dataset, if possible.
3. Prior information on the dataset if possible, such as the quality of the
data and key features of the data. This can be useful in choosing a proper
algorithm for the problem, and in application of regularization techniques
in the training.
4. Proper computer software modules and/or effective algorithms.
5. A computer, preferably connected to the internet.

1.7 Types of Data

Data are the key to any data-based models. There are many types of data
available for different types of problems that one may make use of as follows:

• Images: photos from cameras (more often now cellphones), images

obtained from the open websites, computer tomography (CT), X-ray,
ultrasound, Magnetic resonance imaging (MRI), etc.
• Computer-generated data: data from proven physics-law-based mod-
els, other surrogate models, other reliable trained machine learning
models, etc.
• Text: unclassified text documents, books, emails, webpages, social media
records, etc.
• Audio and video: audio and video recordings.
12 Machine Learning with Python: Theory and Applications

Note that the quality and the sampling domain of the dataset play
important roles in training reliable machine learning models. Use of a trained
model beyond the data sampling domain requires a special caution, because
it can go wrong unexpectedly, and hence be very dangerous.

1.8 Relation Between Physics-Law-based and

Data-based Models

Machine learning models are in general slow learners, fast predictors,

while physics-law-based models do not need to learn (using existing laws),
but are slow in prediction. This is because the strategies for physics-law-
based models and those for data-based models are quite different. ML models
use datasets to train the parameters, but physics-law-based models use laws
to determine the parameters.
However, at the detailed computational methodology level, many tech-
niques used in both models are in fact the same or quite similar. For example,
when we express a variable as a function of other variables, both models
use basis functions (polynomial, or radial basis function (RBF), or both).
In constructing objective functions, the least squares error formulation is
used in both. In addition, the regularization methods used are also quite
similar. Therefore, one should not study these models in total isolation. The
ideas and techniques may be deeply connected and mutually adaptable. This
realization can be useful in better understanding and further development
of more effective methods for both models, by exchanging the ideas and
techniques from one to another. In general, for physics-law-based computa-
tional methods, such as the general form of meshfree methods, we understand
reasonably well why and how a method works in theory [3]. Therefore, we are
quite confident about what we are going to obtain when a method is used
for a problem. For data-based methods, however, this is not always true.
Therefore, it is of importance to develop fundamental theories for data-based
methods. The author made some attempts [21] to reveal the relationship
between physics-law-based and data-based models, and to establish some
theoretical foundation for data-based models. In this book, we will try to
discuss the similarities and differences, when a computational method is
used in both models.

1.9 This Book

This book offers an introduction to general topics on machine learning. Our

focus will be on the basic concepts, fundamental theories, and essential
 Introduction 13

computational techniques related to creation of various machine learn-

ing models. We decided not to provide a comprehensive document for
all the machine learning techniques, models, and algorithms. This is
because the topic of machine learning is very extensive and it is not possible
to be comprehensive in content. Also, it is really not possible for many read-
ers to learn all the content. In addition, there are in fact plenty of documents
and codes available publicly online. There is no lack of material, and there is
no need to simply reproduce these materials. In the opinion of the author, the
best learning approach is to learn the most essential basics and build a strong
foundation, which is sufficient to learn other related topics, methods, and
algorithms. Most importantly, readers with strong fundamentals can even
develop innovative and more effective machine models for their problems.
Based on this philosophy, the highlights of the book that cannot be found
easily or in good completion in the open literature are listed as follows, many
of which are the outcomes of author’s study in the past years:

1. Detailed discussion on and demonstration of predictability for arbitrary

linear functions of the basic hypothesis used in major ML models.
2. Affine transformation properties and their demonstrations, affine space,
affine transformation unit, array, chained arrays, roles of the weights and
biases, and roles of activation functions for deepnet construction.
3. Examination of predictability of high-order functions and a Universal
Prediction Theory for deepnets.
4. A concept of data-parameter converter, parameter encoding, and unique-
ness of the encoding.
5. Role of affine transformation in SVM, complete description of SVM
formulation, and the kernel trick.
6. Detailed discussion on and demonstration of activation functions,
Neural-Pulse-Unit (NPU), leading to the Universal Approximation
Theorem for wide-nets.
7. Differentiation of a function with respect to a vector and matrix, leading
to automatic differentiation and Autograd.
8. Solution Existence Theory, effects of parallel data-points, and pre-
dictability of the solution against the label.
9. Neurons-Samples Theory gives, for the first time, a general rule of thumb
on relationship between the number of data-points and the number
neurons in a neural network (or the total pseudo-dimensions of affine
spaces involved).
10. Detailed discussion on and demonstration of Tikhonov regularization
effects.
14 Machine Learning with Python: Theory and Applications

The author has made substantial effort to write Python codes to demonstrate
the essential and difficult concepts and formulations, which allows readers
to comprehend each chapter earlier. Based on the learning experience of the
author, this can make the learning more effective.
The chapters of this book are written, in principle, readable indepen-
dently, by allowing some duplicates. Necessary cross-references between
chapters provided are kept minimum.

1.10 Who May Read This Book

The book is written for beginners interested to learn the basics of machine
learning, including university students who have completed their first
year, graduate students, researchers, and professionals in engineering and
sciences. Engineers and practitioners who want to learn to build machine
learning models may also find the book useful. Basic knowledge of college
mathematics is helpful in reading this book smoothly.
This book may be used as a textbook for undergraduates (3rd year or
senior) and graduate students. If this book is adopted as a textbook, the
instructor may contact the author (liugr100@gmail.com) directly for some
homework and course projects and solutions.
Machine learning is still a fast developing area of research. There still exist
many challenging problems, which offer ample opportunities for research to
develop new methods and algorithms. Currently, it is a hot topic of research
and applications. Different techniques are being developed every day, and
new businesses are formed constantly. It is the hope of the author that this
book can be helpful in studying existing and developing machine learning
models.

1.11 Codes Used in This Book

The book has been written using Jupiter Notebook with codes.
Readers who purchased the book may contact the author directly
(mailto:liugr100@gmail.com) to request a softcopy of the book with codes
(which may be updated), free for academic use after registration. The
conditions for use of the book and codes developed by the author, in both
hardcopy and softcopy, are as follows:

1. Users are entirely at their own risk using any of part of the codes and
techniques.
 Introduction 15

2. The book and codes are only for your own use. You are not allowed to
further distribute without permission from the author of the code.
3. There will be no user support.
4. Proper reference and acknowledgment must be given for the use of the
book, codes, ideas, and techniques.

Note that the handcrafted codes provided in the book are mainly for
studying and better understanding the theory and formulation of ML
methods. For production runs, well-established and well-tested packages
should be used, and there are plenty out there, including but not limited
to Scikit learn, PyTouch, TensorFlow, and Keras. Also, our codes provided
are often run with various packages/modules. Therefore, care is needed when
using these codes, because the behavior of the codes often depends on the
versions of Python and all these packages/modules. When the codes do not
run as expected, version mismatch could be one of the problems. When this
book was written, the versions of Python and some of the packages/modules
were as follows:

• Python 3.6.13 :: Anaconda, Inc.

• Jupyter Notebook (web-based) 6.3.0
• TensorFlow 2.4.1
• keras 2.4.3
• gym 0.18.0

When issues are encountered in running a code, readers may need to

check the versions of the packages/modules used. If Anaconda Navigator
is used, the versions of all these packages/modules installed with the Python
environment are listed when the Python environment are highlighted. You
can also check the versions of a package in a code cell of the Jupyter
Notebook. For example, to check the version of the current environment
of Python, one may use

!python -V # ! is used to execute an external command

Python 3.6.13 :: Anaconda, Inc.

To check the version of a package/module, one may use

• import package name

• print(‘package name version’,package name)
16 Machine Learning with Python: Theory and Applications

For example,

import keras
print('keras version',keras.__version__)
import tensorflow as tf
print('tensorflow version',tf.version.VERSION)

keras version 2.4.3

tensorflow version 2.4.1

If the version is indeed an issue, one would need to either modify the code
to fit the version or install the correct version in your system, by may be
creating an alternative environment. It is very useful to query on the web
using the error message, and solutions or leads can often be found. This is
the approach the author often takes when encountering an issue in running
a code. Finally, this book has used materials and information available on
the web with links. These links may change over time, because of the nature
of the web. The most effective way (and often used by the author) to dealing
with this matter is to use keywords to search online, if the link is lost.

References

[1] G.R. Liu and S.S. Quek, The Finite Element Method: A Practical Course,
Butterworth-Heinemann, London, 2013.
[2] G.R. Liu and T.T. Nguyen, Smoothed Finite Element Methods, Taylor and Francis
Group, New York, 2010.
[3] G.R. Liu, Mesh Free Methods: Moving Beyond the Finite Element Method, Taylor
and Francis Group, New York, 2010.
[4] G.R. Liu and Gui-Yong Zhang, Smoothed Point Interpolation Methods: G Space
Theory and Weakened Weak Forms, World Scientific, New Jersey, 2013.
[5] G.R. Liu and X. Han, Computational Inverse Techniques in Nondestructive Evalua-
tion, Taylor and Francis Group, New York, 2003.
[6] F. Rosenblatt, Principles of Neurodynamics: Perceptrons and the Theory of
Brain Mechanisms, New York, 1962. https://books.google.com/books?id=7FhRAA
AAMAAJ.
[7] D.E. Rumelhart, G.E. Hinton and R.J. Williams, Learning Internal Representations
by Error Propagation, 1986.
[8] G.R. Liu, FEA-AI and AI-AI: Two-way deepnets for real-time computations for both
forward and inverse mechanics problems, International Journal of Computational
Methods, 16(08), 1950045, 2019.
[9] G.R. Liu, S.Y. Duan, Z.M. Zhang et al., TubeNet: A special trumpetnet for explicit
solutions to inverse problems, International Journal of Computational Methods,
18(01), 2050030, 2021. https://doi.org/10.1142/S0219876220500309.
 Introduction 17

[10] Fukushima Kunihiko, Neocognitron: A self-organizing neural network model for a

mechanism of pattern recognition unaffected by shift in position, Biological Cyber-
netics, 36(4), 193–202, Apr 1980. https://doi.org/10.1007%2Fbf00344251.
[11] D. Ciregan, U. Meier and J. Schmidhuber, Multi-column deep neural networks
for image classification, 2012 IEEE Conference on Computer Vision and Pattern
Recognition, 2012.
[12] M.V. Valueva, N.N. Nagornov, P.A. Lyakhov et al., Application of the residue number
system to reduce hardware costs of the convolutional neural network implementation,
Mathematics and Computers in Simulation, 177, 232–243, 2020.
[13] Duan Shuyong, Ma Honglei, G.R. Liu et al., Development of an automatic lawnmower
with real-time computer vision for obstacle avoidance, International Journal of
Computational Methods, Accepted, 2021.
[14] Duan Shuyong, Lu Ningning, Lyu Zhongwei et al., An anchor box setting technique
based on differences between categories for object detection, International Journal of
Intelligent Robotics and Applications, 6, 38–51, 2021.
[15] M. Warren and P. Walter, A logical calculus of ideas immanent in nervous activity,
Bulletin of Mathematical Biophysics, 5, 127–147, 1943.
[16] J. Schmidhuber, Habilitation Thesis: An Ancient Experiment with Credit Assignment
Across 1200 Time Steps or Virtual Layers and Unsupervised Pre-training for a
Stack of Recurrent NNs, 1993, TUM. https://people.idsia.ch//∼juergen/habilitation/
node114.html.
[17] Yu Yong, Si Xiaosheng, Hu Changhua et al., A review of recurrent neural networks:
LSTM cells and network architectures, Neural Computation, 31(7), 1235–1270,
2019. https://direct.mit.edu/neco/article/31/7/1235/8500/A-Review-of-Recurrent-
Neural-Networks-LSTM-Cells.
[18] L. Shi, F. Wang, S. Duan et al., Two-way TubeNets uncertain inverse methods for
improving positioning accuracy of robots based on interval, The 11th International
Conference on Computational Methods (ICCM2020), 2020.
[19] Duan Shuyong, Shi Lutong, G.R. Liu et al., An uncertainty inversion technique using
two-way neural network for parameter identification of robot arms, Inverse Problems
in Science & Engineering, 29, 3279–3304, 2021.
[20] Duan Shuyong, Wang Li, G.R. Liu et al., A technique for inversely identifying joint-
stiffnesses of robot arms via two-way TubeNets, Inverse Problems in Science &
Engineering, 13, 3041–3061, 2021.
[21] G.R. Liu, A neural element method, International Journal of Computational Methods,
17(07), 2050021, 2020.
MACHINE LEARNING
WITH PYTHON
 Chapter 2

Basics of Python

This chapter discusses basics of Python language for coding machine learning
models. Python is a very powerful high-level programming language with
the need for compiling, but with some level of efficiency of machine-level
language. It has become the top popular tool for the development of tools and
applications in the general area of machine learning. It has rich libraries for
open access, and new libraries are constantly being developed. The language
itself is powerful in terms of functionality. It is an excellent tool for effective
and productive coding and programming. It is also fast, and the structure
of the language is well built for making use of bulky data, which is often the
case in machine learning.
This chapter is not a formal training on Python, but just to help readers
have a smoother start in learning and practicing the materials in the later
chapters. Our focus will be on some useful simple tricks that are familiar
to the author, and some behavior subtleties that often affect our coding in
ML. Readers familiar with Python may simply skip this chapter. We will
use the Jupyter Notebook as the platform for the discussions, so that the
documentation and demonstration can be all in a single file.
You may go online and have the Jupyter Notebook installed at, for
example, https://www.anaconda.com/distribution/, where you can have the
Jupyter Notebook and Python installed at the same time, and maybe along
with another useful Python IDE (Integrated Development Environment)
called Spyder. In my laptop, I have all three pieces ready to use.
A Jupyter Notebook consists of “cells” of different types: cells for codes
and cells for text called “makedown”. Each cell is framed with color borders,
and the color shows up when the cell is clicked on. A green color border
indicates that this cell is in the input mode, and one can type and edit the
contents. Pressing “ctrl + Enter” within the cell, the green border changes

19
20 Machine Learning with Python: Theory and Applications

to blue color, indicating that this cell is formatted or executed, and may
produce an outcome. Double clicking on the blue framed cell sets it back to
the input mode. The right vertical border is made thicker for better viewing.
This should be sufficient for us to get going. One will get more skills (such
as adding cells, deleting cells, and converting cell types) by playing and
navigating among the menu bars on the top of the Notebook window.
Googling the open online sources is excellent for getting help when one
has a question. The author does this all the time. Sources of the reference
materials include the following:

• https://docs.python.org/3.7/
• https://docs.scipy.org/doc/numpy/reference/?v=20191112052936
• https://medium.com/ibm-data-science-experience/markdown-for-jupyte
r-notebooks-cheatsheet-386c05aeebed
• https://www.dataquest.io/blog/jupyter-notebook-tips-tricks-shortcuts/
• https://www.python.org/about/
• https://www.learnpython.org/
• https://en.wikipedia.org/wiki/Python (programming language)
• https://www.learnpython.org/en/Basic Operators+
• https://www.python.org/+
• https://jupyter.org/
• https://www.youtube.com/watch?v=HW29067qVWkb
• https://pynative.com/

The following lists the details of versions of modules related to jupyter-

notebook in the current installation in the author’s laptop (Use “!” to execute
an external command):

!jupyter --version
jupyter core : 4.7.1
jupyter-notebook : 6.3.0
qtconsole : not installed
ipython : 7.16.1
ipykernel : 5.3.4
jupyter client : 6.1.12
jupyter lab : not installed
nbconvert : 6.0.7
ipywidgets : 7.6.3
nbformat : 5.1.3
traitlets : 4.3.3
 Basics of Python 21

2.1 An Exercise

Let us have a less conventional introduction here. Different from other books
on computer languages, we start the discussion on how to make use of our
own codes that we may develop during the course of study.
First, we “import” the Python system library or module from external or
internal sources, so that functions (also called methods) there can be used
in our code for interaction with your computer system. The most important
environment setting is the path.

import sys # import an external module "sys" which

# provides tools for accessing the computer system.
sys.path.append('grbin')
# I made a code in folder grbin in the current
# working directory, and want to use it later.
#print(sys.path)
# check the current paths To execute this or any
# cell use Ctrl-Enter (hold on Ctrl and press Enter)

Note that “#” in a code cell starts a comment line. It can be put anywhere
in a line in a code. Everything in the line behind # becomes comments, and
is not treated as a part of the code.
One may remove “#” in front of print(sys.path), execute it, and a number
of paths will be printed. Many of them were set during the installations
of the system and various codes, including the Anaconda and Python.
“grbin” in the current working directory has just been added in using the
sys.path.append().
When multiple lines of comments are needed, we use “doc-strings” as
follows:
'''Inside here are all comments with multiple lines. It is \
a good way to convey information to users, co-programmers. \
Use a backslash to break a line.'''

'Inside here are all comments with multiple lines. It is a

good way to convey information to users, co-programmers.
Use a backslash to break a line.'

Just for demonstration purposes, we now import our own “module” (a

Python file named as grcodes.py) “grcodes”, and then give it an alias “gr” for
easy reference later, when accessing the attributes, functions, classes, etc.,
inside the module.
22 Machine Learning with Python: Theory and Applications

import grcodes as gr # a Python code grcodes.py in 'grbin'.

The following cell contains the Python code “grcodes.py”. Readers may
create the “grcodes.py” file and put it in the folder “grbin” (or any other
folder), so that the cell above can be executed and “gr.printx()” can be used.

from __future__ import print_function # import external module

import sys
# Define function
def printx(name):
""" This prints out both the name and its value together.
usage: name = 88.0; printx('name') """
frame=sys._getframe(1)
print(name,'=',repr(eval(name,frame.f_globals,frame.f_locals)))

Let us try to use a function in the imported module grcodes, using its
alias gr.

x = 1.0 # Assign x a value.

print(x) # The Python built-in print() function prints
# the value of the given argument x.
gr.printx('x') # a function from the gr module. It prints the
# argument name, and its value at the same time.
1.0
x = 1.0

help(gr.printx) # Find out the usage of the gr.printx function

Help on function printx in module grcodes:

printx(name)
This prints out both the name and its value together.
usage: name = 88.0; printx('name')

Nice. I have actually completed a simple task of printing out “x” using
Python, and in two ways. The gr.printx function is equivalent to doing the
following:

print('x=',x) # you must type the same x twice

x= 1.0

Notice in this case that you must type the same x twice, which gives room
for error. A good code should have as little as possible repetition, allowing
 Basics of Python 23

easy maintenance. When a change is needed, the programmer (or others

using or maintaining the code) shall just need to do it once.
Alternatively, one can import functions from a module in the following
alternative manner:

from grcodes import printx # you may import more functions

# by adding the function names separated with ",".
#from grcodes import * # Import everything from grcodes
# This is not a very good practice, because it can
# lead to problems when some function names in
# grcodes happened to be the same as those in the code.

In this case, we can now use the imported functions as if it is written in

the current file (notebook).

gr.printx('x')
printx('x') # Notice that "gr." is no longer needed.
x = 1.0
x = 1.0

2.2 Briefing on Python

Now, what is Python? Python was created by Guido van Rossum and first
released in 1991. Python’s design philosophy emphasizes code “readability”.
It uses an object-oriented approach aiming to help programmers to write
clear, less repetitive, logical codes for small- and large-scale projects that
may have teams of people working together.
Python is open source, and its interpreters are available for many
operating systems. A global community of programmers develops and
maintains CPython, an open-source reference implementation. A non-
profit organization, the Python Software Foundation, manages and directs
resources for Python and CPython development.
The language’s core philosophy is summarized in the document The Zen
of Python (PEP 20), which includes aphorisms such as the following:

• Beautiful is better than ugly.

• Explicit is better than implicit.
• Simple is better than complex.
• Complex is better than complicated.
• Readability counts.
24 Machine Learning with Python: Theory and Applications

Guido van Rossum manages Python development projects together with

a steering council. There are two major Python versions, Python 2 and
Python 3, and they are quite different. Python 3.0, released 2008, was a
major revision. It is not completely backward-compatible with Python 2.
Due to the number of codes written for Python 2, support for Python 2.7
(the last release in the 2.x series) was extended to 2020. At the time of
writing this book, Python 3.9 had already been released. This tutorial uses
Python 3.6 because it supports more existing libraries and modules.
There are a huge number (probably in the order of hundreds) of computer
programming languages developed so far. The author’s first experience with
computer programming languages was in the 1970s, when learning BASICS
for programming. He used ALGOL60 later and then FORTRAN for a long
time from the 1970s till today, along with limited use of Matlab, C, C++,
and now Python. Any programming language has a complicated syntax and
deeply organized logic. For a user like the author, the best approach to learn a
computer programming language is via examples and practice, while paying
attention to the syntax, property, and the behavior. For a beginner, following
examples is probably the best approach to get started. This will be the
guidance in writing this section of the book. For rigorous syntax, readers may
read the relevant documentations that are readily available online. We will
give a lot of examples, with explanations in the form of comments (as a
programmer often does). All these examples may be directly executed within
this book while reading, so that readers can have a real-time feeling for
easy observation of the behavior and hence comprehension. Readers may
also make use of it via a simple copy and paste to form his/her notebook.
Because of this example-based approach, the discussions on different topics
may jump a little here and there.
To write and execute a Python code, one often uses an IDE. Jupyter
Notebook, PyCharm, and Spyder are among the popular IDEs. In this
book, we use Jupyter Notebook (https://jupyter.org/) via the distributor
Anaconda (https://www.anaconda.com/). Jupyter Notebook can be used
not only as an IDE but also as a nice document generator allowing
description text (markdown cells) and code cells to be edited together in
one document. The lecture notes used in the author’s machine learning
course have also been mostly developed using Jupyter Notebook. The
documents created using Jupyter Notebook can be exported in various types
of documents, including ascii doc, html, latex, markdown, pdf, Python(.py),
and slides(.slides.html). Readers and students may use this notebook as a
 Basics of Python 25

template for your documents (project reports, homeworks, etc.), if so desired.

If one needs to use a spelling check when typing in the markdown cells in
a Jupyter Notebook, the following commands should be executed in the
Anaconda Prompt:

• %pip install jupyter contrib nbextensions.

• %jupyter contrib nbextension install --user.
• %jupyter nbextension enable spellchecker/main.

This would mark the misspelled words for you (but will not provide
suggestions). Other necessary modules with add-on functions may also be
installed in a similar manner.
This book covers in a brief manner a tiny portion of Python.

2.3 Variable Types

Python is said to be object oriented. Every variable in Python is an object. It

is “dynamically typed”: the variable type is determined at the point when it
is typed. You do not need to declare variables before using them, or declare
their types. It has some basic types of variables: Numbers and Strings. These
variables can stand alone, or form a Tuple, List, Dictionary, Set, Numpy
Arrays, etc. They all can be subjected to various operations (arithmetic,
boolean, logical, formatting, etc.) in a code. Note that the variable is loosely
defined, meaning it could be a Tuple, List, Dictionary, etc. For example, a
List can be in a List, a Tuple in a List, or a List in a Tuple.

2.3.1 Numbers

Python supports three types of numbers — integers, floating point

numbers, and complex numbers.
To define an integer variable, one can simply assign it with an integer.

my_int = 48 # by this assignment, my_int becomes an integer

print(my_int); printx('my_int')

48
my int = 48
26 Machine Learning with Python: Theory and Applications

type(my_int) # Check the type of the variable. print() is not

# needed, if it is the last line in the cell

int

my_int = 5.0 # by this my_int becomes now a float

type(my_int)

float

my_complex=8.0+5.0j # by this my_complex is a complex number

print(my_complex)
printx('my_complex')

(8+5j)
my complex = (8+5j)

my_int, my_float, my_string = 20, 10.0, "Hello!"

if my_string == "Hello!": # comparison operators: ==, !=, <, <=, >, >=
print("A string: %s" % my_string) # Indented 4 spaces

if isinstance(my_float, float) and my_float == 10.0:

#isinstance():returns if an object is an instance of a class
print("This is a float, and it is %f" % my_float)

if isinstance(my_int, int) and my_int == 20:

print("This is an integer, and it is: %d" % my_int)

A string: Hello!
This is a float, and it is 10.000000
This is an integer, and it is: 20

To list all variables, functions, modules currently in the memory, try this:

#%whos # you may remove "#" and try this

 Basics of Python 27

The type of a variable can be converted:

my_float = 5.0 # by this assignment, my_float becomes a float

print(my_float)
my_float = float(6)
# create a float, by converting an integer, using float()
print(my_float)
print(int(7.0)) # float is converted to integer.
printx('int(7.0)')

5.0
6.0
7
int(7.0) = 7

To check the memory address of a variable, use

a = 1.0
print('a=',a, 'at memory address:',id(a))

a= 1.0 at memory address: 1847993237600

b = a
print('b=',b, 'at memory address: ',id(b))

b= 1.0 at memory address: 1847993237600

Notice that ‘b’ has the same address of ‘a’.

a, b = 2.0, 3.0
print('a=',a, 'at memory address: ',id(a))
print('b=',b, 'at memory address: ',id(b))

a= 2.0 at memory address: 1847974064064

b= 3.0 at memory address: 1847974063944

Notice the change in address when the value of a variable changes.

28 Machine Learning with Python: Theory and Applications

2.3.2 Underscore placeholder

n1=100000000000
n_1=100_000_000_000 # for easy reading
print('Yes, n1 is same as n_1') if n1==n_1 else print('No')
# Ternary if Statement
n2=1_000_000_000
print('Total=',n1+n2)
print('Total=',f'{n1+n2:,}') # f-string (Python3.6 or later)
total=n1+n2
print('Total=',f'{total:_}')

Yes, n1 is same as n 1
Total= 101000000000
Total= 101,000,000,000
Total= 101 000 000 000

2.3.3 Strings

Strings are bits of text, which are very useful in coding in generating labels
and file names for outputs. Strings can be defined with anything between
quotes. The quote can be either a pair of single quotes or a pair of double
quotes.

my_string = "How are you?"

# spring is defined, the characters in it can be indexed
print(my_string, my_string[0:3],my_string[5],my_string[10:])
my_string = 'Hello,' + " hello!" + " I am here."
# note "+" operator for strings is concatenation
print(my_string)

How are you? How r u?

Hello, hello! I am here.

Although both single and double quotes can all be used, when there are
apostrophes in a string, one should use double quotes, or these apostrophes
would terminate the string if single quotes are used and vice versa. For
example,
 Basics of Python 29

my_string = "Do not worry, just use double quotes to 'escape'."

print(my_string)

Do not worry, just use double quotes to 'escape'.

One may exchange the role of these two types of quotes:

my_string = 'Do not worry about "double quotes".

print(my_string)

Do not worry about "double quotes"

One shall refer to the Python documentation, when needing to include things
such as carriage returns, backslashes, and Unicode characters. Below are
some more handy and clean operators applied to numbers and strings. You
may try it out and get some experience.

one, two, three = 1, 2, 3 # Assign values to variables.

summation = one + two + three
print('summation=',summation) # printx('summation')

summation= 6

one, two, three = 1, 2, 3.0 # variable type can be mixed!

summation = one + two + three
print('Summation=',summation)

Summation= 6.0

one3 = two3 = three3 = 3 # Assign a same value to variables

print(one3, two3, three3)

3 3 3
30 Machine Learning with Python: Theory and Applications

More handy operations:

hello, world = "Hello,", "world!"

helloworld = hello + " " + world + "!!" # concatenate strings
print(helloworld, ' ', hello + " " + world)
lhw=len(helloworld) # length of the string, counting the space
# and the punctuations.
print('The length of the "helloword" is',lhw)

Hello, world!!! Hello, world!

The length of the "helloword" is 15

You can split the string to a list of strings, each of which is a word.

the_words = helloworld.split(" ") # creates a list of strings

# Similar operations on Lists later
print("Split the words of the string: %s" % the_words)
print('Joined together again with a space as separator:','
'.join(the_words))

Split the words of the string: ['Hello,', 'world!!!']

Joined together again with a space as separator: Hello, world!!!

To find a letter (character) in a string, try this:

my_string = "Hello world!"

print('"o" is right after the',my_string.index("o"),\
'th letter.') # "\" is used to break a line
print('The first letter "l" is right after the', \
my_string.index("l"), 'nd letter.')

"o" is right after the 4 th letter.

The first letter "l" is right after the 2 nd letter.

Do not like the white-spaces between “4” and “th” “2” and “nd”? use string
concatenation:
 Basics of Python 31

print('The position of the letter "o" is right after the ' +

str(my_string.index("o")) + 'th letter.')
# "+" concatenate
print('The 1st letter "l" is right after the ' +
str(my_string.index("l")) + 'nd letter.')

The position of the letter "o" is right after the 4th letter.
The 1st letter "l" is right after the 2nd letter.

You may need to find the frequency of each element in a list.

from collections import Counter # import Counter module.

my_list = ['a','a','b','b','b','c','d','d','d','d','d']
count = Counter(my_list) # Counter object is a dictionary
print(count) # of frequencies of each element in the list
# See also Dictionary later

Counter({'d': 5, 'b': 3, 'a': 2, 'c': 1})

print('The frequency of "b" is', count['b'])

# frequency of an element indexed by its key

The frequency of "b" is 3

Note Python (and many other programming languages) starts counting at 0

instead of 1.
We list below more operations that can be useful.

Conversion between uppercase and lowercase of a string

my_string = "Hello world!"

print(my_string.upper(),my_string.lower(),my_string.title())
# convert to uppercase and lowercase, respectively.

HELLO WORLD! hello world! Hello World!

• Reversion of a string using slicing (also see section on Lists).

32 Machine Learning with Python: Theory and Applications

my_string = "ABCDEFG"
reversed_string = my_string[::-1]
print(reversed_string)

GFEDCBA

The title() function of string class

my_string = "my name is professor g r liu"

new_string = my_string.title()
print(new_string)

My Name Is Professor G R Liu

Use of repetitions

n = 8
my_list = [0]*n
print(my_list)

[0, 0, 0, 0, 0, 0, 0, 0]

my_string = "abcdefg "

print(my_string*2) #concatenated n times and then print out

abcdefg abcdefg

lotsofhellos = "Hello " * 5 #concatenate 5 times

print(lotsofhellos)

Hello Hello Hello Hello Hello

my_list = [1,2,3,4,5]
print(my_list*2) #concatenate 2 times and then print out

[1, 2, 3, 4, 5, 1, 2, 3, 4, 5]
 Basics of Python 33

Length of a given argument using len()

print('The length of "ABCD" is', len('ABCD'))

The length of "ABCD" is 4

print('The length of',my_string,'is', len(my_string))

The length of abcdefg is 8

even_numbers, odd_numbers= [2,4,6,8], [1,3,5,7]

length = len(even_numbers) #get the length using len()

all_numbers = odd_numbers + even_numbers #concatenation
print(all_numbers,' The original length is',
length, '. The new length is',len(all_numbers))

[1, 3, 5, 7, 2, 4, 6, 8] The original length is 4 . The new

length is 8

Sort the elements using sorted()

print(sorted('BACD'),sorted('ABCD',reverse=True))
print(sorted(all_numbers), sorted(all_numbers,reverse=True))

['A', 'B', 'C', 'D'] ['D', 'C', 'B', 'A']

[1, 2, 3, 4, 5, 6, 7, 8] [8, 7, 6, 5, 4, 3, 2, 1]

Multiplying each element in a list by a same number

original_list, n = [1,2,3,4], 2
new_list = [n*x for x in original_list]
# list comprehension for element-wise operations
print(new_list)

[2, 4, 6, 8]
34 Machine Learning with Python: Theory and Applications

Generating index for a list using enumerate()

my_list = ['a', 'b', 'c']

for index, value in enumerate(my_list):
print('{0}: {1}'.format(index+1, value))

1: a
2: b
3: c

for index, value in enumerate(my_list): # generate indices

print(f'{index+1}: {value}') # f-string

1: a
2: b
3: c

Error exception tracks code while avoiding stop execution

a, b = 1,2
try:
print(a/b) # exception raised when b=0
except ZeroDivisionError:
print("division by zero")
else:
print("no exceptions raised")
finally:
print("Regardless of what happened, run this always")

0.5
no exceptions raised
Regardless of what happened, run this always

Get the memory size in bytes

import sys #import sys module

num = "AAA"
print('The memory size is %d'%sys.getsizeof(num),'bytes')
# memory size of string
 Basics of Python 35

The memory size is 52 bytes

num = 21099
print('The memory size is %d'%sys.getsizeof(num),'bytes')
# memory size of integer

The memory size is 28 bytes

num = 21099.0
print('The memory size is %d'%sys.getsizeof(num),'bytes')
# memory size of float

The memory size is 24 bytes

Check whether the string starts with or ends with something

astring = "Hello world!"

print(astring.startswith("Hello"),astring.endswith("asdf"),\
astring.endswith("!"))

True False True

The first one printed True, as the string starts with “Hello”. The second one
printed False, as the string certainly does not end with “asdf”. The third
printed True, as the string ends with “!”. Their boolean values are useful
when creating conditions. More such functions:

my_string="Hello World!"
my_string1="HelloWorld"
my_string2="HELLO WORLD!"
print (my_string.isalnum()) #check if all char are numbers
print (my_string1.isalpha()) #check if all char are alphabetic
print (my_string2.isupper()) #test if string is upper case

False
True
True
36 Machine Learning with Python: Theory and Applications

my_string3, my_string4, my_string5="hello world!"," ", "8888a"

print (my_string3.istitle()) #test if contains title words
print (my_string3.islower()) #test if string is lower case
print (my_string4.isspace()) #test if string is spaces
print (my_string5.isdigit()) #test if string is digits

False
True
True
False

Checking the type and other attributes of variables

n, x, s = 8888,8.0, 'string'
print (type(n), type(x), type(s)) # check the type of an object
print (len(s),len(str(n)),len(str(x)))

<class 'int'> <class 'float'> <class 'str'>

6 4 3

2.3.4 Conversion between types of variables

When one of the variables is a floating point number in an operation with

other integers, the variable becomes a floating point number.

a = 2
print('a=',a, ' type of a:',type(a))
b = 3.0; a = a + b
print('a=',a, ' type of a:',type(a))
print('b=',b, ' type of a:',type(b))

a= 2 type of a: <class 'int'>

a= 5.0 type of a: <class 'float'>
b= 3.0 type of a: <class 'float'>

The type of a variable can be converted to other types.

 Basics of Python 37

n, x, s = 8888,8.5, 'string'
sfn = str(n) #integer to string
print(sfn,type(sfn))
sfx = str(x) #float to string
print(sfx,type(sfx))

8888 <class 'str'>

8.5 <class 'str'>

xfn = float(n) #integer to float

print(xfn,type(xfn))
nfx = int(x) #float to integer
print(nfx,type(nfx))

8888.0 <class 'float'>

8 <class 'int'>

#a = int('Hello') # string to integer: produces ValueError

#a = int('8.5') # string to integer: produces ValueError
a = int('85') # works,'85' is converted to an integer
print(a,type(a))

85 <class 'int'>

a = float('85') # how about this one?

print(a,type(a))

85.0 <class 'float'>

8.0 + float("8.0") #try this

16.0

a = int(False) # check this out

print(a,type(a))

0 <class 'int'>
38 Machine Learning with Python: Theory and Applications

However, operators with mixed numbers and strings are not permitted, and
it triggers a TypeError:

my_mix = my_float + my_string

----------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-67-8c4f7138852e> in <module>
----> 1 my mix = my float + my string

TypeError: unsupported operand type(s) for +: 'float' and 'str'

2.3.5 Variable formatting

Formatting is very useful in printing out variables. Python uses C-style string
formatting to create new, formatted strings. The “%” operator is used to
format a set of variables enclosed in a “tuple”, which is a fixed size list (to
be discussed later). It produces a normal text at a location given by one
of the “argument specifiers” like “%s”, “%d”, “%f”, etc. The best way to
understand this is through examples.

name = "John"
print("Hello, %s! how are you?" % name) # %s is for string
# for two or more argument specifiers, use a tuple:
name, age = "Kevin", 23
print("%s is %d years old." % (name,age)) # %d is for digit
Hello, John! how are you?
Kevin is 23 years old.

print(f"{name} is {age} years old.")

# f-string Python 3.6 or later.
Kevin is 23 years old.

Any object that is not a string (a list for example) can also be formatted
using the %s operator. The %s operator formats the object as a string using
the “repr” method and returns it. For example:

list1,list2,x=[1,2,3],['John','Ian'],21.5 # multi- assignment

print("List1:%s; List2:%s\n x=%s,x=%f,x=%.3f,x=%e" \
% (list1, list2, x, x, x, x))
 Basics of Python 39

List1:[1, 2, 3]; List2:['John', 'Ian']

x=21.5,x=21.500000,x=21.500,x=2.150000e+01

print(f"List1:{list1};List2:{list2};x={x},x={x:.2f}, x={x:.3e}")
# powerful f-string

List1:[1, 2, 3]; List2:['John', 'Kevin']; x=21.5,x=21.50,x=2.150e+01

Often used formatting argument specifiers (if not using f-string):

• %s - String (or any object with a string representation, like numbers).

• %d - Integers.
• %f - Floating point numbers.
• %.f - Floating point numbers with a fixed number-of-digits to the right of
the dot.
• %e - scientific notation: a float multiplied by the specified power of 10.
• %x/%X - Integers in hex representation (lowercase/uppercase).

2.4 Arithmetic Operators

Addition, subtraction, multiplication, and division operators can be used

with numbers.

2.4.1 Addition, subtraction, multiplication, division, and power

+ − ∗ / //(floor division) % (remainder of integer division or modulo) ∗∗

number = 1 + 2 * 3 / 4.0
print(number)

2.5

The modulo (%) operator returns the integer remainder of the division:
dividend % divisor = remainder.

numerator, denominator = 11, 2

floor = numerator // denominator #floor division
print(str(numerator)+'//'+str(denominator)+ '=', floor)
remainder = numerator % denominator
print(str(numerator) +'%'+ str(denominator) +'=', remainder)
print(floor*denominator + remainder)
40 Machine Learning with Python: Theory and Applications

11//2= 5
11%2= 1
11

Using two multiplication symbols makes a power relationship.

squared, cubed = 7 ** 2, 2 ** 3
print('7 ** 2 =', squared, ', and 2 ** 3 =',cubed)

7 ** 2 = 49 , and 2 ** 3 = 8

bwlg_XOR = 7^2
print(bwlg_XOR) # ^ is XOR (bitwise logic gate) operator!

Python allows simple swap operation between two variables.

a, b = 100, 200
print('a=',a,'b=',b)
a, b = b, a # swapping without using a "mid-man"
print('a=',a,'b=',b)

a= 100 b= 200
a= 200 b= 100

2.4.2 Built-in functions

Python provides a number of build-in functions and types that are always
available. For a quick glance, see the following table, or find more details at
https://docs.python.org/3/library/functions.html.
 Basics of Python 41

#help(all) # to find out what a builtin function does

2.5 Boolean Values and Operators

Boolean values are two constant objects: True and False. When used as an
argument to an arithmetic operator, they behave like the integers 0 and 1,
respectively. The built-in function bool() can be used to cast any value to a
Boolean. The definitions are as follows:

print(bool(5),bool(-5),bool(0.2),bool(-0.1),bool(str('a')),
bool(str('0')))
# True True True True True True
print(bool(0),bool(-0.0)) # These are all zero
# False False
print(bool(''),bool([]),bool({}),bool(())) # all empty (0)
# False False False False

True True True True True True

False False
False False False False

bool() returns False, only if the value is zero or the container is empty.
Otherwise, True. Note that str(‘0’) is neither zero nor empty.
42 Machine Learning with Python: Theory and Applications

Boolean operators include “and” and “or”.

print(True and True, False or True, True or True,)

# True True True
print(False and False, False and True)
# False False

True True True

False False

2.6 Lists: A diversified variable type container

We already saw Lists a few times. This Section gives more details. A list is
a collection of variables, and it is very similar to arrays (see Numpy Array
section for more details). A list may contain any type of variables, and as
many variables as one likes. These variables are held in a pair of square
brackets [ ]. Lists can be iterated over for operations when needed. It is
one of the “iterables”. Let us look at the following examples.

2.6.1 List creation, appending, concatenation, and updating

x_list = [] # Use [] to define a placeholder for x_list.

# It is empty but with an address assigned.
print('x_list=',x_list)
print(hex(id(x_list))) # memory address in hexadecimal

x list= []
0x1ae44e9f548

x_list.append(1) # 1 is appended as the 0th member in this list

x_list.append(2) # 2 is appended as the 1st member
x_list.append(3.) # Variable type changed!
print(x_list[0]) # prints 1, the 0th element ...
print(x_list[1]) # prints 2
print(x_list[2]) # prints 3
print(x_list) # print all in the list

3.0
[1, 2, 3.0]
 Basics of Python 43

for x in x_list: # prints out 1,2,3.0 in an iteration

print(x, end=',')

print('\n')
x_list2 = x_list*2
# concatenation of 2 x_list (not element-wise
# multiplication!) this creates an independent new x_list2
print(x_list2)

1,2,3.0,

[1, 2, 3.0, 1, 2, 3.0]

print(id(x_list),id(x_list2)) # addresses are different

1847992120648 1847993139592

id(x_list[1]) # Again, print() function is not needed

# because this is the last line in the cell

1594536160

x_list3 = x_list # assignment is a "pointer" to x_list3

print(x_list,' ',x_list3,)

[1, 2, 3.0] [1, 2, 3.0]

print(id(x_list),id(x_list3)) # They share the same address

1847992120648 1847992120648

x_list4 = x_list.copy() # copy() function creates x_list4

# it is a new independent list
print(x_list,' ',x_list4)

[1, 2, 3.0] [1, 2, 3.0]

44 Machine Learning with Python: Theory and Applications

print(id(x_list),id(x_list4)) # x_list4 has its own address

1847992120648 1847993186760

x_list[0] = 4.0 # Assign the 0th element a new value

print(x_list)

[4.0, 2, 3.0]

print(x_list3,' ',x_list4) # x_list3 is changed with x_list,

# because assignment creates a "pointer". x_list4 is not
# changed, because it was created using copy() function.

[4.0, 2, 3.0] [1, 2, 3.0]

print(x_list2) # Changes to x_list has no affect

[1, 2, 3.0, 1, 2, 3.0]

Creating a list by unpacking a string of digits:

num = 19345678
list_of_digits=list(map(int, str(num))) #list iterable
print(list_of_digits)
list_of_digits=[int(x) for x in str(num)] #list comprehension
print(list_of_digits)

[1, 9, 3, 4, 5, 6, 7, 8]
[1, 9, 3, 4, 5, 6, 7, 8]

2.6.2 Element-wise addition of lists

Element-wise addition of lists needs a little trick. The best ways, including
the use of numpy arrays, will be discussed in the list comprehension section.
Here, we use a primitive method to achieve this.
 Basics of Python 45

list1, list2= [20, 30, 40], [5, 6, 8]

print (list1, ' ', list2, ' ', list1+list2)
print ("Original list 1: " + str(list1))
print ("Original list 2: " + str(list2))

print ('"+" is not addition, but concatenation:',list1+list2)

[20, 30, 40] [5, 6, 8] [20, 30, 40, 5, 6, 8]

Original list 1: [20, 30, 40]
Original list 2: [5, 6, 8]
"+" is not addition, it is concatenation: [20, 30, 40, 5, 6, 8]

# We shall use a for-loop to achieve element-wise addition:

add_list = []
for i in range(0, len(list1)): # for-loop to add up one-by-one!
add_list.append(list1[i] + list2[i])

print ("Element-wise addition of 2 lists is: " + str(add_list))

Element-wise addition of 2 lists: [25, 36, 48]

id(add_list[0]) # check the address of the list

1594536896

id(list1[0])

1594536736

add_list = []
for i1,i2 in zip(list1,list2): # for-loop and zip() to add it up
add_list.append(i1+i2)

print ("The element-wise addition of 2 lists: ",add_list)

The element-wise addition of 2 lists: [25, 36, 48]

46 Machine Learning with Python: Theory and Applications

2.6.3 Slicing strings and lists

Slicing is a useful and efficient operation to manipulate parts of a string, list,

or array (to be discussed later). Our discussion starts from slicing strings,
and then lists.

my_string = "Hello world!"

#123456789TET # conventional order
print('0123456789TE') # ordering in Python
print (my_string)
print('5th=',my_string[4]) # take the 5th character
print('7-11th=',my_string[6:11]) # 7th to 11th

0123456789TE
Hello world!
5th= o
7-11th= world

print('[6:-1]=',my_string[6:-1])
# "-1" for the last slice from the 6th to (last-1)th
print('[:]=',my_string[:]) # all characters in the string
print('[6:]=',my_string[6:]) # slice from 7th to the end
print('[:-1]=',my_string[:-1]) # to (last-1)th

[6:-1]= world
[:]= Hello world!
[6:]= world!
[:-1]= Hello world

my_string = "Hello world!"

#123456789TET # conventional order
print('[3:9:2]=',my_string[3:9:2]) # 4th to 9th step 2
# Syntax:[start:stop:step]
my_string = "Hello world!"

[3:9:2]= l o
 Basics of Python 47

Using a negative step, we can easily reverse a string, as we have seen earlier:

my_string = "Hello world!"

print('string:',my_string)
print('[::-1]=',my_string[::-1]) # all but from the last

string: Hello world!

[::-1]= !dlrow olleH

In summary, if we just have one number in the brackets, the slicing takes
the character at the (number +1)th position. This is because Python counts
from zero. A colon stands for all available. If it is used alone, the slice is the
entire string. If there is a number on its left, the slice is from the number
to the right-end, and vice versa. A negative number means it counts the
number but is from the right-end: −3 means “the 3rd character from the
right-end”. One can also use the step option for skipping.

Note that when accessing a string with an index which does not exist, it
generates an exception of IndexError.

print('[14]=',my_string[14]) # index out of range error

-----------------------------------------------------------
IndexError Traceback (most recent call last)
<ipython-input-102-024f17c69a4f> in <module>
----> 1 print('[14]=',my string[14]) # gives an index out of range error

IndexError: string index out of range

print('[14:]=',my_string[14:]) # This will not give an

# error, but gives nothing: nothing can be sliced

[14:]=

The very similar rules detailed above for strings apply also to a list, by
treating a variable in the list as a character.
48 Machine Learning with Python: Theory and Applications

# Create my_list that contains mixed type variables:

list2=[]
my_list=[0, 1, 2, 3,'4E', 5, 6, 7,[8,8], 9]
# 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 -> indices
# 0, 1, 2, 3, 4, 5, 6, 7, 8, 9 -> Python indices
#-10 -9,-8,-7,-6, -5,-4,-3, -2, -1 -> Python reverse indices

print(my_list[0:10:1]) # [start:end:step] end is inclusive!

print(my_list[:]) # A colon stands for all variable

[0, 1, 2, 3, '4E', 5, 6, 7, [8, 8], 9]

[0, 1, 2, 3, '4E', 5, 6, 7, [8, 8], 9]

print(my_list[0:]) # from (0+1)st to the end

print(my_list[1:]) # from (1+1)th to the end
print(my_list[8:]) # from the (8+1)th to the end
print(my_list[8:9]) # Gives a list in list

[0, 1, 2, 3, '4E', 5, 6, 7, [8, 8], 9]

[1, 2, 3, '4E', 5, 6, 7, [8, 8], 9]
[[8, 8], 9]
[[8, 8]]

print(my_list[:]) # A colon stands for all variable

print(my_list[:3]) # 0,1,2,
print(my_list[:1]) # from 1st to 1st: 0
print(my_list[:0]) # from 1st to 0th: empty []
print(my_list[-1]) # reads out the last: 9
print(my_list[-1:]) # Slices out the last:
print(my_list[::-1]) # reverse the list

[0, 1, 2, 3, '4E', 5, 6, 7, [8, 8], 9]

[0, 1, 2]
[0]
[]
9
[9]
[9, [8, 8], 7, 6, 5, '4E', 3, 2, 1, 0]
 Basics of Python 49

When accessing a list with an index that does not exist, it generates an
exception of IndexError.

#print(my_list[11]) # will give an index out of range error

print(my_list[10:]) # from the (10+1)th to the end: no more

# there, not out of range, but an empty list

[]

2.6.4 Underscore placeholders for lists

nlist = [10, 20, 30, 40, 50, 6.0, '7H'] # Mixed variables
_, _, n3,_,*nn = nlist # when only the 3rd is needed,
# skip one, and then the rest
print ('n3=',n3, 'the rest numbers',*nn)

nlist = [10, 20, 30, 40, 50, 60, 70]

_, _, n3, *nn, nlast = nlist # The 3rd, last and the rest
# in between are needed
print('n3=',n3,', the last=', nlast,', and all the rest numbers',*nn)

n3= 30 the rest numbers 50 6.0 7H

n3= 30 , the last= 70 , and all the rest numbers 40 50 60

2.6.5 Nested list (lists in lists in lists)

nested_list = [[11, 12], ['2B',22], [31, [32,3.2]]]

# A nested list of mixed types of variables
printx('nested_list')
print(len(nested_list)) #number of sub-lists in nested_list

nested list = [[11, 12], ['2B', 22], [31, [32, 3.2]]]

3

print(nested_list[0]) #1st sub-list in the nested_list

print(nested_list[1]) #2nd sub-list in the nested_list
print(nested_list[2]) #3rd sub-list in the nested_list
print(nested_list) #print all for easy viewing
50 Machine Learning with Python: Theory and Applications

[11, 12]
['2B', 22]
[31, [32, 3.2]]
[[11, 12], ['2B', 22], [31, [32, 3.2]]]

print(nested_list[0][0]) #1st element in 1st sub-list

print(nested_list[0][1]) #2nd element in 1st sub-list

11
12

print(nested_list)
print(nested_list[1][0]) # Try this: what would this be?
print(nested_list[2][1]) #?
print(nested_list[2][1][0]) #?

[[11, 12], ['2B', 22], [31, [32, 3.2]]]

2B
[32, 3.2]
32

2.7 Tuples: Value preserved

After the discussion about the List, discussing Tuples becomes straightfor-
ward. This is because they are essentially the same, and the major difference
is as follows:

• A Tuple is usually enclosed with (), but a List is with [].

• A Tuple is immutable, but a List is mutable. This means that tuples
cannot be changed after they are created. Values in Tuples are preserved.

Because a Tuple is immutable, we use it to store data that needs to be

preserved. Thus, its use is very much limited. It is used to store constants
preventing them from being changed. Also, operating on Tuples is faster.
Except these differences, a Tuple behaves like a List. It can be accessed
via index, iterated over, and assigned to other variables. Below are some
examples.
 Basics of Python 51

ttuple = (10, 20, 30, 40, 50, 6.0, '7H') # create a Tuple
gr.printx('ttuple') # print(ttuple)
aa = ttuple[0]
print('aa=',aa)
print(ttuple[1], ' ',ttuple[6],' ',ttuple[-1])

ttuple = (10, 20, 30, 40, 50, 6.0, '7H')

aa= 10
20 7H 7H

for i, data in enumerate(ttuple):

# use enumerate function to get both index and content
if i < 3:
print(i, ':', data)

0 : 10
1 : 20
2 : 30

# ttuple[2] = 300 # this gives an error

The above may be all we need to know about Tuples. We now discuss
another useful data structure in Python.

2.8 Dictionaries: Indexable via keys

A dictionary is a data type similar to a list. It contains paired keys and

values. The key is a string and can be used for indexing. The value can
be any type of object: a string, a number, a list, etc. Because the key and
value are paired, each value stored in a dictionary can be accessed using the
corresponding key. A dictionary does not contain any duplicated keys. The
values may be with duplication.

2.8.1 Assigning data to a dictionary

For example, phone numbers can be assigned to a dictionary in the following

format:
52 Machine Learning with Python: Theory and Applications

phonebook1 = {} # placeholder for a dictionary

phonebook1["Kevin"] = 513476565
phonebook1["Richard"] = 513387234
phonebook1["Jim"] = 513682762
phonebook1["Mark"] = 513387234 # A duplicated value
gr.printx('phonebook1')
print(phonebook1)

phonebook1 = {'Kevin': 513476565, 'Richard': 513387234,

'Jim': 513682762, 'Mark': 513387234}
{'Kevin': 513476565, 'Richard': 513387234, 'Jim': 513682762,
'Mark': 513387234}

A dictionary can also be initialized in the following means:

phonebook2 = {'Joanne': 656477456, 'Yao': 656377243,

'Das': 656662798}
print(phonebook2)

{'Joanne': 656477456, 'Yao': 656377243, 'Das': 656662798}

phonebook0 = {
"John" : [788567837,788347278],
# this is a list, John has 2 phonenumbers
'Mark': 513683222,
'Joanne': 656477456
}
print(phonebook0)

{'John': [788567837, 788347278], 'Mark': 513683222, 'Joanne':

656477456}

2.8.2 Iterating over a dictionary

Like lists, dictionaries can be iterated over. Because keys and values are
recorded in pairs, we may use for-loop to access them.
 Basics of Python 53

for name, number in phonebook1.items():

print("Phone number of %s is %d" % (name, number))

Phone number of Kevin is 513476565

Phone number of Richard is 513387234
Phone number of Jim is 513682762
Phone number of Mark is 513387234

for key, value in phonebook1.items():

print(key, value)

Kevin 513476565
Richard 513387234
Jim 513682762
Mark 513387234

for key in phonebook1.keys():

print(key)

Kevin
Richard
Jim
Mark

for value in phonebook1.values():

print(value)

513476565
513387234
513682762
513387234

2.8.3 Removing a value

To delete a pair of records, we use the build-in function del or pop, by using
the keys.
54 Machine Learning with Python: Theory and Applications

phonebook2 = {'Joanne': 656477456, 'Yao': 656377243,

'Das': 656662798}
del phonebook2["Yao"]
print(phonebook2)
value = phonebook2.pop("Das")
print(phonebook2)
print('value for poped out key',value)

{'Joanne': 656477456, 'Das': 656662798}

{'Joanne': 656477456}
value for poped out key 656662798

2.8.4 Merging two dictionaries

First, use update() method.

phonebook1.update(phonebook2) # phonebook1 is updated

print(phonebook1)

{'Kevin': 513476565, 'Richard': 513387234, 'Jim': 513682762,

'Mark': 513387234, 'Joanne': 656477456}

Now, use a simpler means called double-star. This allows one to create a 3rd
new dictionary that is a combination of two dictionaries, without affecting
the original two.

phonebook3 = {**phonebook1, **phonebook0}

print(phonebook3)

{'Kevin': 513476565, 'Richard': 513387234, 'Jim': 513682762,

'Mark': 513683222,
'Joanne': 656477456, 'John': [788567837, 788347278]}

Duplicated keys (if any) are removed in the new dictionary:

dict_1 = {'Apple': 7, 'Banana': 5}

dict_2 = {'Banana': 3, 'Orange': 4} #'Banana' is in dict_1
 Basics of Python 55

combined_dict = {**dict_1, **dict_2}

#'Banana' in dict_1 will be replaced in the new dictionary
print(combined_dict)
{'Apple': 7, 'Banana': 3, 'Orange': 4}

2.9 Numpy Arrays: Handy for scientific computation

Numpy arrays are similar to Lists, and much easier to work with for scientific
computations. Operations on Numpy arrays are usually much faster for bulky
data.

2.9.1 Lists vs. Numpy arrays

(1) Similarities:

• Both are mutable (the elements there can be added and removed after
the creation. A mutating operation is also called “destructive”, because it
modifies the list/array in place instead of returning a new one).
• Both can be indexed.
• Both can be sliced.

(2) Differences:

• For using arrays, one needs to import Numpy module, but lists are
build-in.
• Array works for element-wise operations, but lists cannot (need some
coding to do that).
• Data types in an array must be the same, but a list can have different types
of data (part of the reason why element-wise operation is not generally
accepted).
• Numpy array can be multi-dimensional.
• Operations with arrays are, in general, much faster than those on lists.
• Storing arrays uses less memory than storing lists.
• Numpy arrays are more convenient to use for mathematical operations
and machine learning algorithms.

2.9.2 Structure of a numpy array

We first brief on the structure of a numpy array in comparison with a list
we discussed earlier. To start the discussion, we import the numpy package.
56 Machine Learning with Python: Theory and Applications

import numpy as np # Import numpy & give it an alias np

#dir(np) #try this (remove #)

x1 = np.array([28, 3, 28, 0]) # a one-dimensional (1D) numpy array

print('x1=',x1) # A numpy array looks like a list.
gr.printx('x1') # This specifies that it is an array.

x1 = [28 3 28 0]
x1 = array([28, 3, 28, 0])

As shown above, a numpy array is “framed” in a pair of square brackets

(same as a list).

x2 = np.array([[51,22.0],[0,0],(18+9j,3.)]) # mixed types

print('x2=',x2) # All become complex-valued

x2= [[51.+0.j 22.+0.j]

[ 0.+0.j 0.+0.j]
[18.+9.j 3.+0.j]]

This is a 2D numpy array. It is framed in a double pair of square brackets.

A list does not have multi-dimensionality, except in the form of nesting: lists
in lists.
We can also create numpy arrays using lists. In the following, we create first
two lists, and then create Numpy arrays using these two lists:

list_w = [57.5, 64.3, 71.6, 68.2] # list, peoples' weights (Kg)

list_h = [1.5, 1.6, 1.7, 1.65] # list heights (m)
print('list_w=',list_w, '; list_h= ',list_h)

list w= [57.5, 64.3, 71.6, 68.2] ; list h= [1.5, 1.6, 1.7, 1.65]

narray_w = np.array(list_w) # convert list to numpy array

narray_h = np.array(list_h)
print('narray_w=',narray_w, '; narray_h= ',narray_h)

narray w= [57.5 64.3 71.6 68.2] ; narray h= [1.5 1.6 1.7 1.65]
 Basics of Python 57

Let us create a function that prints out the information of a given numpy array.

def getArrayInfo(a):
'''Get the information about a given array:
getArrayInfo(array)'''
print('elements of the first axis of the array:',a[0])
print('type:',type(a))
print('number of dimensions, a.ndim:', a.ndim)
print('a.shape:', a.shape)
print('number of elements, a.size:', a.size)
print('a.dtype:', a.dtype)
print('memory address',a.data)

help(getArrayInfo) # may try this

Help on function getArrayInfo in module main :

getArrayInfo(a)
Get the information about a given array: getArrayInfo(array)

We see here that ”’ ”’ useful to provide a simple instruction for the use of a created
function. Let us now use it to get the information for nist w.

getArrayInfo(narray_w)

elements of the first axis of the array: 57.5

type: <class 'numpy.ndarray'>
number of dimensions, a.ndim: 1
a.shape: (4,)
number of elements, a.size: 4
a.dtype: float64
memory address <memory at 0x000001AE7D549408>

We note that narray w is 1D in dimension, and has a shape of (4,) meaning it

has four entries. The shape of a numpy array is given in a tuple.
Slicing also works for a numpy array, in the similar way as for lists. Let us take
a slice for an array.

print(list_w[1:3]) # a slice between the 2nd and 3rd elements

[64.3, 71.6]

print(narray_w[1:3]) # a slice between the 2nd and 3rd elements

[64.3 71.6]
58 Machine Learning with Python: Theory and Applications

Let us now append an element to both the list and the numpy array.

# For lists, we use:

list_w.append(59.8)
print(list_w)
# For numpy array we shall use:
print(np.append(narray_w,59.8))

[57.5, 64.3, 71.6, 68.2, 59.8]

[57.5 64.3 71.6 68.2 59.8]

print(list_w,' ',narray_w)
print(type(list_w),' ',type(narray_w))
print(len(list_w),' ', narray_w.ndim) # Use len() to get the length

[57.5, 64.3, 71.6, 68.2, 59.8] [57.5 64.3 71.6 68.2]

<class 'list'> <class 'numpy.ndarray'>
5 1

nwh = (narray_w,narray_h) # This forms a tuple of np arrays

print(nwh)

(array([57.5, 64.3, 71.6, 68.2]), array([1.5 , 1.6 , 1.7 , 1.65]))

To form a multi-dimensional array, we may use the following (more on this later):

arr = np.array([narray_w,narray_h])
arr

array([[57.5 , 64.3 , 71.6 , 68.2 ],

[ 1.5 , 1.6 , 1.7 , 1.65]])

getArrayInfo(arr)

elements of the first axis of the array: [57.5 64.3 71.6 68.2]
type: <class 'numpy.ndarray'>
number of dimensions, a.ndim: 2
a.shape: (2, 4)
number of elements, a.size: 8
a.dtype: float64
memory address <memory at 0x000001AE44FB12D0>
 Basics of Python 59

We note that arr is of dimension 2, and has a shape of (2, 4), meaning it has
two entries along axis 0 and 4 entries along axis 1. We see again that the shape
of a numpy array is given in a tuple. A multi-dimensional numpy array can be
transported:

arrT = arr.T
print(arrT)
getArrayInfo(arrT) # see the change in shape from (2,4) to (4,2)

[[57.5 1.5 ]
[64.3 1.6 ]
[71.6 1.7 ]
[68.2 1.65]]
elements of the first axis of the array: [57.5 1.5]
type: <class 'numpy.ndarray'>
number of dimensions, a.ndim: 2
a.shape: (4, 2)
number of elements, a.size: 8
a.dtype: float64
memory address <memory at 0x000001AE44FB12D0>

It is seen that the dimension remains 2, but the shape is changed from (2,4) to
(4,2). The value of an entry in a numpy array can be changed.

arr = np.array([[57.5 , 64.3 , 71.6 , 68.2 ],

[ 1.5 , 1.6 , 1.7 , 1.65]])
arrb = arr
printx('arrb')
arr[0,0]= 888.0 # change is done to arr only
printx('arrb')

arrb = array([[57.5 , 64.3 , 71.6 , 68.2 ],

[ 1.5 , 1.6 , 1.7 , 1.65]])
arrb = array([[888. , 64.3 , 71.6 , 68.2 ],
[ 1.5 , 1.6 , 1.7 , 1.65]])

Notice the behavior of the array created via an assignment: changes in an array
will affect the other. This behavior was observed for lists. To create an independent
array, use copy() function.

arr = np.array([[57.5 , 64.3 , 71.6 , 68.2 ],

[ 1.5 , 1.6 , 1.7 , 1.65]])
arrc = arr.copy() #This is expensive. Do it only it is necessary
printx('arrc')
arr[0,0]= 77.0
printx('arr')
printx('arrc')
60 Machine Learning with Python: Theory and Applications

arrc = array([[57.5 , 64.3 , 71.6 , 68.2 ],

[ 1.5 , 1.6 , 1.7 , 1.65]])
arr = array([[77. , 64.3 , 71.6 , 68.2 ],
[ 1.5 , 1.6 , 1.7 , 1.65]])
arrc = array([[57.5 , 64.3 , 71.6 , 68.2 ],
[ 1.5 , 1.6 , 1.7 , 1.65]])

2.9.3 Axis of a numpy array

Axis is an important concept for numpy array operations. A 1D array has axis 0, a
2D array has two axes, 0 and 1, and so on. The definition is given in Fig. 2.1.

Multidimensional numpy array structure and axes:

We can now use an axis to stack up arrays to form new arrays, as follows:

arr=np.stack([narray_w,narray_h],axis=0)
#stack up 1D arrays along axis 0
print(arr)

[[57.5 64.3 71.6 68.2 ]

[ 1.5 1.6 1.7 1.65]]

We can use np.ravel to flatten an array.

print(arr)
rarr=np.ravel(arr)
print(rarr)
getArrayInfo(rarr)

Figure 2.1: Picture modified from that in “Introduction to Numerical Computing with
NumPy”, SciPy 2019 Tutorial, by Alex Chabot-Leclerc.
 Basics of Python 61

[[57.5 64.3 71.6 68.2 ]

[ 1.5 1.6 1.7 1.65]]
[57.5 64.3 71.6 68.2 1.5 1.6 1.7 1.65]
elements of the first axis of the array: 57.5
type: <class 'numpy.ndarray'>
number of dimensions, a.ndim: 1
a.shape: (8)
number of elements, a.size: 8
a.dtype: float64
memory address <memory at 0x000001AE7D5494C8>

It is seen that the dimension is changed from 2 to 1, and the shape is changed
from (2,4) to (8,).
In machine learning computations, we often perform summation of entries of an
array along an axis of the array. This can be done easily using the np.sum function.

print(arr)
print('Column-sum:',np.sum(arr,axis=0),np.sum(arr,axis=0).shape)
print('row-sum:',np.sum(arr,axis=1),np.sum(arr,axis=1).shape)

[[57.5 64.3 71.6 68.2 ]

[ 1.5 1.6 1.7 1.65]]
Column-sum: [59. 65.9 73.3 69.85] (4,)
row-sum: [261.6 6.45] (2,)

Notice that the dimension of the summed array is reduced.

2.9.4 Element-wise computations

Element-wise computation using Numpy arrays is very handy, which is very much
different from that for lists.

print('listwh=',list_w+list_h) # + is a concatenation for lists.

listwh= [57.5, 64.3, 71.6, 68.2, 59.8, 1.5, 1.6, 1.7, 1.65]

print('narraywh=',narray_w+narray_h)
# + is element-wise addition for numpy arrays.

narraywh= [59. 65.9 73.3 69.85]

Let us compute the weights in pounds, using 1kg = 2.20462 lbs.

print(narray_w * 2.20462) # element-wise multiplication

[126.76565 141.757066 157.850792 150.355084]

62 Machine Learning with Python: Theory and Applications

Let us compute the Body Mass Index or BMI using these narrays.

bmi = narray_w / narray_h ** 2 # formula to compute the BMI

print(bmi)

[25.55555556 25.1171875 24.77508651 25.05050505]

This includes element-wise power operation and division as well.

# lbmi = list_w / list_h ** 2 # would this work? Try it!

We discussed element-wise operations for lists earlier, and used special operations
(list comprehension) and special functions such as zip(). The alternative is the
“numpy-way”. This is done by first converting the lists to numpy arrays, then
performing the operations in numpy using these arrays, and finally converting the
results back to a list. When the lists are large in size, this numpy-way can be much
faster, because all these operations can be performed in bulk in numpy, without
element-by-element accessing of the memories.

import numpy as np
list1 = [20, 30, 40, 50, 60]
list2 = [4, 5, 6, 2, 8]
(np.array(list1) + np.array(list2)).tolist()

[24, 35, 46, 52, 68]

The results are the same as those we obtained before using special list element-
wise operations.

2.9.5 Handy ways to generate multi-dimensional arrays

In machine learning and mathematical computations in general, multi-dimensional
arrays are frequently used, because one has to deal with big data frequently. Numpy
supports the necessary functions (tools) to generate, manipulate, and operate multi-
dimensional arrays.

np.arange(2, 8, 0.5, dtype=np.float) # equally spaced values

array([2. , 2.5, 3. , 3.5, 4. , 4.5, 5. , 5.5, 6. , 6.5, 7. , 7.5])

np.linspace(1., 4., 6)
# arrays `with a specified number of elements with equal value spacing

array([1. , 1.6, 2.2, 2.8, 3.4, 4. ])

 Basics of Python 63

a = np.array([1.,2.,3.])
a.fill(9.9) # all entries with the same value
print(a)
[9.9 9.9 9.9]

x = np.empty((3, 4)) # shape (dimension) of (3,4) specified,

# without initializing entries
print(x)
[[2. 2.5 3. 3.5]
[4. 4.5 5. 5.5]
[6. 6.5 7. 7.5]]

x = np.zeros((6,6)) # initialized with zero

x
array([[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0.]])

y = np.ones((2,2))*2 # a 2 by 2 array with 1.0 in all the entries

print(y)
[[2. 2.]
[2. 2.]]

x[3:5,3:5] += y # Assign y to a sliced portion in x

print(x)
[[0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0.]
[0. 0. 0. 2. 2. 0.]
[0. 0. 0. 2. 2. 0.]
[0. 0. 0. 0. 0. 0.]]

2.9.6 Use of external package: MXNet

For computations in machine learning, a number of useful packages/modules/
libraries have been developed for creating large-scale machine learning models.
MXNet is one of those. We will also make use of it in this book. Because it is
an external package, MXNet needs to be installed in your computer system using
pip.
pip install mxnet
64 Machine Learning with Python: Theory and Applications

Note that if an error message like “No module named ‘xyz’ ” is encountered,
which is likely during this course using our codes, one shall perform the installation
of “xyz” module in a similar way, so that all the functions and variables
defined there can be made use of. Note also that there are a huge number of
modules/libraries/packages openly available, and it is not possible to install all of
them. The practical way is installing it only when it is needed. One may encounter
issues in installations, many of which are related to compatibility of versions of the
involved packages. Searching online for help can often resolve these issues, because
the chance is high that someone has already encountered similar issues earlier, and
the huge online community has already provided some solution.
After mxnet module is installed, we import it to our code.

import mxnet as mx
mx.__version__
'1.7.0'

np.set_printoptions(formatter={'float': '{: 0.3f}'.format})

from mxnet import nd # import Mxnet library/package
# see https://gluon.mxnet.io for details
x = nd.empty((3, 4)) # Try also x = nd.ones((3, 4))
print(x)
print(x.size) # print the size (number of elements) of a
# multi-dimensional mxnet or nd.array

[[ 0.000 0.000 0.000 0.000]

[ 0.000 0.000 0.000 0.000]
[ 0.000 0.000 0.000 0.000]]
<NDArray 3x4 @cpu(0)>
12

We often create arrays with randomly sampled values when working with neural
networks. In such cases, we initialize an array using a standard normal distribution
with zero mean and unit variance. For example,

y = nd.random_normal(0, 1, shape=(3, 4)) # 0 mean, variance 1

printx('y')
y
y =
[[-0.681 -0.135 0.377 0.410]
[ 0.571 -2.758 1.076 -0.614]
[ 1.831 -1.147 0.054 -2.507]]
<NDArray 3x4 @cpu(0)>
[[-0.681 -0.135 0.377 0.410]
[ 0.571 -2.758 1.076 -0.614]
 Basics of Python 65

[ 1.831 -1.147 0.054 -2.507]]

<NDArray 3x4 @cpu(0)>

Element-wise operations of addition, multiplication, and exponentiation all work

for multi-dimensional arrays.

x=nd.exp(y)
x
[[ 0.506 0.873 1.458 1.507]
[ 1.771 0.063 2.934 0.541]
[ 6.239 0.318 1.055 0.081]]
<NDArray 3x4 @cpu(0)>

Often used matrix (2D array) transpose can be obtained as follows:

y.T
[[-0.681 0.571 1.831]
[-0.135 -2.758 -1.147]
[ 0.377 1.076 0.054]
[ 0.410 -0.614 -2.507]]
<NDArray 4x3 @cpu(0)>

Now, we can multiply matrices with comparable dimensions. Now we can

multiply matrices with comparable dimensions using dot-product in both numpy
and MXNet:

nd.dot(x, y.T) # x: 3×4; y.T: 4×3

[[ 0.705 -1.476 -3.776]
[ 0.114 3.662 1.970]
[-3.860 3.774 10.910]]
<NDArray 3×3 @cpu(0)>

Note that nd arrays behave differently from np arrays. They do not usually
work together without proper conversion. Therefore, special care is need. When
strange behavior is observed, one may print out the variable to check the array type.
The same is generally true when numpy arrays work with arrays in other external
modules, because the array objects are, in general, different from one module to
another. One may use asnumpy() to convert an nd-array to an np-array, when so
desired. Given below is an example (more on this later):

np.dot(x.asnumpy(), y.T.asnumpy())
# convert nd array to np array and then use numpy np.dot()
array([[ 0.705, -1.476, -3.776],
[ 0.114, 3.662, 1.970],
[-3.860, 3.774, 10.910]], dtype=float32)
66 Machine Learning with Python: Theory and Applications

2.9.7 In-place operations

In machine learning, we frequently deal with bigdata. To avoid expensive and very
complicated moving data operations, we prefer in-place operations. Let us first take
a look at the following computations and the locations of the data:

print('id(y) before operation:', id(y))

y = y + x # x and y must be shape compatible
print('id(y) after operation:', id(y))# location of y changes
id(y) before operation: 1847926752760
id(y) after operation: 1847462282128

For in-place operations, we do this:

print('id(y) before operation:', id(y))

y[:] = x + y
# addition first, put it in a temporary buffer, then copy to y[:]
print('id(y) after operation:', id(y)) # memory of y remains the same

id(y) before operation: 1847462282128

id(y) after operation: 1847462282128

To perform an in-place addition without using even a temporary buffer, we do

this in MXNet:

print('id(y) before operation:', id(y))

print(nd.elemwise_add(x, y, out=y)) # for mxnet nd.arrays
print('id(y) after operation:', id(y))
# memory location un-changed
id(y) before operation: 1847462282128

[[ 0.837 2.485 4.752 4.931]

[ 5.883 -2.568 9.878 1.009]
[ 20.547 -0.194 3.220 -2.263]]
<NDArray 3x4 @cpu(0)>
id(y) after operation: 1847462282128

If we do not plan to reuse x, then the result can be assigned to x itself. We may
do this in MXNet:

print('id(x) before operation:', id(x))

x += y
print(x)
print('id(x) after operation:', id(x))
id(x) before operation: 1847462202168
 Basics of Python 67

[[ 1.343 3.358 6.210 6.438]

[ 7.653 -2.504 12.811 1.550]
[ 26.785 0.124 4.275 -2.182]]
<NDArray 3x4 @cpu(0)>
id(x) after operation: 1847462202168

2.9.8 Slicing from a multi-dimensional array

To read the second and third rows from x, we do this:

print(x)
x[1:3] # read the second and third rows from x
[[ 1.343 3.358 6.210 6.438]
[ 7.653 -2.504 12.811 1.550]
[ 26.785 0.124 4.275 -2.182]]
<NDArray 3x4 @cpu(0)>

[[ 7.653 -2.504 12.811 1.550]

[ 26.785 0.124 4.275 -2.182]]
<NDArray 2x4 @cpu(0)>

x[1:2,1:3] # read the 2nd raw, 2nd to 3rd column from x

[[-2.504 12.811]]
<NDArray 1x2 @cpu(0)>

x[1,2] = 88.0 # change the value at the 2nd raw 3rd column
print(x)
x[1:2,1:3] = 88.0
# change the values from the 2nd raw and 2nd to 3rd column
print(x)
[[ 1.343 3.358 6.210 6.438]
[ 7.653 -2.504 88.000 1.550]
[ 26.785 0.124 4.275 -2.182]]
<NDArray 3x4 @cpu(0)>

[[ 1.343 3.358 6.210 6.438]

[ 7.653 88.000 88.000 1.550]
[ 26.785 0.124 4.275 -2.182]]
<NDArray 3x4 @cpu(0)>

2.9.9 Broadcasting
What would happen if one adds a vector (1D array) y to a matrix (2D array) X? In
Python, this can be done, and is often done in machine learning. Such an operation
is performed using a procedure called “broadcasting”: the low-dimensional array
68 Machine Learning with Python: Theory and Applications

is duplicated along any axis with dimension 1 to match the shape of the high-
dimensional array, and then the desired operation is performed.

import numpy as np
y = np.arange(6) # y has an initial of shape (6), or (1,6)
print('y = ', y,'Shape of y:', y.shape)
x = np.arange(24)
print('x = ', x,'Shape of x:', x.shape)

y = [0 1 2 3 4 5] Shape of y: (6,)
x = [ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
20 21 22 23]
Shape of x: (24,)

X = np.reshape(x,(4,6)) # reshape to (4,6), to match (6,) of y

print(X.shape)
print('X = \n', X,'Shape of X:', X.shape)
print('X + y = \n', X + y) # y's shape expands to (4,6) by copying
# the data along axis 0,then the addition

(4, 6)
X =
[[ 0 1 2 3 4 5]
[ 6 7 8 9 10 11]
[12 13 14 15 16 17]
[18 19 20 21 22 23]] Shape of X: (4, 6)
X + y =
[[ 0 2 4 6 8 10]
[ 6 8 10 12 14 16]
[12 14 16 18 20 22]
[18 20 22 24 26 28]]

print('shape of X is:',X.shape, ' shape of y is:',y.shape)

print(np.dot(X,y))

shape of X is: (4, 6) shape of y is: (6,)

[ 55 145 235 325]

z = np.reshape(X,(2,3,4))
print (z)

[[[ 0 1 2 3]
[ 4 5 6 7]
[ 8 9 10 11]]
 Basics of Python 69

[[12 13 14 15]
[16 17 18 19]
[20 21 22 23]]]

a = np.arange(12).reshape(3,4)
a.fill(100)
a

array([[100, 100, 100, 100],

[100, 100, 100, 100],
[100, 100, 100, 100]])

z + a
array([[[100, 101, 102, 103],
[104, 105, 106, 107],
[108, 109, 110, 111]],

[[112, 113, 114, 115],

[116, 117, 118, 119],
[120, 121, 122, 123]]])

a = np.arange(4)
a.fill(100)
a
array([100, 100, 100, 100])

z + a
array([[[100, 101, 102, 103],
[104, 105, 106, 107],
[108, 109, 110, 111]],

[[112, 113, 114, 115],

[116, 117, 118, 119],
[120, 121, 122, 123]]])

Broadcasting Rules: https://docs.scipy.org/doc/numpy/user/basics.broadcasting.

html.
When operating on two arrays, NumPy compares their shapes element-wise. It
starts with the trailing dimensions and works its way forward. Two dimensions are
compatible when
1. they are equal, or
2. one of them is 1.
70 Machine Learning with Python: Theory and Applications

If these conditions are met, the dimensions are compatible, otherwise not compatible
hence a ValueError. Figure 2.2 shows some examples.
Broadcasting operations:

Figure 2.2: Picture modified from that in “Introduction to Numerical Computing with
NumPy”, SciPy 2019 Tutorial, by Alex Chabot-Leclerc.

2.9.10 Converting between MXNet NDArray and NumPy

Converting between MXNet NDArrays and NumPy arrays is easy. The converted
arrays do not share memory.

import numpy as np
x = np.arange(24).reshape(4,6)
y = np.arange(6)

array([[ 0, 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10, 11],
[12, 13, 14, 15, 16, 17],
[18, 19, 20, 21, 22, 23]])

from mxnet import nd

ndx = nd.array(x)
npa = ndx.asnumpy()
type(npa)

numpy.ndarray
 Basics of Python 71

npa

array([[ 0.000, 1.000, 2.000, 3.000, 4.000, 5.000],

[ 6.000, 7.000, 8.000, 9.000, 10.000, 11.000],
[ 12.000, 13.000, 14.000, 15.000, 16.000, 17.000],
[ 18.000, 19.000, 20.000, 21.000, 22.000, 23.000]],
dtype=float32)

ndy = nd.array(npa)
ndy

[[ 0.000 1.000 2.000 3.000 4.000 5.000]

[ 6.000 7.000 8.000 9.000 10.000 11.000]
[ 12.000 13.000 14.000 15.000 16.000 17.000]
[ 18.000 19.000 20.000 21.000 22.000 23.000]]
<NDArray 4x6 @cpu(0)>

To figure out the detailed differences between the MXNet NDArrays and the
NumPy arrays, one may refer to https://gluon.mxnet.io/chapter01 crashcourse/nd
array.html.

2.9.11 Subsetting in Numpy

Another feature of numpy arrays is that they are easy to subset.

print(bmi[bmi > 25]) # Print only those with BMI above 25

[ 25.556 25.117 25.051]

2.9.12 Numpy and universal functions (ufunc)

NumPy is a useful library for the Python for large-scale numerical computations
including but not limited to machine learning. It supports efficient operations for
bulky data in multi-dimensional arrays (ndarrays) of large size. It also offers a large
collection of high-level mathematical functions to operate on these arrays [1]. In
2005, Travis Oliphant created NumPy by incorporating features of Numarray into
Numeric, which was originally created by Jim Hugunin with several other developers
with extensive modifications. More details can be found at https://en.wikipedia.
org/wiki/NumPy.
The Numpy documentation states that “a universal function (or ufunc for short)
is a function that operates on ndarrays in an element-by-element fashion, supporting
array broadcasting, type casting, and several other standard features. A ufunc is
a “vectorized” wrapper for a function that takes a fixed number of specific inputs
and produces a fixed number of specific outputs. In NumPy, universal functions
are instances of the numpy.ufunc class (to be discussed later). Many of the built-in
functions are implemented in compiled C code. The basic ufuncs operate on scalars,
but there is also a generalized kind for which the basic elements are sub-arrays
72 Machine Learning with Python: Theory and Applications

(vectors, matrices, etc.), and broadcasting is done over other dimensions. One can
also produce custom ufunc instances using the frompyfunc factory function”.
More details of unfuncs can be found at scipy documentation (https://docs.sc
ipy.org/doc/numpy/reference/ufuncs.html). Given below are two of many ufuncs.
exp(x, /[, out, where, casting, order, . . . ]) Calculate the exponential of all elements
in the input array. log(x, /[, out, where, casting, order, . . . ]) Natural logarithm,
element-wise. One may use the following for more details.

# help(np.log) # use this to find out what a ufunc does

2.9.13 Numpy array and vector/matrix

Numpy uses primarily the structure of arrays. The behavior of a numpy array is
similar, and yet quite different from the conventional concepts of vector and matrix
that we learned from linear algebra. It can be quite confusing and even frustrating
during coding and debugging. The author has bothered by such differences quite
frequently in the past. The items below are some of the issues one may need to pay
attention to.

(1) Numpy array is in principle multidimensional

First, we note that numpy array is far more than 1D vector and 2D matrix. By
default, we can have as many as 32D, and that can also be changed. Thus, numpy
array is extremely powerful, and is a data structure works well for complex machine
learning models that use multidimensional dataset and data flows. All the element-
wise operations, broadcasting, handling of flows of large volume data, etc. work very
efficiently. It, however, does not follow precisely the concept of vectors, matrices
that we established in the conventional linear algebra and most frequently used.
This is essentially the root of confusion in many cases, in the author’s opinion.
Understanding the following key points is a good start to mitigate the confusion.

(2) Numpy 1D array vs. vector

Numpy 1D array is similar to the usual vector concept in linear algebra. The
difference is that 1D array does not distinguish row or column vector. It is just
a 1D array with a shape of (n,), where n is the length of the array. It behaves
largely like the row vector, but not quite. For example, transpose() has no effect on
it. This is because what the transpose() function does is to swap two axises of an
array with two or more axises.
The column vector in linear algebra should be treated in numpy as a special case
of 2D array with only one column. One can create an array like the column vector
in linear algebra, by adding an additional axis to the 1D array. See the following
examples.
 Basics of Python 73

a = np.array([1, 2, 3]) # create an 1D array with length 3

print(a, a.T, a.shape, a.T.shape) # its shape will be (3,)

[1 2 3] [1 2 3] (3,) (3,)

As shown, “.T” for transpose() has no effect to the 1D array.

an = a[:,np.newaxis] # add in a newaxis, shape becomes (3, 1)

print(an, an.shape) # it's a column vector, a special 2D array

[[1]
[2]
[3]] (3, 1)

print(an.T, an.T.shape) # .T works, it comes row vector

[[1 2 3]] (1, 3)

The axis added array becomes a 2D array, and the transpose() function works. It
creates a “real” row vector that is a special case of a 2D array in numpy.
The same can also be achieved using the following tricks.

a1 = a.reshape(a.shape+(1,)) # reshape, adds one more dimension

print(a1, a1.shape)
aN = a[:,None] # adds in a new dimension without elements
print(aN, aN.shape)

[[1]
[2]
[3]] (3, 1)
[[1]
[2]
[3]] (3, 1)

Adding an axis can be done at any axis:

a0 = a[np.newaxis,:] # the new axis is added to the 0-axis

print(a0, a0.shape) # shape becomes (1,3), an row vector

[[1 2 3]] (1, 3)

74 Machine Learning with Python: Theory and Applications

print(an+a0) # interesting trick to create a Hankel matrix

[[2 3 4]
[3 4 5]
[4 5 6]]

Once knowing how to create arrays equivalent to the usual row and column
vectors in the conventional linear algebra, we shall be much more comfortable in
debugging codes when encountering strange behavior.
Another often encountered example is when solving linear system equations.
From the conventional linear algebra, we know that the right-hand-side (rhs) should
be a column vector, and the solution also should be a column vector. In using
numpy.linalg.solve() for a linear algebraic equation, we can feed in with a 1D array
as the rhs vector, and it will return a solution that is also a 1D array. Of course,
we can also feed in with a column vector (a 2D array with only one column). In
that case we will get the solution in a 2D array with only one column. We shall see
examples in Section 3.1.11, and many cases in later chapters. These behaviors are
all expected and correct in numpy

(3) Numpy 2D array vs. matrix

Numpy 2D array is similar to the usual matrix in linear algebra largely, but not
quite. For example, the matrix multiplication in linear algebra is often done in
numpy using the dot-product, such as np.dot() or the “@” operator in numpy
(version 3.5 or later). The “*” operator is an element-wise multiplication, as shown
in Section 2.9.4. We will see more examples in Chapter 3 and later chapters. Also,
some operations to a numpy array can results in dimension change. For example,
when mean() is applied to an array, the dimension is reduced. Thus, care is required
on which axis is collapsed.
Note that there is a numpy matrix class (see definition of class in Section 2.14)
that can be used to create matrix objects. These objects behavior quite similar as
the matrix in linear algebra. We try not to use it, because it will be deprecated one
day, as announced in the online document numpy-ref.pdf.
Based on the author’s limited experience, once we are aware of these differences
and behavior subtleties (more discussion later when we encountered one), we can
then pay attention to the behavior subtleties. It is often helpful to frequently
checking the shape of the arrays. This allows us to work more effectively with
powerful numpy arrays, including performing proper linear algebra analysis. At this
moment, it is quite difficult to discuss the theorems in linear algebra using 1D, 2D
or higher dimensional array concepts. In the later chapters, we will still follow the
general rules, principles, and use the terms of vector and matrix in the conventional
linear algebra, because many theoretical findings and arguments are based on it. A
vector refers generally to a 1D numpy array, and a matrix to a 2D numpy array.
When examine the outcomes in numpy codes, we shall notice the behavior subtleties
of the numpy arrays.
 Basics of Python 75

2.10 Sets: No Duplication

A set is a list, but with no duplicate entries.

sentence1 = "His first name is Mark and Mark is his first name"
words1 = sentence1.split() # use split() to form a set of words
print(words1) # whole list of these words is printed
word_set1 = set(words1) # convert to a set
print(word_set1) # print. No duplication

['His', 'first', 'name', 'is', 'Mark', 'and', 'Mark', 'is', 'his',

'first', 'name']
{'and', 'His', 'is', 'first', 'Mark', 'his', 'name'}

Using a set to get rid of duplication is useful for many situations. Many other
useful operations can be applied to sets. For example, we may want to find the
intersection of two sets. To show this, let us create a new list and then a new set.

sentence2 = "Her first name is Jane and Jane is her first name"
words2 = sentence2.split()
print(words2) # whole list of these words is printed.
word_set2 = set(words2) # convert to a set
print(word_set2) # print. No duplication

['Her', 'first', 'name', 'is', 'Jane', 'and', 'Jane', 'is', 'her',

'first', 'name']
{'her', 'Jane', 'Her', 'is', 'first', 'and', 'name'}

2.10.1 Intersection of two sets

print(word_set1.intersection(word_set2)) #intersection of two sets

{'first', 'and', 'name', 'is'}

2.10.2 Difference of two sets

print(word_set1.difference(word_set2))
{'his', 'His', 'Mark'}

This finds words in word set1 that are not in word set2.
We may also want to find the words in word set2 that are not in word set1 as
follows:

print(word_set1.symmetric_difference(word_set2))
{'her', 'His', 'Jane', 'Her', 'Mark', 'his'}
76 Machine Learning with Python: Theory and Applications

Can we do similar operations to lists? Try this.

#print(words1.intersection(words2)) # intersection of two list?

# No. It throws an AttributeError.

2.11 List Comprehensions

We used list comprehension for particular situations a few times. List comprehension
is a very powerful tool for operations on all iterables including lists and numpy
arrays. When used on a list, it creates a new list based on another list, in a single,
readable line.
In the following example, we would like to create a list of integers which specify
the length of each word in a sentence, but only if the word is not “the”. The natural
way to do this is as follows:

sentence="Raises the Sun and comes the light" # Create a string.

words = sentence.split() # Create a list of words using split().
word_lengths = [] # Empty list for the lengths of the words
words_nothe = [] # Empty list for the words that are not "the"
for word in words:
if word != "the":
word_lengths.append(len(word))
words_nothe.append(word)

print(words_nothe, ' ',word_lengths)

['Raises', 'Sun', 'and', 'comes', 'light'] [6, 3, 3, 5, 5]

With a list comprehension, we simply do this:

words = sentence.split()
word_lengths = [len(word) for word in words if word != "the"]
words_nothe = [word for word in words if word != "the"]
print(words_nothe, ' ',word_lengths)
['Raises', 'Sun', 'and', 'comes', 'light'] [6, 3, 3, 5, 5]

The following is even better:

words_nothe2=[] # Empty list to hold the lists of word & length

# for words that are not "the", lists in list
words_nothe2=[[word,len(word)] for word in words if word != "the"]
# one may use () instead of the inner []
print(words_nothe2)
[['Raises', 6], ['Sun', 3], ['and', 3], ['comes', 5], ['light', 5]]
 Basics of Python 77

The following example uses list comprehension to numpy arrays:

import numpy as np
np.set_printoptions(formatter={'float': '{: 0.3f}'.format})
fx = []
x = np.arange(-2, 2, 0.5) # Numpy array with equally spaced values
fx = np.array([-(-xi)**0.5 if xi < 0.0 else xi**0.5 for xi in x])
# Creates a piecewise function (such as activation function).
# The created list is then converted to a numpy array.
print('x =',x); print('fx=',fx)

x = [-2.000 -1.500 -1.000 -0.500 0.000 0.500 1.000 1.500]

fx= [-1.414 -1.225 -1.000 -0.707 0.000 0.707 1.000 1.225]

2.12 Conditions, “if ” Statements, “for” and “while” Loops

In machine learning programming, one frequently uses conditions, “if” statements,

“for” and “while” loops. Boolean variables are used to evaluate conditions. The
boolean values True or False are returned when an expression is compared or
evaluated.

2.12.1 Comparison operators

Comparison operators include ==, <, <=, >, >=, !=. The == operator compares
the values of both the operands and checks for value equality, while “!=” checks for
“inequality”, and others check for inequality and/or equality.

x = 2
print(x == 2) # The comparison result in a boolean value: True
print(x == 3) # The comparison result in a boolean value: False
print(x < 3) # The comparison result in a boolean value: True

True
False
True

x = 2
if x == 2:
print("x equals 2!")
else:
print("x does not equal to 2.")

x equals 2!
78 Machine Learning with Python: Theory and Applications

name, age = "Richard", 18

if name == "Richard" and age == 18: # if-block!
print("He is famous. His name is", name, \
"and he is only",age,"years old.")
He is famous. His name is Richard and he is only 18 years old.

temp_critical = 48.0 #unit: degree Celsius (C).

current_temp = 50.0
if current_temp >= temp_critical:
print("The current temperature is", current_temp, \
"degree C. It is above the critical temperature of",\
temp_critical,"degree C. Actions are needed. ")
else:
print("The current temperature is", current_temp, \
"degree C. It is below the critical temperature of",\
temp_critical, "degree C. No action is needed for now.")

# Notice the use of block, and "\" to break a long line.

The current temperature is 50.0 degree C. It is above the
critical temperature of 48.0 degree C. Actions are needed.

2.12.2 The “in” operator

The “in” operator is used to check if a specified object exists within an iterable
object container, such as a list.
name1, name2 = "Kevin", 'John'
groupA= ["Kevin", "Richard"] # a list with two strings.
if name1 in groupA:
print("The person's name is either",groupA[0], "or", groupA[1])

if name2 in groupA:
print("The person's name is either",groupA[0], "or", groupA[1])
else:
print(name2, "is not found in group A")

The person's name is either Kevin or Richard

John is not found in group A

2.12.3 The “is” operator

Unlike the double equals operator “==”, the “is” operator does not check the values
of the operands. It checks whether both the operands refer to the same object or
not. For example,
 Basics of Python 79

x, y = ['a','b'], ['a','b']
z = y # makes z pointing to y

print(x == y,' x=',x,'y=',y,hex(id(x)),hex(id(y))) # Prints True,

# because the values in x and y are equal

print(x is y,'x=',x,'y=',y,hex(id(x)),hex(id(y))) # Prints False,

# because x and y with different ID

print(y is z,' y=',y,'z=',z,hex(id(y)),hex(id(z))) # True, y is z

True x= ['a', 'b'] y= ['a', 'b'] 0x1ae25570188 0x1ae25570208

False x= ['a', 'b'] y= ['a', 'b'] 0x1ae25570188 0x1ae25570208
True y= ['a', 'b'] z= ['a', 'b'] 0x1ae25570208 0x1ae25570208

y[1]='x' # change the 2nd element

print('After change one value in y')
print(x == y,'x=',x,'y=',y,hex(id(x)),hex(id(y))) # Prints False,
# their values are no longer equal

print(x is y,'x=',x,'y=',y,hex(id(x)),hex(id(y)))
# False, x is # NOT y

print(y is z,' y=',y,'z=',z,hex(id(y)),hex(id(z))) # True, y is z:

# they change together!

After change one value in y

False x= ['a', 'b'] y= ['a', 'x'] 0x1ae25570188 0x1ae25570208
False x= ['a', 'b'] y= ['a', 'x'] 0x1ae25570188 0x1ae25570208
True y= ['a', 'x'] z= ['a', 'x'] 0x1ae25570208 0x1ae25570208

y.append(x)
print('After change one value in y')
print(x == y,'x=',x,'y=',y,hex(id(x)),hex(id(y))) # Prints False,
# their values are no longer equal
print(x is y,'x=',x,'y=',y,hex(id(x)),hex(id(y)))
# False, x is # NOT y
print(y is z,' y=',y,'z=',z,hex(id(y)),hex(id(z))) # True, y is z:
# they change together!

After change one value in y

False x= ['a', 'b'] y= ['a', 'x', ['a', 'b'], ['a', 'b']]
0x1ae25570188 0x1ae25570208
80 Machine Learning with Python: Theory and Applications

False x= ['a', 'b'] y= ['a', 'x', ['a', 'b'], ['a', 'b']]

0x1ae25570188 0x1ae25570208
True y= ['a', 'x', ['a', 'b'], ['a', 'b']] z= ['a', 'x', ['a',
'b'], ['a', 'b']] 0x1ae25570208 0x1ae25570208

2.12.4 The ‘not’ operator

Using “not” before a boolean expression inverts the value of the expression.

print(not False) # Prints out True

print(not False == False) # Prints out False

True
False

The “not” operator can also be used with “is” and “in”: “is not”,“not in”:

x, y = [1,2,3], [1,2,3]
z = y # makes z pointing to same object as y
print(x == y,' x=',x,'y=',y,hex(id(x)),hex(id(y))) # True,
# because the values in x and y are equal

print(x is not y,'x=',x,'y=',y,hex(id(x)),hex(id(y))) # False, the

# values in x in NOT y
print(y is z,' y=',y,'z=',z,hex(id(y)),hex(id(z))) # True, y is z

True x= [1, 2, 3] y= [1, 2, 3] 0x1ae25570348 0x1ae255702c8

True x= [1, 2, 3] y= [1, 2, 3] 0x1ae25570348 0x1ae255702c8
True y= [1, 2, 3] z= [1, 2, 3] 0x1ae255702c8 0x1ae255702c8

2.12.5 The “if ” statements

As any other language, the “if” statement is often used for programming, we have
already seen some earlier. The following are more examples for using Python’s
conditions in the “if” statement with code blocks:

temp_good = 22.0 # unit: degree Celsius (C).

temp_now1 = 48.0 # Readers may change this value and try
statement1=(temp_now1<=(temp_good+10.0))&(temp_now1>=(temp_good-10.0))
# comfortable
statement2 = temp_now1 < (temp_good-10.0) # cold
statement3 = temp_now1 > (temp_good+10.0) # hot
if statement1 is True: # do not forget ":"
print("Okay, it is",temp_now1,"degree C. Comfortable. Let's go.")
pass # do something
elif statement2 is True: # do not forget ":"
print("No, it is",temp_now1,"degree C. Too cold. We cannot go.")
pass # do something else
 Basics of Python 81

elif statement3 is True:

print("No, it",temp_now1,"degree C. Too hot. We cannot go.")
pass # do something else
else:
print("Let's check the temperature.") # do another thing
pass
No, it 48.0 degree C. Too hot. We cannot go.

Note that there is no limit in Python on how many blocks one can use in an if
statement.

2.12.6 The “for” loops

There are two types of loops in Python, for and while. We have used “for” loops
already. We discuss it here in more detail together with the while loops. A for loop
iterates over a given iterable sequence. The starting and stopping points and the
step-size are controlled by the sequence, as shown in the following example. It is
sequence controlled.

primes = [2, 3, 5] # define a list (that is iterable)

for prime in primes: # do not forget ":"
print(prime, end=' ')
2 3 5

For loops can iterate over a sequence of numbers using the “range” function. The
range() is a built-in function which returns a range object: a sequence of integers.
It generates integer numbers between the given start-integer and the stop-integer.
It is generally used to iterate over with for loops.

print('Member in range(10) are')

for n in range(10): # Syntax: range(start, stop[, step])
print(n, end=',')

print('\nMembers in range(3, 8) are')

for n in range(3, 8): # Default step is 1
print(n, end=',')

Member in range(10) are

0,1,2,3,4,5,6,7,8,9,
Members in range(3, 8) are
3,4,5,6,7,

print('Members in range(3, 10, 2) are')

for n in range(-3, 10, 2): # starting from a negative value
print(n, end=',')
82 Machine Learning with Python: Theory and Applications

print('\nMembers in range(10, -3, -2) are')

for n in range(10, -3, -2): # reverse range
print(n, end=',')

Members in range(3, 10, 2) are

-3,-1,1,3,5,7,9,
Members in range(10, -3, -2) are
10,8,6,4,2,0,-2,

For a given list of five numbers, let us display each element that doubled, using a
for loop and range() function.

print("Double the numbers in a list, using for-loop and range()")

given_list = [10, 30, 40, 50]
for i in range(len(given_list)):
print("Index["+str(i)+"]","Value in the given list is",
given_list[i],", and its double is", given_list[i]*2)

Double the numbers in a list, using for-loop and range()

Index[0] Value in the given list is 10 , and its double is 20
Index[1] Value in the given list is 30 , and its double is 60
Index[2] Value in the given list is 40 , and its double is 80
Index[3] Value in the given list is 50 , and its double is 100

The range() function returns an immutable sequence object of integers, so it is

possible to convert a range() output to a list, using the list class. For example,

print("Converting python range() output to a list")

list_rng = list(range(-10,12,2))
print(list_rng)
Converting python range() output to a list
[-10, -8, -6, -4, -2, 0, 2, 4, 6, 8, 10]

2.12.7 The “while” loops

The “while” loops repeat as long as a certain boolean condition is met. The condition
controls the operations. For example,

#To print out 0,1,2,3,4,5,6,7,8,9,

count = 0
while count < 10: # do not forget ":"
print(count, end=',')
count += 1 # This is the same as count = count + 1
0,1,2,3,4,5,6,7,8,9,
 Basics of Python 83

“break” and “continue” statements: break is used to exit a “for” loop or a “while”
loop, whereas continue is used to skip the current block, and return to the “for” or
“while” statement.

# print all digits limited by a given number.

count = 0
while True:
print(count,end=',')
count += 1
if count >= 10:
break
print('\n')
# Prints out only even numbers: 0,2,4,6,8,
for n in range(10): # for-loop to control the range
if n % 2 != 0: # Check condition, control what to print
continue # continue to count on (in the for-loop)
print(n,end=',')

0,1,2,3,4,5,6,7,8,9,

0,2,4,6,8,

When the loop condition fails, then the code part in “else” is executed. If break
statement is executed inside the for loop, then the “else” part is skipped. Note that
the “else” part is executed even if there is a continue statement before it.

# Prints out 0,1,2,3,4 and then it prints "count value reached 5"
count=0
nlimit = 5
while(count<nlimit):
print(count, end=',')
count +=1
else:
print("count value reached %d" %(nlimit))

# Prints out 1,2,3,4

for i in range(1, 10):
if(i%5 == 0): # modulo division (%)
break
print(i, end=',')
else:
print("This is not printed because for-loop is terminated due\
to the break but not due to fail in condition")

0,1,2,3,4,count value reached 5

1,2,3,4,
84 Machine Learning with Python: Theory and Applications

2.12.8 Ternary conditionals

The following 4-line code

condition=True
if condition:
x=1
else:
x=0
print(x)
1

can be written in one line, with ternary conditionals:

condition=True
x=1 if condition else 0
print(x)
1

It is simple, readable and dry. Thus conditionals are frequently used in Python.

2.13 Functions (Methods)

Functions offer a convenient way to divide a code into useful blocks that can be called
unlimited times when needed. This can drastically reduce the repeat of codes, and
make codes clean, more readable, and easy to maintain. In addition, functions are
good ways to define interfaces for easy sharing of the codes among programmers.

2.13.1 Block structure for function definition

A function has a “block” structure. Block keywords include those we have already
seen, such as “if”, “for”, and “while”. Functions in Python are defined using the
block keyword “def”, followed by a function name that is also the block’s name.
The function is called using the function name followed by (), which brackets the
arguments, if any. Try this simplest function:

def print_hello(): # do not forget ":"

print("Hello, welcome to this simple function!")
print_hello()
Hello, welcome to this simple function!

2.13.2 Function with arguments

In the simple case given above, no argument is required. Functions are often created
with required arguments that are variables passed from the caller to the function.
 Basics of Python 85

def greeting_student(username, greeting):

print(f"Hello, {username}, greetings! Wish you {greeting}")
greeting_student("Kevin", "a fun journey in using functions!")

Hello, Kevin, greetings! Wish you a fun journey in using functions!

Functions may be created with return values to the caller, using the keyword
“return”.

def sum_two_numbers(a, b):

return a + b

x, y = 2.0, 8.0
apb = sum_two_numbers(x, y)
print(f'{x} + {y} = {apb}') # print('%f + %f = %f'%(x,y,apb))
x, y = 20, 80
apb = sum_two_numbers(x, y)
print(f'{x} + {y} = {apb}, perfect!')
#print('%d + %d = %d: perfect!'%(x,y,apb))
print(f'{x + y}, perfect!') #f-string allow the use of operations

2.0 + 8.0 = 10.0

20 + 80 = 100, perfect!
100, perfect!

Variable scope: LEGB rule (local, enclosing, global, buildings) defines the sequence
for Python to search for a variable. The searching terminates when the variable is
found.

def sum_two_numbers(a, b):

a += 1
print('Inside the function, a=',a)
print('Inside the function x=',x)
return a + b

x, y = 2.0, 8.0
print('Before the function is called,x=',x,'y=',y)
apb = sum_two_numbers(x, y)
print('After the function is called, %f + %f = %f'%(x,y,apb))

Before the function is called,x= 2.0 y= 8.0

Inside the function, a= 3.0
Inside the function x= 2.0
After the function is called, 2.000000 + 8.000000 = 11.000000
86 Machine Learning with Python: Theory and Applications

2.13.3 Lambda functions (Anonymous functions)

Lambda function is a single-line function. It may be the simplest form of functions
and the most useful. The linear functions can be defined simply as

f = lambda x, k, b: k * x + b # do not forget ":"

print(f(1,1,1),f(2,2,2),f(1,2,3))

2 6 5

The quadratic functions can be defined as

f = lambda x, a, b, c: a*x**2 + b*x +c

print(f(2,2,-4,6))

Often Lambda functions are used together with the normal functions, especially for
returning a value where a single-line function comes in handy.

def func2nd_order(a,b,c):
return lambda x: a*x**2 + b*x +c

f2 = func2nd_order(2,-4,6)
print('f2(2)=',f2(2), 'or', func2nd_order(2,-4,6)(2))

f2(2)= 6 or 6

2.14 Classes and Objects

A class is a single entity that encapsulates variables and functions (or methods).
A class is essentially a template to create class objects. One can create unlimited
number of class objects with it, and each class object gets the structure, vari-
ables/attributes and-functions (or methods) from the class. References used in this
section includes the following:

• https://www.python-course.eu/python3 class and instance attributes.php.

• https://realpython.com/instance-class-and-static-methods-demystified/.

2.14.1 A simplest class

class C: # Define a lass named C

''' A simplest possible class named "C" '''
ca = "class attribute" # an attribute defined in the class

help(C) # check out what has been created

 Basics of Python 87

Help on class C in module main :

class C(builtins.object)
| A simplest possible class named "C"
|
| Data descriptors defined here:
|
| dict
| dictionary for instance variables (if defined)
|
| weakref
| list of weak references to the object (if defined)
|
| --------------------------------------------------------------
| Data and other attributes defined here:
|
| ca = 'class attribute'

This example shows how a Class is structured, comments given in ”’ ”’ in the class
definition are used to convey a message to the user, and how an attribute can be
created in the class. We can now use it to observe some behavior of a class attribute
and instance attributes.

i1 = C() # an instance i1 (a class object) is created.

i2 = C() # an instance i2 (another class object) is created.
print('i1.ca=',i1.ca) # prints: 'class attribute!'
print('i2.ca=',i2.ca) # prints: 'class attribute!'
print('C.ca =',C.ca) # prints: 'class attribute!'
#print('ca=',ca) # NameError:'ca' is not defined

i1.ca= class attribute

i2.ca= class attribute
C.ca = class attribute

C.ca = "This is a changed class attribute 'ca'"

# Changing the class attribute, via class
print('C.ca =',C.ca)
print('i1.ca=',i1.ca)
print('i2.ca=',i2.ca)

C.ca = This is a changed class attribute 'ca'

i1.ca= This is a changed class attribute 'ca'
i2.ca= This is a changed class attribute 'ca'
88 Machine Learning with Python: Theory and Applications

Note the values in the instances are also changed.

i1.ca = "This is a changed instance attribute 'ca'"

# Changing an instance attribute
print('i1.ca=',i1.ca) # Changed
print('C.ca =',C.ca) # Unchanged
print('i2.ca=',i2.ca) # Unchanged
i1.ca= This is a changed instance attribute 'ca'
C.ca = This is a changed class attribute 'ca'
i2.ca= This is a changed class attribute 'ca'

The change is only effective for the instance attribute that is changed.

C.ca = "The 2nd changed class attribute 'ca'"

# Changing the class attribute, via class
print('C.ca =',C.ca) # should change accordingly
print('i1.ca=',i1.ca) # Will not change! No longer follow C
# because it made a change after creation
print('i2.ca=',i2.ca) # should change according to class C
# because it has not been changed creation
C.ca = The 2nd changed the class attribute 'ca'
i1.ca= This is a changed instance attribute 'ca'
i2.ca= The 2nd changed class attribute 'ca'

Class attributes and object instance attributes are stored in separate dictionaries:

C.__dict__
mappingproxy({' module ': ' main ',
' doc ': ' A simplest possible class named "C" ',
'ca': "The 2nd changed the class attribute 'ca'",
' dict ': <attribute ' dict ' of 'C' objects>,
' weakref ': <attribute ' weakref ' of 'C' objects>})

i1.__dict__

{'ca': "This is a changed instance attribute 'ca'"}

It is clear that a dictionary has been created when a change is made at the instance
level, departing from the class level.

i2.__dict__

{}

No dictionary has been created, because no change is made at the instance level. It
stays with the class.
 Basics of Python 89

i2.ca = "Make now a change at the instance y to the attribute 'ca'"

i2.__dict__

{'ca': "Make now a change at the instance y to the attribute 'ca'"}

A dictionary has now been created, because the change is made at the instance
level. It departed from the class level. Any future change at the class level to this
attribute will no longer affect the attribute at this instance level.

2.14.2 A class for scientific computation

Let us look at an example for simple scientific computations. We first create a class
called Circle to compute the area of a circle for given radius. The following is the
code:

class Circle:
''' Class "Circle": Compute the area of a circle '''
pi = 3.14159 #class attribute of constants used class-wide
# and class specific
def __init__(self, radius): # a special constructor
# __init__ is executed when the class is called.
# it is used to initiate a class. For this simple
# task we need only one variable: radius.
# "self" is used to reserve an argument place for an
# instance (to-be-created) itself to pass along.
# It is used for all the functions in a class.
self.radius = radius # This allows the instance accessing
# the variable: radius.
def circle_area(self): # function computes circle area
return self.pi * self.radius **2 # pi gets in there via
# the object instance itself

#help(Circle) # check out what has been created

We can now use this class to perform computations.

r = 10
c10 = Circle(r) # create an instance c10. c10 is now passed to
# self inside the class definition
# 10 is passed to the self.radius
print('Circle.pi before re-assignment',Circle.pi)
# access pi via class
90 Machine Learning with Python: Theory and Applications

print('Radius=',c10.radius)
# access via object c10.radius is the self.radius in __init__
print('c10.pi before re-assignment', c10.pi)
# The class attribute is accessed via instance attribute
Circle.pi before re-assignment 3.14159
Radius= 10
c10.pi before re-assignment 3.14159

c10.pi=3.14 # this will change the constant for instance c10

# It will not change the class-wide pi value
print('c10.pi after re-assignment via c10.pi:',c10.pi)
print('Circle.pi after re-assignment via c10.pi:',Circle.pi)
print('circle_area of c10 =',c10.circle_area())
print('circle_area of Circle100=',Circle(100).circle_area())
c10.pi after re-assignment via c10.pi: 3.14
Circle.pi after re-assignment via c10.pi: 3.14159
circle area of c10 = 314.0
circle area of Circle100= 31415.899999999998

It is seen that the Class Circle works well. Let us now create a subclass.

2.14.3 Subclass (class inheritance)

Subclasses can often be used, to take the advantages of the inheritance feature in
Python. This allows us to create new classes by fully making use of an existing class
for its entire structure (attributes and functions), without affecting the ongoing
use of the existing class. It is thus also useful in upgrading the existing programs
because of the reduced duplications.
Assume that the Circle code created above has already been distributed and
used by many. We now decide to create another class to compute the area of a
partial Circle, given the value of the portion of the circle. We can now create a
subclass for this purpose, called P circle, without affecting the use of the already
distributed Circle. The following is the code:

class P_circle(Circle):
# Subclass P_circle referring the base (or parent)
# Circle in (). This establishes the inheritance
''' Subclass "P_circle" based on Class "Cirle": Compute the\
area of a circle portion '''
def __init__(self,radius,portion):
# with 3 attributes: self, radius, and portion.
super().__init__(radius) # This brings in base attributes
# from the base class Circle.
self.portion = portion # Subclass attribute.
 Basics of Python 91

def pcircle_area(self):
# define a function to compute the area of a partial circle
return self.portion*self.circle_area() # New function in
# subclass. The base class Circle is used here.

#help(P_circle) # check out what has been created

Readers may remove “#” in above cell, execute it, and take a moment to read
through this information, and see how the subclass is structured, its connection
with the base class, how self is been used to prepare for connections with the future
objects to be assigned, and what are the attributes and functions that are newly
created and inherited from the base class.

pc10 = P_circle(10.,0.5) # create an object instance using the

# subclass, with argument radius=10 and portion =50%
pc10.pi # we have the same attribute from the base class

3.14159

pc10.radius # we have the same attribute from the base class

10.0

pc10.pcircle_area() # area of a 50% partial circle is computed

157.0795

Let us now make a change to constant pi via subclass instance.

pc10.pi = 3.14
pc10.pi
3.14

It changed. Let us check pi via the base class c10.

c10.pi
3.14

It remains unchanged. Actions to the subclass are not affecting the base class.

2.15 Modules

We now touch upon modules. A module is a piece of Python file that has a specific
functionality. For example, when writing a finite element program, we may write
one module for creating the stiffness matrix and another for solving the system
92 Machine Learning with Python: Theory and Applications

equations. Each module is a separated Python file, which can be written and edited
independently. This helps a lot in organizing and maintaining large programs.
A module in Python is a Python file with .py extension. The file name will be
the module name. Such a module can have a set of functions, classes, and variables
defined. In a module, one can import other modules in the procedure mentioned in
the beginning of this chapter.

2.16 Generation of Plots

Python is also very powerful in generating plots. This is done by import modules
that are openly available. Here, we shall present a simple demo plot of scattered
circles.
First, we import the modules needed.

import numpy as np
import matplotlib.pyplot as plt
# matplotlib.pyplot is a plot function in the matplotlib module
%matplotlib inline
# to have the plot generated inside the notebook
# Otherwise, it will be generated in a new window

We now generate sample data, and then plot 80 randomly generated circles.

n=80
x=np.random.rand(n) # Coordinates randomly generated
y=np.random.rand(n)
colors=np.random.rand(n)
areas=np.pi*(18*np.random.rand(n))**2
# circle radiuses from 0~20 randomly generated
plt.scatter(x,y,s=areas,c=colors,alpha=0.8)
plt.show()

Figure 2.3: Randomly generated circular areas filled with different colors.
 Basics of Python 93

# Plost a curve
x = range(1000)
y = [i ** 2 for i in x]
plt.plot(x,y)
plt.show();

Figure 2.4: Curve for a quadratic function.

x = np.linspace(0, 1, 1000)**1.5
plt.hist(x);

Figure 2.5: An example of a histogram.

2.17 Code Performance Assessment

Performance assessment on a code can be done in two ways. Typical example codes
are given below. Readers may make use of these codes for accessing computational
performance to his/her codes.
94 Machine Learning with Python: Theory and Applications

import time # import time module

import numpy as np

g=list(range(10_000_000))
#print(g)
q=np.array(g,'float64')
#print(q)
start = time.process_time()
sg=sum(g)
t_elapsed = (time.process_time() - start)
print(sg,'Elapsed time=',t_elapsed)

start = time.process_time()
sq=np.sum(q)
t_elapsed = (time.process_time() - start)
print(sq,t_elapsed)

49999995000000 Elapsed time= 0.28125

49999995000000.0 0.03125

%%timeit #use timeit

sg=sum(g)

329 ms ± 70.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

2.18 Summary

With the basic knowledge of Python and its related modules, and the essential
techniques for coding, we are now ready to learn to code for computations and
machine learning techniques. In this process one can also gradually improve his/her
Python coding skills.

Reference
[1] C.R. Harris, K.J. Millman and J.v.d.W. Stéfan et al. Array programming with NumPy,
Nature, 585(7825), 357–362, Sep 2020. http://dx.doi.org/10.1038/s41586-020-2649-2.
 Chapter 3

Basic Mathematical Computations

This chapter discusses typical scientific computations using codes in Python.

We will focus on how numerical data are represented mathematically, how
they are structured or organized, stored, manipulated, operated upon, or
computed in effective manners. Subtleties in those operations in Python will
be examined. At the end of this chapter, techniques for initial treatment
of datasets will also be discussed. The reference materials used in the
chapter include Numpy documentation, Scipy documentation, https://gluon.
mxnet.io/, and https://jupyter.org/. Our discussion shall start with some
basic linear algebra operations on data with a structure of vector, matrix
and tensor.

3.1 Linear Algebra

Linear algebra is most essential for any computations that involve big data,
such as in machine learning. We plan to briefly review basic linear algebraic
operations, through the use of Python programming, using modules that
have already been developed in the Python community at large. We shall go
through the basic concepts, the mathematical notation, data structure, and
the computation procedure. Readers feel free to skim or skip this chapter
if you are already confident in the basic linear algebra computations. Our
discussion will start from the data structure. First, we import necessary
modules and functions.

import sys # import "sys" module

sys.path.append('grbin') # current/relative directory,
# like ..\\..\\code
# Or absolute folder like 'F:\\xxx\\...\\code'

95
96 Machine Learning with Python: Theory and Applications

import grcodes as gr # grcodes module is placed in the

# folder above
from grcodes import printx # import a particular function
import numpy as np # Import Numpy package

We will also use the MXNet package. If not done yet, we shall have
MXNet installed using: pip install mxnet.
After the installation, we import the MXNet.

from mxnet import nd # Import NDArray and give it a alias nd

3.1.1 Scalar numbers

As discussed in Chapter 2, scalar numerical numbers for mathematical

computations have three major types: integer, real number, and complex
number. In Python, such a numerical number is assigned with a unique name
and a given address. It is accessed by calling its name, can be updated, and
used as an argument of a properly defined function (built-in, defined in a code
or in an imported module). In Python programming, all these operations
on a number are straightforward. The most often encountered problem in
computation is that the number may get out of the bound of the digits of
the computer (over- or under-flow), and may become illegal as an argument
for a function. Also, we assume that the numbers generated in Python can
cover the entire real space with the machine accuracy of limit.

3.1.2 Vectors

A vector refers to an object that has more than one component stacked along
one dimension. It can give a physical meaning depending on the type of the
physics problem. For example, a force vector in three-dimensional (3D) space
has three components, each representing the projection of the force vector
onto one of the three axes of the space. It is also known as the degrees of
freedom (DoF). Higher dimensions are encountered in discretized numerical
models, such as the finite element method or FEM (e.g., [1]). In the FEM,
the solid structure is discretized in space with elements and nodes. The DoFs
of such a discretized model depend on the number of nodes, which can be
very large, often in the order of millions. Therefore, we form vectors with
millions of components or entries. In machine learning models, the features
and labels can be written in vector form.
 Basic Mathematical Computations 97

In this chapter, we will not discuss much about the physical problems.
Instead, we discuss general aspects of a vector in abstract, and issues on
computational operations that we may perform to the vector for a given
coordinate system. The DoF for a vector is also referred as the length. A
vector of length p has a shape denoted in Python as (p’).

p = 15 # Length of the vector

x = nd.arange(p) # Create a vector that is an nd-array
# using nd.arange() function
gr.printx('x') # x is now a vector with n components
printx('x') # Use the printx function directly
print(x.shape)
x =
[ 0. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14.]
<NDArray 15 @cpu(0)>
x =
[ 0. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14.]
<NDArray 15 @cpu(0)>
(15,)
Note that in mathematics, the notation of vector is often presented as a
column vector. In Python, is shows as a raw vector.

print(x.T) # Transpose of x
print(x.T.shape)
[ 0. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14.]
<NDArray 15 @cpu(0)>
(15,)
In mathematics, the transpose of a (column) vector shall become a
raw vector, and vice versa. In MXNet NDArray, The transformed vector
is stored in the same place but marked as transpose, so that operations,
such as multiplication, can be performed properly. This saves operations for
physically copying and moving the data, improving efficiency. To confirm
this, we just print out the addresses, as follows.

print(id(x))
print(id(x.T)) # Transpose of x

2212173184080
2212173184080
98 Machine Learning with Python: Theory and Applications

It is clear that in NDArray, a vector is a vector. One does not distinguish

whether it is a column or row vector. Its transport is just a marker on it and
only one set of data is stored. This is an example that MXNet pays special
attention to not moving the data in memory unnecessarily. We do not seem
to observe this behavior in numpy arrays, as shown below.

xnp = np.arange(15) # an Numpy array is generated

print(id(xnp), xnp.shape)
print(id(xnp.T),xnp.T.shape) # address is changed
2212173195344 (15,)
2212173195424 (15,)

The shape of the numpy array is unchanged (still structured as row-vector

or vector), but its transpose is given a separated address.

3.1.3 Matrices
A matrix refers to an object that has more than one dimensions of data
structure, in which each dimension has more than one component. It can be
viewed as stacks of vectors of some length. It again can have a physical mean-
ing depending on the type of physics problem. For example, the stiffness (and
mass) matrix created based on a discretized numerical model for a solid struc-
ture, such as the FEM, has a two-dimensional (2D) structure. In each of the
dimensions, the number of components is the same as that of the DoFs. The
whole matrix is a kind of spatially distributed “stiffness” of the structure [1].
In machine learning models, the input data points in the feature space, and
learning parameters in the hypothesis space may be written in a matrix form.
Again, we will not discuss much about the physical problem here.
Instead, we discuss general aspects of a matrix in abstract, and issues on
computational operations that we may perform to the matrix. Such an
abstract matrix can be presented in arrays of multi-dimensions known as
shape in Python defined in Chapter 2.

A = x.reshape((3, 5)) # Create a 2D matrix by reshaping

# a 1D array
print("A=", A, "\n A.T=", A.T)
A=
[[ 0. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]
 Basic Mathematical Computations 99

[10. 11. 12. 13. 14.]]

<NDArray 3x5 @cpu(0)>

A.T=
[[ 0. 5. 10.]
[ 1. 6. 11.]
[ 2. 7. 12.]
[ 3. 8. 13.]
[ 4. 9. 14.]]
<NDArray 5x3 @cpu(0)>

print(A.shape, A.T.shape)

(3, 5) (5, 3)

The shape or the dimension of the matrix A is 3 by 5, and that of its

transpose becomes (5, 3).
As discussed in Chapter 2 in detail, each of the components (or entries)
in the vector or matrix can be accessed by indexing or slicing.

print('A[1, 2] = ', A[1, 2]) # via index

print('row 2', A[2, :]) # slice the 3rd row (count from 0)
print('column 1', A[:, 1]) # slice the 2nd column
A[1, 2] =
[7.]
<NDArray 1 @cpu(0)>
row 2
[10. 11. 12. 13. 14.]
<NDArray 5 @cpu(0)>
column 1
[ 1. 6. 11.]
<NDArray 3 @cpu(0)>

The matrix can also be transposed.

A.T # A transpose, shape becomes 5 by 3

print(id(A), id(A.T))
2212173182960 2213649918832

It is found that the transposed matrix has its own address also in MXNet.
100 Machine Learning with Python: Theory and Applications

3.1.4 Tensors
The term of “Tensor” requires some clarification. In mathematics or physics,
tensor has a specific well-defined meaning. It refers to a structured data
(a single number, a vector, or a multi-dimensional matrix) that obeys
a certain tensor transformation rule under coordinate transformations.
Therefore, tensor is a very special group of structured data or object, and
not all matrices can be called a tensor. In fact, most of them are not. So long
as the tensor transformation rules are obeyed, it can be classed in different
orders. Scalars are 0th order tensors, vectors are 1st-order tensors, and 2D
matrices are 2nd-order tensors, and so on.
Having said that, in the machine learning (ML) community, however, any
matrix with dimension higher than 2 is called a tensor. It can be viewed as
stacks of matrices of same shape. This ML tensor seems to carry a meaning
of big data that needs to be structured in high dimensions. The ML tensor
is now used as a general way of representing an array with an arbitrary
dimension or arbitrary number of axes. The use of ML tensors becomes
more convenient when dealing with, for example, images that can have 3D
data structures, with axes corresponding to the height, width, and the three
color (RGB) channels. In numpy, a tensor is a multidimensional array.
Because there is usually no such a coordinate transformation performed in
machine learning, there will be no possible confusion caused in our discussion
in this book. From now onwards, we will call the ML tensor a tensor,
with the understanding that it may not obey the real-tensor transformation
rules and we do not perform such a transformation in machine learning
programming.
We now use nd.range() and then the reshape() to create a 3D nd-array.

X = nd.arange(24).reshape((2, 3, 4))
print('X.shape =', X.shape)
print('X =', X)

X.shape = (2, 3, 4)
X =
[[[ 0. 1. 2. 3.]
[ 4. 5. 6. 7.]
[ 8. 9. 10. 11.]]
 Basic Mathematical Computations 101

[[12. 13. 14. 15.]

[16. 17. 18. 19.]
[20. 21. 22. 23.]]]
<NDArray 2x3x4 @cpu(0)>
Element-wise operations are applicable to all tensors.

A = nd.arange(8).reshape((2, 4))
B = nd.ones_like(A)*8 # get shape of A, assign uniform entries
print('A =', A, '\n B =', B)
print('A + B =', A + B, '\n A * B =', A * B)

A =
[[0. 1. 2. 3.]
[4. 5. 6. 7.]]
<NDArray 2x4 @cpu(0)>
B =
[[8. 8. 8. 8.]
[8. 8. 8. 8.]]
<NDArray 2x4 @cpu(0)>
A + B =
[[ 8. 9. 10. 11.]
[12. 13. 14. 15.]]
<NDArray 2x4 @cpu(0)>
A * B =
[[ 0. 8. 16. 24.]
[32. 40. 48. 56.]]
<NDArray 2x4 @cpu(0)>

3.1.5 Sum and mean of a tensor

x = nd.arange(5)
print(x)

[0. 1. 2. 3. 4.]
<NDArray 5 @cpu(0)>

nd.sum(x) # summation of all entries/elements: 0+1+2+3+4=10

102 Machine Learning with Python: Theory and Applications

[10.]
<NDArray 1 @cpu(0)>

X = nd.ones(15).reshape(3,5)
X
[[1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1.]]
<NDArray 3x5 @cpu(0)>

nd.sum(X) # summation of all entries/elements

[15.]
<NDArray 1 @cpu(0)>

print(nd.mean(X), nd.sum(X)/X.size) # same as nd.mean()

[1.]
<NDArray 1 @cpu(0)>
[1.]
<NDArray 1 @cpu(0)>

3.1.6 Dot-product of two vectors

Dot-product may be one of the most, if not the most, widely used operations
in scientific computation including machine learning. We discussed about it
briefly in Section 2.9. Here, we shall discuss more on its use for vectors that
may have different data structure resulting in some subtleties, as mentioned
in Section 2.9.13
Given two vectors a and b, their dot-product is often written in linear
algebra as a b or a · b. Essentially, it is just a sum of their element-wise
products, which results in a scalar. This implies that the shape of a and b
must be compatible: both have the same length. Let us see some examples.

a = nd.arange(5)
b = nd.ones_like(a) * 2 #This ensures the compatibility
print(f"a={a},a.shape={a.shape} \nb={b},b.shape={b.shape}")
printx('nd.dot(a, b)')
printx('nd.dot(a, b).shape')
print(f"np.dot(a, b)={np.dot(a.asnumpy(), b.asnumpy())}")
 Basic Mathematical Computations 103

print(f"np.dot(a.T,b)={np.dot(a.asnumpy().T, b.asnumpy())}")
printx('np.dot(a.asnumpy(), b.asnumpy()).shape')
a=
[0. 1. 2. 3. 4.]
<NDArray 5 @cpu(0)>,a.shape=(5,)
b=
[2. 2. 2. 2. 2.]
<NDArray 5 @cpu(0)>,b.shape=(5,)
nd.dot(a, b) =
[20.]
<NDArray 1 @cpu(0)>
nd.dot(a, b).shape = (1,)
np.dot(a, b)=20.0
np.dot(a.T,b)=20.0
np.dot(a.asnumpy(), b.asnumpy()).shape = ()

Note that the transpose() to the vector will have not effect, because
transpose() in numpy swaps the axises of a 2D array. An numpy 1D array
has a shape of (n,) and hence no action can be taken. A numpy 1D array is
not treated as a matrix, as discussed in Section 2.9.13. When b is a column
vector, a special case of 2D array, it then has two axises like a matrix. The
dot-product a · b is same as the matrix-product ab (where a is defined as
a row vector and b a column vector), in terms of the scalar value resulted.
Thus, in our formulation, we do not distinguish them mathematically, and
we often use the following equality.
a · b = ab (3.1)
In computations in numpy, however, there are some subtleties. The dot-
product of two (1D array) vectors gives a scalar, and the dot-product of a
(1D array) vector with a column vector gives an 1D array with the same
scalar as the sole element. In NDArray, such subtleties are not observed.
Readers may examine the following code carefully to make sense of this.

b_c = b.reshape(-1, 1) # convert a 1D array to a column vector

print(b_c, 'b_c.shape=', b_c.shape)
printx('np.dot(a.asnumpy(),b.asnumpy())') # np dot-product (scalar)
printx('np.dot(a.asnumpy(),b_c.asnumpy())') # np dot-product (array)
print(a.asnumpy()@b_c.asnumpy()) # np matrix-product (array)
printx('nd.dot(a, b)') # nd dot-product (array)
printx('nd.dot(a, b_c)') # nd dot-product (array)
printx('nd.dot(a, b_c).shape')
104 Machine Learning with Python: Theory and Applications

[[2.]
[2.]
[2.]
[2.]
[2.]]
<NDArray 5x1 @cpu(0)> b_c.shape= (5, 1)
np.dot(a.asnumpy(),b.asnumpy()) = 20.0
np.dot(a.asnumpy(),b_c.asnumpy()) = array([20.], dtype=float32)
[20.]
nd.dot(a, b) =
[20.]
<NDArray 1 @cpu(0)>
nd.dot(a, b_c) =
[20.]
<NDArray 1 @cpu(0)>
nd.dot(a, b_c).shape = (1,)

As seen, all these give the same scalar value but in a different data
structure.
The dot-product of two column vectors (special matrices) a and b of
equal length is written linear algebra as a b or b a, which gives the same
scalar (but in a 2D array or a matrix with only one element).

a_c = a.reshape(-1, 1) # convert a 1D array to a column vector

print(a_c),
printx('np.dot(a_c.asnumpy().T,b_c.asnumpy())') # scalar in 2D array
printx('np.dot(b_c.asnumpy().T,a_c.asnumpy())')
printx('np.dot(b_c.asnumpy().T,a_c.asnumpy()).shape')

[[0.]
[1.]
[2.]
[3.]
[4.]]
<NDArray 5x1 @cpu(0)>
np.dot(a_c.asnumpy().T,b_c.asnumpy()) = array([[20.]], dtype=float32)
np.dot(b_c.asnumpy().T,a_c.asnumpy()) = array([[20.]], dtype=float32)
np.dot(b_c.asnumpy().T,a_c.asnumpy()).shape = (1, 1)

To access the scalar value in a 2D array of shape (1, 1), simply use.

print(np.dot(a_c.asnumpy().T,b_c.asnumpy())[0][0])

20.0
 Basic Mathematical Computations 105

One may use flatten() to make a column vector back to a row vector.

printx('np.dot(a_c.asnumpy().flatten(),b_c.asnumpy())')
printx('np.dot(a_c.asnumpy().flatten(),b_c.asnumpy()).shape')
np.dot(a_c.asnumpy().flatten(),b_c.asnumpy()) = array([20.], dtype=float32)
np.dot(a_c.asnumpy().flatten(),b_c.asnumpy()).shape = (1,)

To access the scalar value in a 1D array of shape (1,), simply use,

printx('np.dot(a_c.asnumpy().flatten(),b_c.asnumpy())[0]')
np.dot(a_c.asnumpy().flatten(),b_c.asnumpy())[0] = 20.0

One may use ravel() to convert a multidimensional array to a 1D array (in

this case no copy of the array is made).

printx('np.dot(a_c.asnumpy().T,b_c.asnumpy()).ravel()')
printx('np.dot(a_c.asnumpy().T,b_c.asnumpy()).ravel()[0]')
np.dot(a_c.asnumpy().T,b_c.asnumpy()).ravel() = array([20.], dtype=float32)
np.dot(a_c.asnumpy().T,b_c.asnumpy()).ravel()[0] = 20.0

3.1.7 Outer product of two vectors

Given two vectors a and b, the outer product a⊗b becomes a matrix; in
its ij position, the element is ai bj . Thus, the shapes of a and b are always
compatible.

a = np.arange(3)
b = np.ones(5) * 2
print(a, b)
print('np.outer=\n',np.outer(a, b))
[0 1 2] [2. 2. 2. 2. 2.]
np.outer=
[[0. 0. 0. 0. 0.]
[2. 2. 2. 2. 2.]
[4. 4. 4. 4. 4.]]

A matrix (2D array) is created using two 1D arrays of arbitrary lengths,

with the help of the np.outer() function. One can achieve the same results
using the @ operator, but a needs to be a column vector with shape (n, 1)
106 Machine Learning with Python: Theory and Applications

and b needs to be a row vector with shape (1, m). Readers may try this
as an exercise. Note that although we may get the same results, using the
built-in np.outer() is recommended, because it is usually much faster, and
does not need additional operations. This recommendation applies to all the
other similar situations.

3.1.8 Matrix-vector product

When the dimensionality (shape) is compatible (or made compatible via
“broadcasting”), one can obtain a matrix-vector product, using the np.dot()
function.
A35 = nd.arange(15).reshape(3,5)
d5 = nd.ones(A35.shape[1]) # get the 2nd element of shape A35
print(A35,A35.shape, d5, d5.shape)
f = nd.dot(A35, d5) # shape compatible:[3,5]X[5]->vector
# of length 3
print(f,f.shape)

[[ 0. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]
[10. 11. 12. 13. 14.]]
<NDArray 3x5 @cpu(0)> (3, 5)
[1. 1. 1. 1. 1.]
<NDArray 5 @cpu(0)> (5,)

[10. 35. 60.]

<NDArray 3 @cpu(0)> (3,)

#nd.dot(d5,A35) # shape error: [5]X[3,5] not compatible

d3 = nd.ones(A35.shape[0])
nd.dot(d3,A35) # this works:[3]X[3,5]-> vector of length 5
[15. 18. 21. 24. 27.]
<NDArray 5 @cpu(0)>

3.1.9 Matrix-matrix multiplication

Further, dot-product can also be used for matrix-matrix multiplications, as
long as the shape is compatible.
 Basic Mathematical Computations 107

A23 = nd.ones(shape=(2, 3))

B35 = nd.ones(shape=(3, 5))
print(A23,B35)
nd.dot(A23, B35) # [2,3]X[3,5]: shape compatible ->[2,5]
[[1. 1. 1.]
[1. 1. 1.]]
<NDArray 2x3 @cpu(0)>
[[1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1.]]
<NDArray 3x5 @cpu(0)>

[[3. 3. 3. 3. 3.]
[3. 3. 3. 3. 3.]]
<NDArray 2x5 @cpu(0)>

#nd.dot(B35,A23) # this would give a dot shape error: [4,5]X[3,4]

In numpy, we have similar ways to perform matrix-matrix dot-product

operation.

import numpy as np
print('np.dot():\n',np.dot(A23.asnumpy(),B35.asnumpy()))
print('numpy @ operator:\n',A23.asnumpy() @ B35.asnumpy())
np.dot():
[[3. 3. 3. 3. 3.]
[3. 3. 3. 3. 3.]]
numpy @ operator:
[[3. 3. 3. 3. 3.]
[3. 3. 3. 3. 3.]]

Care is needed in dealing with matrix-vector, matrix-matrix operations,

because of the requirement of dimension compatibility. It is important to
always check the consistency of the dimensions of all the terms in the same
equation. Readers may need to struggle a while to get used to it. Operations
between one-dimensional arrays (row vectors) in Python are rather simpler,
because there is only on dimension to check and we usually use only the
dot-product (inner product).
108 Machine Learning with Python: Theory and Applications

3.1.10 Norms
A norm is used in Python to measure how “big” a vector or matrix is. There
are types of norm measures, but they all produce a non-negative value for
the measure. The most often used or default one is the L2-norm. It is the
square root of the sum of the squared elements in the vector or matrix or
tensor. For matrices, it is often called the Frobenius norm. The computation
is by calling a norm() function:

d = nd.ones(9) # create an array (vector)

print(d,nd.sum(d))
printx('nd.norm(d)') # use nd.norm()
print(np.linalg.norm(d.asnumpy())) # use numpy linalg.norm()
[1. 1. 1. 1. 1. 1. 1. 1. 1.]
<NDArray 9 @cpu(0)>
[9.]
<NDArray 1 @cpu(0)>
nd.norm(d) =
[3.]
<NDArray 1 @cpu(0)>
3.0

Note the difference that the nd.norm returns an NDArray, but

np.linalg.norm gives a float.

#help(nd.norm) # when wondering, use this

print(nd.norm(A23),np.sqrt(6*1**2)) # nd.norm() for matrix,

# default L2
print(np.linalg.norm(A23.asnumpy())) # numpy linalg.norm()
# for matrix
[2.4494898]
<NDArray 1 @cpu(0)> 2.449489742783178
2.4494898

print(nd.norm(A23,ord=2, axis=1)) # nd.norm() for matrix,

# along axis 1
print(np.linalg.norm(A23.asnumpy(),ord=2, axis=1))
# numpy linalg.norm() for matrix
 Basic Mathematical Computations 109

[1.7320508 1.7320508]
<NDArray 2 @cpu(0)>
[1.7320508 1.7320508]

The L1-norm of a vector is the sum of the absolute value of the elements
in a vector. The L1-norm of a matrix can be defined as the maximum of
L1-norm of column vectors of the matrix. For computing the L1-norms of a
vector, we use the following:

printx('nd.sum(nd.abs(d))') # use nd.norm() for vector

printx('nd.norm(d,1)')
nd.sum(nd.abs(d)) =
[9.]
<NDArray 1 @cpu(0)>
nd.norm(d,1) =
[9.]
<NDArray 1 @cpu(0)>

print(np.sum(np.abs(d.asnumpy()))) # numpy for vector

print(np.linalg.norm(d.asnumpy(),1)) # np.linalg.norm() for vector
9.0
9.0

print(np.linalg.norm(A23.asnumpy(),1)) # np.linalg.norm()
# for matrix
2.0

3.1.11 Solving algebraic system equations

We can use numpy.linalg.solve to solve a set of linear algebraic system
equations given as follows:

KD = F (3.2)

where K is a given positive-definite (PD) square matrix (stiffness matrix in

FEM, for example), F is a given vector (nodal forces in FEM), and D is the
unknown vector (the nodal displacements). The K matrix is a symmetric-
positive-definite (SPD) for well-posed FEM models [1].
110 Machine Learning with Python: Theory and Applications

import numpy as np
np.set_printoptions(formatter={'float': '{: 0.3f}'.format})

K = np.array([[1.5, 1.], [1.5, 2.]]) # A square matrix

print('K:',K)
F = np.array([1, 1]) # one may try F = np.array([[2], [1]])
print('F:',F)
D = np.linalg.solve(K,F)
print('D:',D)

K: [[ 1.500 1.000]
[ 1.500 2.000]]
F: [1 1]
D: [ 0.667 0.000]

If careful, one should see that the input F is a 1D numpy array, and the
result is also a 1D array, which is not the convention of linear algebra, as
discussed in Section 2.9.13. One can also purposely define a column vector
(a 2D array with only on column) following the convention of linear algebra,
and get the solution. In this case, however, the returned solution will be the
same, but is in a column vector. Readers may try this as an exercise.
Note that solving linear algebraic system equations numerically can be
very time consuming and expensive, especially for large systems. With
the development of computer hardware and software in the past decades,
numerical algorithms for solving linear algebraic systems are well developed.
The most effective solver for very large systems uses iterative methods.
It converts the problem of solving algebraic equations to a minimization
problem with a properly defined error residual function as a cost or
loss function. A gradient-based algorithm, such as the conjugate gradient
methods and Krylov methods, can be used to minimize the residual error.
These methods are essentially the same as those used in machine learning.
The numpy.linalg.solve uses routines from the widely used efficient Linear
Algebra PACKage or LAPACK.
For a matrix that is not square, we shall use the least-square solvers
for a best solution with minimized least-square error. The function in
numpy is numpy.linalg.lstsq(). We will see examples later when discussing
interpolations.
 Basic Mathematical Computations 111

3.1.12 Matrix inversion

Readers may notice that the solution to Eq. (3.2) can be written as

D = K−1 F (3.3)

where K−1 is the inverse of K. Therefore, if one can compute K−1 , the
solution is simply a matrix-vector product. Indeed, for small systems, this
approach does work, and is used by many. We now use numpy.linalg.inv()
to compute the inverse of a matrix.

from numpy.linalg import inv # import the inv() function

Kinv = inv(K)
print(Kinv)
[[ 1.333 -0.667]
[-1.000 1.000]]

print(np.allclose(np.dot(K, Kinv), np.eye(2)))

print(np.allclose(np.dot(Kinv, K), np.eye(2)))
True
True

The solution to Eq. (3.2) is obtained as follows:

D = np.dot(Kinv,F)
print('D:',D)
D: [ 0.667 0.000]

which is the same as the one obtained earlier.

Multiple matrices can be inverted at once.

a = np.array([[[1., 2.], [3., 4.]], [[1, 3], [3, 5]]])

print(a)
inv(a)
[[[ 1.000 2.000]
[ 3.000 4.000]]

[[ 1.000 3.000]
[ 3.000 5.000]]]
112 Machine Learning with Python: Theory and Applications

array([[[-2.000, 1.000],
[ 1.500, -0.500]],

[[-1.250, 0.750],
[ 0.750, -0.250]]])

a = np.array([[1.5, 1.], [1.5, 1.]])

# A singular matrix, because its
print(a) # two columns are parallel
#ainv = inv(a) # This would give a Singular matrix error
[[ 1.500 1.000]
[ 1.500 1.000]]

Note that computing numerically the inverse of a matrix of large size is

much more expensive, compared to solving the algebraic system equations.
Therefore, one would like to avoid the computation of inverse matrix. In
many cases, we can change the matrix inversion problem to a set of problems
of solving algebraic equations.
In machine learning computations, one may encounter matrices that
are singular, which do not have an inverse, leading to breakdown in
computations. More often, the matrices are nearly singular, which may allow
the computation to continue but can lead to serious error, showing as some
unexpected, strange behavior. When such behavior is observed, the chance
is high that the system matrix may be “bad-conditioned”, and one should
check for possible errors in the data or the formulation procedure that may
lead to the nearly singular system matrix. If the problem is rooted in the data
itself, one may need to clean up the data or check for data error. After this
is exhausted, one may resort to mathematical means, one of which is the use
of singular value decomposition (SVD) to get the best possible information
from the data. See later in this chapter on SVD.
The key point we would like to make here is that the most important
factor that controls whether we can solve an algebraic system equation for
quality solution is the property (or characteristics or condition) of the system
matrix. Therefore, studying the property of a matrix is of fundamental
importance.
Eigenvalues (if they exist) and their corresponding eigenvectors are the
characteristics of a matrix.
 Basic Mathematical Computations 113

3.1.13 Eigenvalue decomposition of a matrix

A diagonalizable matrix can have an eigenvalue decomposition, which gives

a set of eigenvalues and the corresponding eigenvectors. The original matrix
can be decomposed into a diagonal matrix with eigenvalues at the diagonal
and a matrix consisting of eigenvectors. In particular, for real symmetric
matrices, eigenvalue decomposition is useful, and the computation can be
fast, because the eigenvalues are all real and the eigenvectors can be made
real and orthonormal. Consider a real symmetric square matrix A that is
positive-definite (PD). It has the following eigenvalue decomposition:

A = VΛV (3.4)

where  stands for transpose, Λ is a diagonal matrix with eigenvalues at the

diagonal, and matrix V is formed with eigenvectors corresponding to the
eigenvalues. It is an orthonormal matrix, because

VV = I (3.5)

which implies also that the inverse of V equals its transpose.

V−1 = V (3.6)

In addition, once a matrix is decomposed, computing the inverse of it

becomes trivial simply by matrix multiplications. To see this, we start from
the definition of inverse of a matrix, AA−1 = I, and use Eqs. (3.4) and (3.6),
which leads to
A−1 = VΛ−1 V (3.7)

Because the matrix is PD, its inverse exists and all the eigenvalues are
nonzero (positive-definite). The inverse of the diagonal matrix Λ is simply
the same diagonal matrix with diagonal terms replaced by the reciprocals of
the eigenvalues.
Eigenvalue decomposition can be viewed as a special case of SVD. For
general matrices that we often encounter in machine learning, the SVD is
more widely used for matrix decomposition and will be discussed later in
this chapter, because it exists for all matrices.
In this section, let us see an example of how the eigenvalues and the
corresponding eigenvectors can be computed in Numpy.
114 Machine Learning with Python: Theory and Applications

import numpy as np
from numpy import linalg as lg # import linalg module
A = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]]) # Identity matrix
print('A=',A)
e, v = lg.eig(A)
print('Eigenvalues:',e)
print('Eigenvectors:\n',v)

A= [[1 0 0]
[0 1 0]
[0 0 1]]
Eigenvalues: [ 1.000 1.000 1.000]
Eigenvectors:

[[ 1.000 0.000 0.000]

[ 0.000 1.000 0.000]
[ 0.000 0.000 1.000]]

It is clearly seen that the identity matrix has three eigenvalues all of 1, and
their corresponding eigenvectors are three linearly independent unit vectors.
Let us look at a more general symmetric matrix.

A = np.array([[1, 0.2, 0], [0.2, 1, 0.5], [0, 0.5, 1]])

# Symmetric A
print('A:\n',A)
e, v = lg.eig(A)
print('Eigenvalues:',e, '\n Eigenvectors:\n',v)

A:
[[ 1.000 0.200 0.000]
[ 0.200 1.000 0.500]
[ 0.000 0.500 1.000]]
Eigenvalues: [ 1.539 1.000 0.461]
Eigenvectors:
[[-0.263 0.928 0.263]
[-0.707 -0.000 -0.707]
[-0.657 -0.371 0.657]]

We obtain three eigenvalues and their corresponding eigenvectors, and

they are all real numbers. These eigenvectors are orthonormal. To see this,
lets compute:
 Basic Mathematical Computations 115

print(np.dot(v,v.T))
[[ 1.000 0.000 0.000]
[ 0.000 1.000 -0.000]
[ 0.000 -0.000 1.000]]

This means that all these three eigenvectors are orthogonal with each
other, and the dot-product between them becomes a unit. We are now ready
to recover the original matrix A, using these eigenvalues and eigenvectors.

print(A)
lamd = np.eye(3)*e
A_recovered = v@lamd@v.T
print(A_recovered)
[[ 1.000 0.200 0.000]
[ 0.200 1.000 0.500]
[ 0.000 0.500 1.000]]
[[ 1.000 0.200 0.000]
[ 0.200 1.000 0.500]
[ 0.000 0.500 1.000]]

It is clear that matrix A is recovered with the machine error. This

means that the information in matrix A is fully kept in its eigenvalues and
eigenvectors.
Next, we use the eigenvalues and eigenvectors to compute its inverse,
using Eq. (3.7).

lamd_inv = np.eye(3)/e
A_inv = v@lamd_inv@v.T
print(A_inv)
[[ 1.056 -0.282 0.141]
[-0.282 1.408 -0.704]
[ 0.141 -0.704 1.352]]

which is the same as that obtained directly using the numpy.linalg.inv()

function:

from numpy.linalg import inv

print(inv(A))
116 Machine Learning with Python: Theory and Applications

[[ 1.056 -0.282 0.141]

[-0.282 1.408 -0.704]
[ 0.141 -0.704 1.352]]

Let us now compute the eigenvalues and eigenvectors of an asymmetric

matrix.
A = np.array([[1,-0.2, 0], [0.1, 1,-0.5], [0, 0.3, 1]])
# Symmetric A
print('A:\n',A)
e, v = lg.eig(A)
print('Eigenvalues:',e, '\n Eigenvectors:\n',v)
A:
[[ 1.000 -0.200 0.000]
[ 0.100 1.000 -0.500]
[ 0.000 0.300 1.000]]
Eigenvalues: [1.+0.j 1.+0.4123j 1.-0.4123j]
Eigenvectors:
[[-0.9806+0.j 0. +0.3651j 0. -0.3651j]
[ 0. +0.j 0.7528+0.j 0.7528-0.j ]
[-0.1961+0.j -0. -0.5477j -0. +0.5477j]]

We now see one real eigenvalue, but the other two eigenvalues are
complex valued. These two complex eigenvalues are conjugates of each other.
Similar observations are made for the eigenvectors. We conclude that a real
asymmetric matrix can have complex eigenvalues and eigenvectors. Complex
valued matrices shall in general have complex eigenvalues and eigenvectors.
A special class of complex valued matrices called the Hermitian matrix (self-
adjoint matrix) has real eigenvalues. This example shows that the complex
space is geometrically closed, but the real space is not. An n by n real matrix
should have n eigenvalues (and eigenvectors), but they may not be all in the
real space. Some of them get into the complex space (that with the real space
as its special case).

3.1.14 Condition number of a matrix

The condition number of a matrix is a measure of the “level” of singularity.
There are a number of options of norms for computing the condition number.
It is always larger than or equal to 1 for any measure. This implies that the
best condition of a matrix is 1, which is the condition number of any unit
matrix that has no (the lowest) singularity. Any other matrix shall have some
 Basic Mathematical Computations 117

level of singularity. The condition number for a matrix with the highest level
of singularity is infinite. Let us see some examples.

from numpy import linalg as lg

A = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]])
# A unit matrix
# Clearly, it has 3 eigenvalues of all 1.0
print(A, '\n Condition number of A=',lg.cond(A))
# or lg.cond(A,2)
# Option 2-> L2 more measure
[[1 0 0]
[0 1 0]
[0 0 1]]
Condition number of A= 1.0

Because matrix A is a unit matrix, we got a condition number of 1, as

expected. Another example is as follows:

A = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 0]])

# A singular matrix
# It has 2 eigenvalues of all 1.0, and 1 eigenvalue of 0
print(A, '\n Condition number of A=',lg.cond(A))
[[1 0 0]
[0 1 0]
[0 0 0]]
Condition number of A= inf

Because matrix A is singular, its condition number is inf which is a numpy

number for infinity, as expected. More examples are as follows:

A = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 10.]])

# Last entry is 10.
print(A, '\n Condition number of A=',lg.cond(A))
[[ 1.000 0.000 0.000]
[ 0.000 1.000 0.000]
[ 0.000 0.000 10.000]]
Condition number of A= 10.0

The condition number of this A is 10.0, which is 10.0/1.0. Again, the

condition number is the ratio of the largest eigenvalue and the smallest
eigenvalue. We can now conclude that if the largest eigenvalue of a matrix is
118 Machine Learning with Python: Theory and Applications

very large or the smallest eigenvalue of the matrix is very small, the matrix
is likely singular, depending on their ratio.
This finding implies that a normalization to a matrix (which is often
done in machine learning) will not in theory change its condition number.
It may help in reducing the loss of significant digits (because of the limited
presentation of floats in computer hardware).

3.1.15 Rank of a matrix

We mentioned rank before. If a square matrix has a full rank, it implies that
all the columns (or rows) are mutually linearly independent. Such a matrix is
not singular. For a singular matrix, its rank should be less than full, which is
called rank deficiency. One can further ask what the level of rack deficiency
is. The answer is the deficit number of its rank. Essentially, if the matrix has
a rank of 2, we know that two linearly independent vectors can be formed
using these columns (or rows) in the matrix.
For a non-square matrix, the full rank is the number of its columns or
rows, whichever is smaller. Similarly, it can also have rank deficiency if the
rank is smaller than the full rank.
Let us now examine it in more detail using Numpy.

from numpy.linalg import matrix_rank, eig

print('Rank=',matrix_rank(np.eye(4))) # Identity matrix
Rank= 4

It is seen that the identity matrix has a shape of 4 × 4. It has a full rank.

A = np.array([[1,-0.2, 0], [0.1, 1,-0.5], [0.1, 1,-0.5]])

# singular A
print(A, '\n Rank=', matrix_rank(A))
[[ 1.000 -0.200 0.000]
[ 0.100 1.000 -0.500]
[ 0.100 1.000 -0.500]]
Rank= 2

This singular matrix has two linearly independent columns, and hence
a rank of 2. It has a rank deficiency of 1. Thus, it should also have a zero
eigenvalue, as shown below. If the matrix has a rank deficiency of n, it shall
have n zero eigenvalues. This is easy checked out using Numpy.
 Basic Mathematical Computations 119

np.set_printoptions(formatter={'float': '{: 0.3f}'.format})

eig(A)
(array([ 0.956, 0.544, -0.000]),
array([[ 0.955, -0.296, 0.088],
[ 0.209, -0.675, 0.438],
[ 0.209, -0.675, 0.894]]))

A = np.array([[1, -0.2, 0], [0.1, 1, -0.5]])

print('A:',A, '\n Rank=',matrix_rank(A))
A: [[ 1.000 -0.200 0.000]
[ 0.100 1.000 -0.500]]
Rank= 2

The matrix has only two rows, and a rank of 2. It has a full rank.

3.2 Rotation Matrix

For two-dimensional cases, the coordinate transformation (rotation) matrix

can be given as follows:
 
cos θ − sin θ
T= (3.8)
sin θ cos θ

where T is the transformation (or rotation) matrix, and θ is the rotation

angle. A given vector (displacement, force, for example) can be written with
two components in the coordinate system as follows:

d = [u v] (3.9)

The new coordinates of the vector dθ that is rotated by θ can be computed

using the rotation matrix.

import numpy as np
theta = 45 # Degree
thetarad = np.deg2rad(theta)
c, s= np.cos(thetarad), np.sin(thetarad)
T = np.array([[c, -s],
[s, c]])
print('Transformation matrix T:\n',T)
120 Machine Learning with Python: Theory and Applications

Transformation matrix T:
[[ 0.707 -0.707]
[ 0.707 0.707]]

d = np.array([1, 0]) # Original vector

T @ d # rotated by theta
array([ 0.707, 0.707])

T @ (T@d) # rotated by 2 thetas

array([-0.000, 1.000])

T @ T # 2 theta rotations
array([[ 0.000, -1.000],
[ 1.000, -0.000]])

T @ (T@(T @ T)) # 4 theta rotations

array([[-1.000, 0.000],
[ 0.000, -1.000]])

T@(T @ (T@(T @ (T @ (T@(T @ T)))))) # 8 theta rotations =

# no rotation
array([[ 1.000, 0.000],
[-0.000, 1.000]])

3.3 Interpolation

Interpolation is a frequently used numerical technique for getting an

approximate date based on known data. Machine learning is in some way
quite similar to the interpolation. This section studies general issues related
to interpolation, using numpy. Interpolation is also known as curve fitting.
We show here some examples of function interpolation and approximation,
using values given at discrete points in a space.
Let us try numpy.interp first. More descriptions can also be found in
Scipy documentation (https://docs.scipy.org/doc/numpy-1.13.0/reference/
generated/numpy.interp.html).
 Basic Mathematical Computations 121

3.3.1 1-D piecewise linear interpolation using numpy.interp

# Data available
xn = [1, 2, 3] # data: given coordinates x
fn = [3, 2, 0] # data: given function values at x
# Query/Prediction f at a new location of x
x = 1.5
f = np.interp(x, xn, fn) # get approximated value at x
print(f'f({x:.3f})≈{f:.3f}')
f(1.500)≈2.500

np.interp(2, xn, fn) # Is it a data-passing interpolation?

2.0

np.interp([0, 1, 1.5, 2.72, 3.14], xn, fn) # querying at

# more points

array([ 3.000, 3.000, 2.500, 0.560, 0.000])

np.interp(4, xn, fn)

0.0

In practice, we know that interpolation can be a quite dangerous

operation, and hence extra care is required, especially when extrapolating.
To avoid or to be made aware of an extrapolation, one can set a warning
when the interpolation occurs outside the domain that the data covers.

out_of_domain = -109109109.0 # A warning number is used

print(np.interp(2.9, xn, fn,right=out_of_domain))
# print out the number when extrapolation occurs
print(np.interp(3.5, xn, fn,right=out_of_domain))
0.20000000000000018
-109109109.0

Interpolation using higher-order polynomials can be more accurate but

can also be a bigger problem. Piecewise linear approximation has often been
found to be much safer and can be very effective, when dense data are
available. Given below is an example using piecewise linear interpolation for
the approximation of a sine function.
122 Machine Learning with Python: Theory and Applications

import matplotlib.pyplot as plt # module for plot the results

x = np.linspace(0, 2*np.pi, 20) # data: x values
y = np.sin(x) # data: function values at x
xvals = np.linspace(0, 2*np.pi, 50) # generate dense x data
# at which the values are
# obtained via interpolation
yinterp = np.interp(xvals, x, y)
plt.plot(x, y, 'o') # plot the original data points
plt.plot(xvals, yinterp, '-x') # plot interpolated data points
plt.show() # show the plots

Figure 3.1: Fitted sine curve.

3.3.2 1-D least-square solution approximation

This is to fit a given set of data with a straight line in x − y plane

y = wx + b (3.10)

We shall determine the gradient w and bias b, using the data pair [xi , yi ]. In
this example, Eq. (3.10) can be rewritten as

y =X·w (3.11)

where X = [x, 1] and w = [w, b]. Now, we can use the np.linalg.lstsq to solve
for w:

w_true,b_true = 1.0, -1.0 # used for generating data

x = np.array([0, 1, 2, 3]) # x value at which data
# will be generated
X = np.vstack([x, np.ones(len(x))]).T # Form the matrix of data
X
 Basic Mathematical Computations 123

array([[ 0.000, 1.000],

[ 1.000, 1.000],
[ 2.000, 1.000],
[ 3.000, 1.000]])

y = w_true*x+b_true+np.random.rand(len(x))/1.0
# generate y data random noise added
print(y)
w, b = np.linalg.lstsq(X, y, rcond=None)[0]
w, b
[-0.686 0.754 1.643 2.702]

(1.1055249814646126, -0.5550392342429096)

#help(np.linalg.lstsq) # to find out the details of this function.

import matplotlib.pyplot as plt

plt.plot(x, y, 'o', label='Original data', markersize=10)
plt.plot(x, w*x + b, 'r', label='Fitted line')
plt.legend()
plt.show()

Figure 3.2: Least square approximation of data via a straight line.

We have, in fact, created the simplest machine learning model, known as

linear regression.
Scipy package offers efficient functions for machine learning computations,
including interpolation. Let us examine some examples that are available
at the Scipy documentation (https://docs.scipy.org/doc/scipy/reference/
tutorial/interpolate.html).
124 Machine Learning with Python: Theory and Applications

3.3.3 1-D interpolation using interp1d

import numpy as np
from scipy import interpolate
from scipy.interpolate import interp1d
import matplotlib.pyplot as plt
x0, xL = 0, 18
x = np.linspace(x0, xL, num=11, endpoint=True)
# x data points
y = np.sin(-x**3/8.0) # y data points
print('x.shape:',x.shape,'y.shape:',y.shape)
f = interp1d(x, y) # linear interpolation
f2 = interp1d(x, y, kind='cubic') # Cubit interpolation
# try also quadratic
xnew = np.linspace(x0,xL,num=41,endpoint=True)
# x prediction points
plt.plot(x, y, 'o', xnew, f(xnew), '-', xnew, f2(xnew), '--')
plt.legend(['data', 'linear', 'cubic'], loc='best')
plt.show()
x.shape: (11,) y.shape: (11,)

Figure 3.3: Interpolation using Scipy.

3.3.4 2-D spline representation using bisplrep

x, y = np.mgrid[-1:1:28j, -1:1:28j] # x, y data grid

z = (x**2+y**2)*np.exp(-2.0*(x*x+y*y+x*y)) # z data
plt.figure()
 Basic Mathematical Computations 125

plt.pcolor(x, y, z, shading='auto') # plot the initial data

plt.colorbar()
plt.title("Function sampled at discrete points")
plt.show()

Figure 3.4: Spline representation using bisplrep, coarse grids.

xnew, ynew = np.mgrid[-1:1:88j, -1:1:88j] # for view at grid

tck = interpolate.bisplrep(x, y, z, s=0) # B-spline
znew = interpolate.bisplev(xnew[:,0], ynew[0,:], tck) # z value
plt.figure()
plt.pcolor(xnew, ynew, znew, shading='auto')
plt.colorbar()
plt.title("Interpolated function.")
plt.show()

Figure 3.5: Spline representation using bisplrep, fine grids.

126 Machine Learning with Python: Theory and Applications

3.3.5 Radial basis functions for smoothing and interpolation

Radial basis functions (RBFs) are useful basis functions for approximation
of functions. RBFs are distance functions, and hence work well for irregular
grids (even randomly distributed points), in high dimensions, and are often
found less prone to overfitting. They are also used for constructing meshfree
methods [2]. In using Scipy, the choices of RBFs are as follows:

• “multiquadric”: sqrt((r/self.epsilon)**2 + 1)
• “inverse”: 1.0/sqrt((r/self.epsilon)**2 + 1)
• “gaussian”: exp(-(r/self.epsilon)**2)
• “linear”: r
• “cubic”: r**3
• “quintic”: r**5
• “thin plate”: r**2 * log(r).

The default is “multiquadric”.

First, let us look at one-dimensional examples.

import numpy as np
from scipy.interpolate import Rbf, InterpolatedUnivariateSpline
import matplotlib.pyplot as plt

# Generate data
np.set_printoptions(formatter={'float': '{: 0.3f}'.format})
x = np.linspace(0, 10, 9)
print('x=',x)
y = np.sin(x)
print('y=',y)
# fine grids for plotting the interpolated data
xi = np.linspace(0, 10, 101)
# use fitpack2 method
ius=InterpolatedUnivariateSpline(x,y) # interpolation
# function
yi = ius(xi) # interpolated values at fine grids
plt.subplot(2, 1, 1) # have 2 sub-plots plotted together
plt.plot(x, y, 'bo') # original data points in blue dots
plt.plot(xi, np.sin(xi), 'r') # original function, red line
 Basic Mathematical Computations 127

plt.plot(xi, yi, 'g') # Spline interpolated, green line

plt.title('Interpolation using univariate spline')
plt.show()
# use RBF method
rbf = Rbf(x, y)
fi = rbf(xi)
plt.subplot(2, 1, 2) # have 2 plots plotted together
plt.plot(x, y, 'bo') # original data points in blue dots
plt.plot(xi, np.sin(xi), 'r') # original function, red line
plt.plot(xi, fi, 'g') # RBF interpolated, green line
plt.title('Interpolation using RBF - multiquadrics')
plt.show()
x= [ 0.000 1.250 2.500 3.750 5.000 6.250 7.500 8.750
10.000]
y= [ 0.000 0.949 0.598 -0.572 -0.959 -0.033 0.938 0.625
-0.544]

Figure 3.6: Comparison of interpolation using spline and radial basis function (RBF).

We now examine some two-dimensional examples.

import numpy as np
from scipy.interpolate import Rbf
import matplotlib.pyplot as plt
from matplotlib import cm
128 Machine Learning with Python: Theory and Applications

# 2-d tests - setup scattered data

x = np.random.rand(108)*4.0-2.0
y = np.random.rand(108)*4.0-2.0
z = (x+y)*np.exp(-x**2-y**2+x*y)
di = np.linspace(-2.0, 2.0, 108)
XI, YI = np.meshgrid(di, di)

# use RBF https://docs.scipy.org/doc/scipy/reference/

# generated/scipy.
# interpolate.Rbf.html#scipy.interpolate.Rbf
rbf = Rbf(x, y, z, epsilon=2)
ZI = rbf(XI, YI)

# plot the result

plt.pcolor(XI, YI, ZI, cmap=cm.jet, shading='auto')
plt.scatter(x, y, 88, z, cmap=cm.jet)
plt.title('RBF interpolation - multiquadrics')
plt.xlim(-2, 2)
plt.ylim(-2, 2)
plt.colorbar();

Figure 3.7: Two-dimensional interpolation using RBF.

RBFs can be used for interpolation over N -dimensions. Below is an

example for 3D.
 Basic Mathematical Computations 129

from scipy.interpolate import Rbf

x, y, z, d = np.random.rand(4, 20)
# randomly generated data 0~1
# 4D, 50 points in each dimension
#print(d) # original data
rbfi = Rbf(x, y, z, d) # RBF interpolator
xi = yi = zi = np.linspace(0, 1, 10)
di = rbfi(xi, yi, zi) # interpolated values
print(di)
[-0.378 -0.058 0.450 0.832 0.825 0.665 0.559 0.512
0.430 0.497]

3.4 Singular Value Decomposition

3.4.1 SVD formulation

Singular value decomposition (SVD) is an essential tool for many numerical
operations on matrices, including signal processing, statistics, and machine
learning. It is a general factorization of a matrix (real or complex) that
may be singular and of any shape. It is very powerful, because any such a
matrix exists an SVD and can be found numerically. It is a generalization of
the eigenvalue decomposition that works for diagonalizable square matrices,
which was discussed earlier in this chapter.
A general (real or complex, square, or not square) m × p matrix A has
the following singular value decomposition:

A = UΣV∗ (3.12)

where * stands for Hessian (transpose of the matrix and conjugate to the
complex values).
• U is an m × m unitary matrix.
• Σ is an m × p rectangular diagonal matrix with non-negative real numbers
on the diagonal entries.
• V is a p × p unitary matrix.
• The diagonal entries σi in Σ are known as the singular values of A.
• The columns of U are called the left-singular vectors of A.
• The columns of V are called the right-singular vectors of A.
More detailed discussions on SVD can be found at Wikipedia (https://
en.wikipedia.org/wiki/Singular value decomposition).
130 Machine Learning with Python: Theory and Applications

3.4.2 Algorithms for SVD

Computation of SVD for large matrices can be very expensive. The
often used SVD algorithm is based on the QR decomposition (https://en.
wikipedia.org/wiki/QR decomposition) and its variations. The basic idea is
to decompose the given matrix into an orthogonal matrix Q and an upper
triangular matrix R. Readers can refer to the Wikipedia page for more details
and the leads there on the related topic. Here, we discuss a simple approach
to compute SVD based on the well-established eigenvalue decomposition.
This approach is not used for numerical computation of SVD. This is because
it uses a normal matrix that with condition number squared, leading to
numerical instability issues for large systems. For our theoretical analysis
and formula derivation, this is not an issue, and thus will be used here. For
this simple approach to work, we would need to impose some condition on
matrix A.
Consider a general m×p matrix A of real numbers with m>p, and assume
it has a rank of p. Such a matrix is often encountered in machine learning.
We first form a normal matrix B:

B = A A (3.13)

which will be a p × p symmetric square matrix (smaller in size). Therefore,

B will be orthogonally diagonalizable. Because matrix A has a rank of p, B
will also be symmetric-positive-definite (SPD). Thus, B has an eigenvalue
decomposition, we perform such a decomposition. The results can be written
in the form of

B = Ve ΛVe (3.14)

where Ve is a p × p orthonormal matrix of p eigenvectors of the B matrix.

Λ is a p × p square diagonal matrix. The diagonal entries are the eigenvalues
that are positive real numbers.
On the other hand, we know that matrix A has an SVD decomposition;
we thus also have
A = UΣV (3.15)

Because matrix A has a rank of p, the singular value in Σ shall all be positive
real numbers. Using Eq. (3.15), we have
   
A A = UΣV UΣV = VΣU UΣV = VΣ2 V = B (3.16)
 Basic Mathematical Computations 131

In the above derivation, we used the fact that U is unitary: U U = I, and

Σ is diagonal (not affected by the transpose). Comparing Eq. (3.16) with
Eq. (3.14), we have
V = Ve (3.17)
√
Σ= Λ (3.18)

Using now Eq. (3.15), and the orthonormal property of V: V V = I, we

have

AV = UΣ (3.19)

Because Σ has all positive eigenvalues, we finally obtain

U = AVΣ−1 (3.20)

It is easy to confirm U is unitary. Finally, if A is rank deficient

(rank(A)< p), matrix B will have zero eigenvalues. In such cases, we simply
discard all the zero-eigenvalues and their corresponding eigenvectors. This
will still give us an SVD in a reduced form, and all the process given above
still holds
Readers may derive the similar set of equations for an m × p matrix A
of real numbers with p > m, and assume it has a rank of m.
The above analysis proves in theory that any matrix has a SVD. In
practical computations, we usually do not use the above procedure to
compute the SVD. This is because the condition number of matrix B is
squared, as seen in Eq. (3.13), leading to numerical instability. The practical
SVD algorithms often use the QR decomposition of A, which avoids forming
B. With the theoretical foundation, we now use some simple examples to
demonstrate the SVD process using Python.

3.4.3 Numerical examples

import numpy as np
a = np.random.randn(3, 6) # matrix with random numbers
print(a)
[[-1.582 0.398 0.572 -1.060 1.000 -1.532]
[ 0.257 0.435 -1.851 0.894 1.263 -0.364]
[-1.251 1.276 1.548 1.039 0.165 -0.946]]
132 Machine Learning with Python: Theory and Applications

u, s, vh = np.linalg.svd(a, full_matrices=True)
print(u.shape, s.shape, vh.shape)
(3, 3) (3,) (6, 6)

print('u=',u,'\n','s=',s,'\n','vh=',vh)
u= [[-0.677 0.283 0.679]
[ 0.202 0.959 -0.198]
[-0.707 0.003 -0.707]]
s= [ 3.412 2.466 1.881]
vh= [[ 0.589 -0.318 -0.544 0.048 -0.158 0.479]
[-0.083 0.216 -0.652 0.227 0.606 -0.318]
[-0.128 -0.382 -0.180 -0.867 0.166 -0.159]
[-0.510 -0.736 -0.110 0.412 -0.115 -0.066]
[ 0.300 -0.343 0.476 0.115 0.723 0.171]
[-0.529 0.218 -0.083 -0.104 0.210 0.781]]

smat = np.zeros((3, 6))

smat[:3, :3] = np.diag(s)
print(smat)
[[ 3.412 0.000 0.000 0.000 0.000 0.000]
[ 0.000 2.466 0.000 0.000 0.000 0.000]
[ 0.000 0.000 1.881 0.000 0.000 0.000]]

np.allclose(a, np.dot(u, np.dot(smat, vh)))

# Is original a recovered?
True

print(a)
print(np.dot(u, np.dot(smat, vh)))
[[-1.582 0.398 0.572 -1.060 1.000 -1.532]
[ 0.257 0.435 -1.851 0.894 1.263 -0.364]
[-1.251 1.276 1.548 1.039 0.165 -0.946]]
[[-1.582 0.398 0.572 -1.060 1.000 -1.532]
[ 0.257 0.435 -1.851 0.894 1.263 -0.364]
[-1.251 1.276 1.548 1.039 0.165 -0.946]]
 Basic Mathematical Computations 133

We note here that the SVD of a matrix keeps the full information of
the matrix: using all these singular values and vectors, one can recover the
original matrix. What if one uses only some of these singular values (and the
corresponding singular vectors)?

3.4.4 SVD for data compression

We can use SVD to compress data, by discarding some (often many) of these
singular values and vectors of the data matrix. The following is an example
of compressing an m × n array of an image data.

from pylab import imshow,gray,figure

from PIL import Image, ImageOps
import matplotlib.pyplot as plt
import numpy as np
%matplotlib inline
np.set_printoptions(formatter={'float': '{: 0.3f}'.format})

A = Image.open('../images/hummingbird.jpg') # open an image

print(np.shape(A)) # check the shape of A, it is a 3D tensor
A = np.mean(A,2) # get 2-D array by averaging RBG values
m, n = len(A[:,0]), len(A[1])
r = m/n # Aspect ratio of the original image
print(r,len(A[1]),A.shape,A.size)
fsize, dpi = 3, 80 # inch, dpi (dots per inch, resolution)
plt.figure(figsize=(fsize,fsize*r), dpi=dpi)
gray()
imshow(A)
U, S, Vh = np.linalg.svd(A, full_matrices=True)
print(U.shape, S.shape, Vh.shape)
# Recover the image
k = 20 # use first k singular values
S = np.resize(S,[m,1])*np.eye(m,n)
Compressed_A=np.dot(U[:,0:k],np.dot(S[0:k,0:k],Vh[0:k,:]))
#print(Compressed_A.shape,'Compressed_A=',Compressed_A)
plt.figure(figsize=(fsize,fsize*r), dpi=dpi)
gray()
imshow(Compressed_A)
134 Machine Learning with Python: Theory and Applications

(405, 349, 3)
1.160458452722063 349 (405, 349) 141345
(405, 405) (349,) (349, 349)

<matplotlib.image.AxesImage at 0x203185a7080>

Figure 3.8: Reproduced image using compressed data in comparison with the original
image.

It is clear that when k = 20 (out of 349) singular values are used, the
reconstructed image is quite close to the original one. Readers may estimate
how much the storage can be saved if one keeps 10% of the singular values
(and the corresponding singular vectors), assuming the reduced quality is
acceptable.
 Basic Mathematical Computations 135

3.5 Principal Component Analysis

Principal component analysis (PCA) is an effective technique to extract

features from datasets. It was invented in 1901 by Karl Pearson [3]. It is a pro-
cedure that converts a dataset of p observations (raw features) of a possibly
correlated to a reduced set of variables (extracted features), using an orthog-
onal transformation. The principal components produced via a PCA are not
linearly correlated and are sorted by their variance values. Principal compo-
nents at the top of the sorted list account for higher variability in the dataset.
It is an effective way to reduce the dimension of feature spaces of datasets,
so that machine learning models such as the neural network can work more
effectively [4, 5]. Note that PCA is known to be sensitive to the relative
scaling of the original observations, see Wikipedia (https://en.wikipedia.
org/wiki/Principal component analysis#cite note-1) for more details.
PCA can be performed at least in two ways. One is to use a regression
approach which finds the set of orthogonal axes in an iterative manner.
The other is to use eigenvalue decomposition algorithms. The following
examples use the 2nd approach.

3.5.1 PCA formulation

Consider a general m × p matrix A of real numbers with m>p. We first form
a matrix B:
B = A A (3.21)

which is a p × p symmetric square matrix with reduced size. Thus, it will be

at least semi-positive-definite, and often SPD. We can perform an eigenvalue
decomposition to it, which gives

B = VΣV (3.22)

These decomposed matrices are as follows:

• V is a p × p orthonormal matrix of p eigenvectors of the B matrix.
• Σ is a p×p square diagonal matrix. The diagonal entries are the eigenvalues
that are non-negative real numbers.
The PCA is then given as

AP CA = AV (3.23)

which is the projection of A on these p orthonormal eigenvectors. It has the

same shape as the original A that is m × p.
136 Machine Learning with Python: Theory and Applications

One may reconstruct A using the following formula, if using all the
eigenvectors:
Ar = AP CA V = AVV = A (3.24)

This is because eigenvectors are orthonormal. It is often that the first few
(ranked by the value of the eigenvalues in descending order) eigenvectors
contain most of the overall information of the original matrix A. In this
case, we can use only a small number of eigenvectors to reconstruct the A
matrix. For example, if we use k  p number of eigenvectors, we have
Ar = AP CA [0 : m, 0 : k]V [0 : k, 0 : p]
(3.25)
= A[0 : m, 0 : p]V[0 : p, 0 : k]V [0 : k, 0 : p] = A
This will, in general, not equal the original A, but can be often very close
to it. In this case, the storage becomes m × k + k × p which can be much
smaller than the original size of m × p. In Eq. (3.25), we used the Python
syntax, and hence it is very close to that in the Python code.
Note that if matrix A has dimensions of m < p, we simply treat its
transpose in the same way mentioned above.
One can also perform a similar analysis by forming a normal matrix B
using the following equation instead:
B = AA (3.26)

which will be an m×m symmetric square matrix of reduced size. Assuming it

is at least semi-positive-definite, we can perform an eigenvalue decomposition
to it, which gives
B = VΣV (3.27)

In this case, these decomposed matrices are as follows:

• V is an m × m orthonormal matrix of m eigenvectors of the B matrix.
• Σ is an m × m square diagonal matrix. The diagonal entries are the
eigenvalues that are non-negative real numbers.
The PCA is then given as
AP CA = V A (3.28)

It has the same shape as the original A that is m×p. One may reconstruct A
using the following formula and all the eigenvectors (that are orthonormal):

Ar = VAP CA = VV A = A (3.29)

 Basic Mathematical Computations 137

We can use only a small number of eigenvectors to reconstruct the A matrix.

For example, if we use k  m number of eigenvectors, we have

Ar = V[0 : k, 0 : m]V [0 : m, 0 : k]AP CA [0 : k, 0 : p]

(3.30)
= A

Note that for large systems, we do not really form the normal matrix
B, perform eigenvalue decomposition, and then compute V numerically.
Instead, the QR transformation type of algorithms are used. This is because
of the instability reasons mentioned in the beginning of Section 3.4.2.

3.5.2 Numerical examples

3.5.2.1 Example 1: PCA using a three-line code

We show an example of PCA code with only three lines. It is from glowing-
python (https://glowingpython.blogspot.com/2011/07/principal-component-
analysis-with-numpy.html), with permission. It is inspired by the function
princomp of matlab’s statistics toolbox and quite easy to follow. We modified
the code to exactly follow the PCA formulation presented above.

import numpy as np
from pylab import plot,subplot,axis,show,figure
def princomp(A):
""" PCA on matrix A. Rows: m observations; columns:
p variables. A will be zero-centered and normalized
Returns:
coeff: eig-vector of A^T A. Row-reduced observations,
each column is for one principal component.
score: the principal component - representation of A in
the principal component space.Row-observations,
column-components.
latent: a vector with the eigenvalues of A^T A.
"""
# eigenvalues and eigenvectors of covariance matrix
# modified. It was:
# M = (A-np.mean(A.T,axis=1))
138 Machine Learning with Python: Theory and Applications

# [latent,coeff] = np.linalg.eig(np.cov(M))
# score = np.dot(coeff.T,M)
A=(A-np.array([np.mean(A,axis=0)])) # subtract the mean
[latent,coeff] = np.linalg.eig(np.dot(A.T,A))
score = np.dot(A,coeff) # projection on the new space
return coeff,score,latent

Let us test the code using a 2D dataset.

# A simple 2D dataset
np.set_printoptions(formatter={'float': '{: 0.2f}'.format})

Data = np.array([[2.4,0.7,2.9,2.5,2.2,3.0,2.7,1.6,1.8,1.1,
1.6,0.9],
[2.5,0.5,2.2,1.9,1.9,3.1,2.3,2.0,1.4,1.0,
1.5,1.1]])
A = Data.T # Note: transpose to have A with m>p
print('A.T:\n',Data)
coeff, score, latent = princomp(A) # change made. It was A.T
print('p-by-p matrix, eig-vectors of A:\n',coeff)
print('A.T in the principal component space:\n',score.T)
print('Eigenvalues of A, latent=\n',latent)
figure(figsize=(50,80))
figure()
subplot(121)
# every eigenvector describe the direction of a principal
# component.
m = np.mean(A,axis=0)
plot([0,-coeff[0,0]*2]+m[0], [0,-coeff[0,1]*2]+m[1],'--k')
plot([0, coeff[1,0]*2]+m[0], [0, coeff[1,1]*2]+m[1],'--k')
plot(Data[0,:],Data[1,:],'ob') # the data points
axis('equal')
subplot(122)
# New data produced using the s
plot(score.T[0,:],score.T[1,:],'*g') # Note: transpose back
axis('equal')
show()
 Basic Mathematical Computations 139

A.T:
[[ 2.40 0.70 2.90 2.50 2.20 3.00 2.70 1.60 1.80
1.10 1.60 0.90]
[ 2.50 0.50 2.20 1.90 1.90 3.10 2.30 2.00 1.40 1.00
1.50 1.10]]
p-by-p matrix, eig-vectors of A:
[[ 0.74 -0.67]
[ 0.67 0.74]]
A.T in the principal component space:
[[ 0.82 -1.79 0.98 0.49 0.26 1.66 0.90 -0.11 -0.37
-1.16 -0.45 -1.24]
[ 0.23 -0.11 -0.33 -0.28 -0.08 0.27 -0.12 0.40 -0.18 -0.01
0.03 0.20]]
Eigenvalues of A, latent=
[ 11.93 0.58]

<Figure size 3600x5760 with 0 Axes>

Figure 3.9: Data process with PCA.

3.5.2.2 Example 2: Truncated PCA

This example is a modified PCA based on the previous code. The test is
done for an image compression application. The code is from glowingpython
(https://glowingpython.blogspot.com/2011/07/pca-and-image-compression-
with-numpy.html), with permission.
140 Machine Learning with Python: Theory and Applications

import numpy as np
def princomp(A,numpc=0):
# computing eigenvalues and eigenvectors of covariance
# matrix A
A = (A-np.array([np.mean(A,axis=0)]))
# subtract the mean (along columns)
[latent,coeff] = np.linalg.eig(np.dot(A.T,A))
#was: A = (A-np.mean(A.T,axis=1)).T # subtract the mean
#was: [latent,coeff] = np.linalg.eig(np.cov(M))

p = np.size(coeff,axis=1)
idx = np.argsort(latent) # sorting the eigenvalues
idx = idx[::-1] # in ascending order
# sorting eigenvectors according to eigenvalues
coeff = coeff[:,idx]
latent = latent[idx] # sorting eigenvalues
if numpc < p and numpc >= 0:
coeff = coeff[:,range(numpc)] # cutting some PCs

#score = np.dot(coeff.T,M) # projection on the new space

score = np.dot(A,coeff)
# projection of the data on the new space
return coeff,score,latent

The following code computes the PCA of matrix A, which is an image in

color scale. It first converts Image A into gray scale. After the PCA is done,
a different reduced number of principal components are used to reconstruct
the image.

from pylab import imread,subplot,imshow,title,gray,figure,

show,NullLocator
from ipykernel import kernelapp as app
from PIL import Image, ImageOps
%matplotlib inline
#A = Image.open('./images/hummingbirdcapsized.jpg')
A = Image.open('../images/hummingbird.jpg') # open an image
#A = ImageOps.flip(B) # flip it if so required
# or use A = imread('./images/hummingbirdcapsized.jpg')
A = np.mean(A,2) # to get a 2-D array
 Basic Mathematical Computations 141

full_pc = np.size(A,axis=1)
# numbers of all the principal components
r = len(A[:,0])/len(A[1])
print(r,len(A[1]),A.shape,A.size)
i = 1
dist = []
figure(figsize=(11,11*r))
for numpc in range(0,full_pc+10,50): # 0 50 100 ... full_pc
coeff, score, latent = princomp(A,numpc)
print(numpc,'coeff, score, latent \n',
coeff.shape, score.shape, latent.shape)
Ar = np.dot(score,coeff.T)+np.mean(A,axis=0)
#was:Ar = np.dot(coeff,score).T+np.mean(A,axis=0)
# difference in Frobenius.norm
dist.append(np.linalg.norm(A-Ar,'fro'))
# showing the pics reconstructed with less than 50 PCs
if numpc <= 250:
ax = subplot(2,3,i,frame_on=False)
ax.xaxis.set_major_locator(NullLocator())
ax.yaxis.set_major_locator(NullLocator())
i += 1
imshow(Ar) #imshow(np.flipud(Ar))
title('PCs # '+str(numpc))
gray()
figure()
imshow(A) #imshow(np.flipud(A))
title('numpc FULL: '+str(len(A[1])))
gray()
show()

1.160458452722063 349 (405, 349) 141345

0 coeff, score, latent
(349, 0) (405, 0) (349,)
50 coeff, score, latent
(349, 50) (405, 50) (349,)
100 coeff, score, latent
(349, 100) (405, 100) (349,)
150 coeff, score, latent
(349, 150) (405, 150) (349,)
200 coeff, score, latent
142 Machine Learning with Python: Theory and Applications

(349, 200) (405, 200) (349,)

250 coeff, score, latent
(349, 250) (405, 250) (349,)
300 coeff, score, latent
(349, 300) (405, 300) (349,)
350 coeff, score, latent
(349, 349) (405, 349) (349,)

Figure 3.10: Images reconstructed using reduced PCA components, in comparison with
the original image.
 Basic Mathematical Computations 143

We can see that 50 principal components give a pretty good quality image,
compared to the original one.
To assess the quality of the reconstruction quantitatively, we compute
the distance of the reconstructed images from the original one in the
Frobenius norm, for a different number of eigenvalues/eigenvectors used in
the reconstruction. The results are plotted in Fig. 3.11, with the x-axis for
the number of eigenvalues/eigenvectors used. The sum of the eigenvalues is
plotted in the blue curve, and the Frobenius norm is plotted in the red curve.
The sum of the eigenvalues relates to the level of variance contribution.

from pylab import plot, axis, cumsum

figure()
perc = cumsum(latent)/sum(latent)
dist = dist/max(dist)
plot(range(len(perc)),perc,'b',range(0,full_pc+10,50), dist,'r')
axis([0,full_pc,0,1.1])
show()

Figure 3.11: Quality of the reconstructed images.

In practical computations, the QR decomposition can be used to compute

the eigenvectors V, to avoid numerical stability, as discussed earlier for SVD.

3.6 Numerical Root Finding

Module scipy.optimize offers a function fsolve() to find roots of a set of

given nonlinear equations defined by f (x) = 0, with estimated locations of
the roots. The fsolve() is a wrapper around the algorithms in MINPACK
that uses essentially a variant of the Newton iteration method (https://en.
wikipedia.org/wiki/Newton%27s method), which finds the root using the
144 Machine Learning with Python: Theory and Applications

Figure 3.12: The Newton Iteration: The function is shown in blue and the tangent line at
local xi is in red. We see that xi gets closer and closer to the root of the function when the
number of iterations i increases (https://en.wikipedia.org/wiki/Newton%27s method#/
media/File:NewtonIteration Ani.gif) under the CC BY-SA 3.0. (https://creativecommons.
org/licenses/by-sa/3.0/) license.

function derivative to approximate the function locally. This process can

be easily viewed from the animation nicely made by Ralf Pfeifer shown in
Fig. 3.12.
Let us define two functions, one with a single variable and another
with two variables, as examples to demonstrate how to find the roots. For
functions with a single variable x, we define the following function that is
often encountered in structural mechanics problems:
import numpy as np
def BeamOnFoundation(bz): # Deflection of a beam on foundation
return np.exp(-bz)*(np.sin(bz)+np.cos(bz))
def f(x): # function whose root to be found
return 2*BeamOnFoundation(x/2)-1-BeamOnFoundation(x)

from scipy.optimize import fsolve

starting_guess = 5 # specify estimated location of the root
x_root=fsolve(f, starting_guess)
print('x_root=',x_root)
np.isclose(f(x_root), 0) # check if f(x_root)=0.0.
x root= [ 1.86]

array([ True])
 Basic Mathematical Computations 145

Let us now consider a set of two functions with two variables.

def f2d(x):
return [x[0]*np.sin(x[1])-5,x[1]*x[0]-x[1]-8]
# x is an array function-1 function-2
x_roots = fsolve(f2d, [3, 2]) # specify 2 estimated roots
print('x_roots=',x_roots)
np.isclose(f2d(x_roots), [0.0, 0.0]) # check if f(x_root)=0.0.

x roots= [ 5.25 1.88]

array([ True, True])

Note in general, a polynomial (or other algebraic equation) can have

complex roots, even though their coefficients are all real. This is another
case where the complex space is geometrically closed, but the real space is
not. A polynomial of nth order should have n roots, but they may not be
all in the real space. Some of them get into the complex space.

3.7 Numerical Integration

Numerical integration is one of the routine operations in computations for

practical problems in sciences and engineering. This is because only simple
functions can be analytically integrated, and one has to resort to numerical
means for real-life problems. Different types of numerical integration tech-
niques have been developed in the past, and numpy made the computation
easy to implement and use. Our discussion on this topic starts from the
classical trapezoid rule that may be familiar to many readers. More reference
materials can be found from the Scipy.integrate documentation (https://
docs.scipy.org/doc/scipy/reference/tutorial/integrate.html) and notebook.
community (https://notebook.community/sodafree/backend/build/ipyth
on/docs/examples/notebooks/trapezoid rule).

3.7.1 Trapezoid rule

The trapezoid rule for definite integration uses the following formula:
 n
b
1s

f (x) dx ≈ (xk − xk−1 ) (f (xk ) + f (xk−1 )) . (3.31)

a 2
k=1

We define a simple polynomial function and sample it in a finite range [a, b]

at ns number of sampling points equally spaced.
146 Machine Learning with Python: Theory and Applications

%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt
np.set_printoptions(formatter={'float': '{: 0.3f}'.format})

def f(x):
return 9*x**3-8*x**2-7*x+6 # define a polynomial function
a, b, n = -1., 2, 400 # large n for plotting the curve
x = np.linspace(a, b, n) # x at n points in [a,b]
y = f(x) # compute the function values

The function is integrated over [a, b], by sampling a small number of

points.

ns = 6 # sample ns points for integration

xint = np.linspace(a, b, ns)
yint = f(xint)

Plot the function curve and the trapezoidal (shaded) areas below it.

plt.plot(x, y, lw=2) # plot the function as a line of width 2

#plt.axis([a, b, 0, 150]) # plot x and y axes
plt.fill_between(xint, 0, yint, facecolor='gray', alpha=0.4)
# plot the shaded area over which the integration is done
plt.text((a+b)/2,12,r"$\int_a^b f(x)dx$", horizontalalignment=\
'center',fontsize=15); # use \ to change line in code

Figure 3.13: Integration using the trapezoidal rule.

 Basic Mathematical Computations 147

The trapezoid integration computes the shaded area. Thus, it is likely an

approximation as shown.

from scipy.integrate import quad # quadrature (integration)

integral, error = quad(f, a, b)
# shall give the results and the error
integral_trapezoid=sum((xint[1:]-xint[:-1])*(yint[1:]+yint[:-1]))/2
# use the trapezoid formula
print("The results should be:", integral, "+/-", error)
print("The results by the trapezoid approximation with",len(xint),
"points is:", integral_trapezoid)
The results should be: 17.25 +/- 1. 9775770133077287e-13
The results by the trapezoid approximation with 6 points are:
18.240000000000002

3.7.2 Gauss integration

Gauss integration (or quadrature) is regarded as one of the most effective
numerical integration techniques. It samples the integrand function at
specific points called the Gauss points and sums up these sampled function
values weighted by the Gauss weights for these points. It can produce exact
values (to machine accuracy) for the integration of a polynomial integrand,
because the Gauss point locations are the roots of the Legendre polynomials
defined in the natural coordinates in [−1, 1].
The Gauss integration is widely applied in numerical integration if the
fixed locations of sampling points are not a concern. It is a standard
integration scheme used in the FEM [1]. Here, we show an example using the
p roots() function available in Scipy module to find the roots of polynomials,
and then carry out the integration.

from pylab import *

from scipy.special.orthogonal import p_roots

def gauss(f,n,a,b):
[x,w] = p_roots(n+1) # roots of the Legendre polynomial
# and weights
G=0.5*(b-a)*sum(w*f(0.5*(b-a)*x+0.5*(b+a)))
# in natural coordinates
# sample the function values at these roots and sum up.
return G
148 Machine Learning with Python: Theory and Applications

def my_f(x):
return 9*x**3-8*x**2-7*x+6 # define a polynomial function
ng = 2
integral_Gauss = gauss(my_f,ng,a,b)
print("The results should be:", integral, "+/-", error)
print("The results by the trapezoid approximation with",
len(xint),"points is:", integral_trapezoid)
print("The results by the Gauss integration with", ng,
'Gauss points:', "points is:", integral_Gauss)

The integral results should be: 17.25 +/-

1.9775770133077287e-13
The results by the trapezoid approximation with 6 points is:
18.240000000000002
The results by the Gauss integration with 2 Gauss points:
points is: 17.250000000000007

It is observed that the Gauss integration gives a much more accurate

solution with a much smaller number of sampling points. In fact, the solution
is exact (within the machine error) for this example because the integrand
is a polynomial of the order of 3. We need only 2 Gauss points to obtain the
exact solution. The general formula for polynomial integrands is ng = n+1 2 ,
where n is the order of the polynomial integrand and ng is the number
of Gauss points needed to obtain the exact solution for the integral. Note
that when the trapezoid integration rule is used with 6 sampling points, the
solution is still quite far off.
For general complicated integrand functions, Gauss integration may not
give the exact solution. The accuracy, however, will still be much better
compared to the trapezoid rule or the rectangular rule (which we did not
discuss, but very similar to the trapezoid rule). In other words, for solutions
of similar accuracy, Gauss integration uses less sampling points.

3.8 Initial data treatment

Finally, let us introduce techniques often used for initial treatment for
datasets. Consider a given training dataset X ∈ Xm×p . In machine learning
models, m is the number of data-points in the dataset, and p is the number of
feature variables. The values of the data are often in a wide range for real-life
problems. For numerical stability reasons, we usually perform normalization
to the given dataset before feeding it to a model. There are mainly two
 Basic Mathematical Computations 149

techniques are used: min-max feature scaling and standard scaling. Such a
scaling or normalization is also called transformation in many ML modules.

3.8.1 Min-max scaling

For formulation for min-max scaling is given as follows.

X − X. min(axis = 0)
Xscaled = (3.32)
X. max(axis = 0) − X. min(axis = 0)

where X. min and X. min will be (row) vectors, and we used the Python
syntax of broadcasting rules and element-wise divisions. This would bring
all values for each feature into [0, 1] range. A more generalized formula that
can bring these values to an arbitrary range of [a, b] is given as follows.

X − X. min(axis = 0)
Xscaled = a + (b − a) (3.33)
X. max(axis = 0) − X. min(axis = 0)

Here we used again the Python syntax so that scalars, vectors and matrix
are all in the same formula.
Once such a scaling transformation to the training dataset is done, X. min
and X. min can be used to perform exactly the same transformation to the
testing dataset to ensure consistency for proper predictions.
The following is a simple code to perform min-max scaling using
Eq.(3.32).

np.set_printoptions(precision=4)
X = [[-1, 2, 8], # an assumed toy training dataset
[2.5, 6, 1.5], # with 4 samples, and 3 features
[3, 11, -6],
[21, 7, 2]]
print(f"Original training dataset X:\n{X}")
X = np.array(X)
X_scaled = (X - X.min(axis=0)) / (X.max(axis=0) - X.min(axis=0))
print(f"Scaled training dataset X:\n{X_scaled}")
print(f"Maximum values for each feature:\n{X.max(axis=0)}")
print(f"Minimum values for each feature:\n{X.min(axis=0)}")

Original training dataset X:

[[-1, 2, 8], [2.5, 6, 1.5], [3, 11, -6], [21, 7, 2]]
Scaled training dataset X:
[[0. 0. 1. ]
[0.1591 0.4444 0.5357]
150 Machine Learning with Python: Theory and Applications

[0.1818 1. 0. ]
[1. 0.5556 0.5714]]

Maximum values for each feature:

[21. 11. 8.]
Minimum values for each feature:
[-1. 2. -6.]

We can now perform the same transformation to the testing dataset using
X. min and X. min of the training dataset.

Xtest = [[-2, 3, 7], # assumed testing dataset

[5, 4, 5.5]]
Xt_scaled = (Xtest - X.min(axis=0)) / (X.max(axis=0) - X.min(axis=0))
print(f"Scaled corresponding testing dataset Xtest:\n{Xt_scaled}")

Scaled corresponding testing dataset Xtest:

[[-0.0455 0.1111 0.9286]
[ 0.2727 0.2222 0.8214]]

The inverse transformation can be done with ease.

X_back = X_scaled*(X.max(axis=0) - X.min(axis=0))+ X.min(axis=0)

print(f"Back transformed training dataset:\n{X_back}")

Xt_back = Xt_scaled*(X.max(axis=0) - X.min(axis=0))+ X.min(axis=0)

print(f"Back transformed testing dataset Xtest:\n{Xt_back}")

Back transformed training dataset:

[[-1. 2. 8. ]
[ 2.5 6. 1.5]
[ 3. 11. -6. ]
[21. 7. 2. ]]
Back transformed testing dataset Xtest:
[[-2. 3. 7. ]
[ 5. 4. 5.5]]

It is clearly seen that the min-max scaling does no harm to the dataset.
One can get it back as needed.
The same min-max scaling can be done using Sklearn.
 Basic Mathematical Computations 151

from sklearn.preprocessing import MinMaxScaler

scaler = MinMaxScaler() # create an instance
scaler.fit(X) # fit with the training dataset
X_scaled = scaler.transform(X) # perform the scaling transformation
print(f"Scaled training dataset X:\n{X_scaled}")
print(f"Maximum values for each feature:\n{scaler.data_max_}")
print(f"Minimum values for each feature:\n{scaler.data_min_}")

Scaled training dataset X:

[[0. 0. 1. ]
[0.1591 0.4444 0.5357]
[0.1818 1. 0. ]
[1. 0.5556 0.5714]]
Maximum values for each feature:
[21. 11. 8.]
Minimum values for each feature:
[-1. 2. -6.]

Xtest = [[-2, 3, 7], # assumed testing dataset

[5, 4, 5.5]]
Xt_scaled = scaler.transform(Xtest)
print(f"Scaled corresponding testing dataset:\n{Xt_scaled}")

Scaled corresponding testing dataset:

[[-0.0455 0.1111 0.9286]
[ 0.2727 0.2222 0.8214]]

X_back = scaler.inverse_transform(X_scaled)
print(f"Back transformed training dataset:\n{X_back}")

Xt_back = scaler.inverse_transform(Xt_scaled)
print(f"\nBack transformed testing dataset Xtest:\n{Xt_back}")

Back transformed training dataset:

[[-1. 2. 8. ]
[ 2.5 6. 1.5]
[ 3. 11. -6. ]
[21. 7. 2. ]]

Back transformed testing dataset Xtest:

[[-2. 3. 7. ]
[ 5. 4. 5.5]]
152 Machine Learning with Python: Theory and Applications

3.8.2 “One-hot” encoding

Many ML datasets use categorical features. For example, a color variable may
have values of “red”, “green”, and “blue”. These values must be converted
to numerical numbers for building a ML model. Consider a single column
feature vector is given originally as [[green], [red], [0], [blue]], one can simply
encode this feature vector as [[1], [2], [0], [3]], where the integers are arbitrary
but distinct. The treatment to dataset coded in this manner will be the same
as any ordinary dataset we discussed before. However, this implies that the
colors are having significance in values, which may not be what we want.
To avoid such a problem, we often use the so-called “one-hot” encoding.
The single column dataset is then encoded to a matrix X with three columns,
as shown in the code below. Thus, one-hot encoding results in a significant
increase in the number of feature vectors, so that the features can all be
made unique, and the categories are not given any value significance. Let us
scale such kind of dataset.

# red green blue

X = [[0, 1, 0 ], # a 'one-hot' training dataset
[1, 0, 0 ],
[0, 0, 0 ],
[0, 0, 1 ]]
scaler.fit(X)
print(f"Original 'one-hot' training dataset X:\n{X}")
X_scaled = scaler.transform(X) # perform the scaling transformation
print(f"Scaled 'one-hot' training dataset X:\n{X_scaled}")
print(f"Maximum values for each feature:\n{scaler.data_max_}")
print(f"Minimum values for each feature:\n{scaler.data_min_}")

Original 'one-hot' training dataset X:

[[0, 1, 0], [1, 0, 0], [0, 0, 0], [0, 0, 1]]
Scaled 'one-hot' training dataset X:
[[0. 1. 0.]
[1. 0. 0.]
[0. 0. 0.]
[0. 0. 1.]]
Maximum values for each feature:
[1. 1. 1.]
Minimum values for each feature:
[0. 0. 0.]

It is seen that the min-max scaling used has not changed anything to the
one-hot dataset, as expected. They do not have value significance.
 Basic Mathematical Computations 153

3.8.3 Standard scaling

When the dataset has a distribution that is close to the normal distribution,
one can use the standard scaling. The formulation for the standard scaling
is given as follows.
X − X.mean(axis = 0)
Xscaled = (3.34)
X.std(axis = 0)
The following is a simple code to perform min-max scaling using
Eq.(3.34).

X = [[-1, 2, 8], # an assumed toy training dataset

[2.5, 6, 1.5],
[3, 11, -6],
[21, 7, 2]]
print(f"Original training dataset X:\n{X}")
X = np.array(X)
X_scaled = (X - X.mean(axis=0)) / X.std(axis=0)
print(f"Standard scaled training dataset X:\n{X_scaled}")
print(f"Mean value for each feature:\n{X.mean(axis=0)}")
print(f"Standard deviation for each feature:\n{X.std(axis=0)}")

Original training dataset X:

[[-1, 2, 8], [2.5, 6, 1.5], [3, 11, -6], [21, 7, 2]]
Standard scaled training dataset X:
[[-0.8592 -1.4056 1.3338]
[-0.4515 -0.1562 0.0252]
[-0.3932 1.4056 -1.4848]
[ 1.7039 0.1562 0.1258]]
Mean value for each feature:
[6.375 6.5 1.375]
Standard deviation for each feature:
[8.5832 3.2016 4.9671]

Note that the values are not confirmed in [−1, 1]. It follows a normal dis-
tribution. We can now perform the same transformation to the corresponding
testing dataset using the fitted instance of the training dataset.

Xtest = [[-2, 3, 7], # an assumed testing dataset

[5, 4, 5.5]]
Xt_scaled = (Xtest - X.mean(axis=0)) / X.std(axis=0)
print(f"Standard scaled corresponding testing dataset Xtest:
\n{Xt_scaled}")
154 Machine Learning with Python: Theory and Applications

Standard scaled corresponding testing dataset Xtest:

[[-0.9757 -1.0932 1.1325]
[-0.1602 -0.7809 0.8305]]

The inverse transformation can be done with ease.

X_back = X_scaled*X.std(axis=0) + X.mean(axis=0)

print(f"Back transformed training dataset:\n{X_back}")

Xt_back = Xt_scaled*X.std(axis=0) + X.mean(axis=0)

print(f"Back transformed corresponding testing dataset Xtest:
\n{Xt_back}")

Back transformed training dataset:

[[-1. 2. 8. ]
[ 2.5 6. 1.5]
[ 3. 11. -6. ]
[21. 7. 2. ]]
Back transformed corresponding testing dataset Xtest:
[[-2. 3. 7. ]
[ 5. 4. 5.5]]

It is clearly seen that the standard scaling does no harm to the dataset.
One can get it back as needed.
The same standard scaling can be done using Sklearn.

from sklearn.preprocessing import StandardScaler

scaler = StandardScaler() # create an instance
scaler.fit(X)
X_scaled = scaler.transform(X)
print(f"Scaled dataset X:\n{X_scaled}")
print(f"Mean values for each feature:\n{scaler.mean_}")
print(f"Standard deviations for each feature:\n{np.sqrt(scaler.var_)}")

Scaled dataset X:
[[-0.8592 -1.4056 1.3338]
[-0.4515 -0.1562 0.0252]
[-0.3932 1.4056 -1.4848]
[ 1.7039 0.1562 0.1258]]
Mean values for each feature:
[6.375 6.5 1.375]
Standard deviations for each feature:
[8.5832 3.2016 4.9671]
 Basic Mathematical Computations 155

Xtest = [[-2, 3, 7], # assumed testing dataset

[5, 4, 5.5]]
Xt_scaled = scaler.transform(Xtest)
print(f"Scaled corresponding testing dataset Xtest:\n{Xt_scaled}")

Scaled corresponding testing dataset Xtest:

[[-0.9757 -1.0932 1.1325]
[-0.1602 -0.7809 0.8305]]

X_back = scaler.inverse_transform(X_scaled)
print(f"Back transformed training dataset:\n{X_back}")

Xt_back = scaler.inverse_transform(Xt_scaled)
print(f"\nBack transformed testing dataset Xtest:\n{Xt_back}")

Back transformed training dataset:

[[-1. 2. 8. ]
[ 2.5 6. 1.5]
[ 3. 11. -6. ]
[21. 7. 2. ]]

Back transformed testing dataset Xtest:

[[-2. 3. 7. ]
[ 5. 4. 5.5]]

Note that the same scaling can be done for the labels in the training
dataset, if they are not probability distribution types of data. When perform
the testing on the trained model or prediction using the trained model, the
labels should be scaled back to the original really data unit.
Also, it is a good practice to take look at the distribution of the data-
points. This is usually done after scaling so that the region of the data-points
is normalized. One may simply plot the so-called kernel density estimation
(KDE) using, for example, seaborn.kdeplot().

References

[1] G.R. Liu and S.S. Quek, The Finite Element Method: A Practical Course, Butterworth-
Heinemann, 2013. London.
[2] G.R. Liu, Mesh Free Methods: Moving Beyond the Finite Element Method, Taylor and
Francis Group, New York, 2010.
156 Machine Learning with Python: Theory and Applications

[3] K. Pearson, On lines and planes of closest fit to systems of points in space, Philosophical
Magazine, 2(1), 559–572, 1901.
[4] G.R. Liu, S.Y. Duan, Z.M. Zhang et al., Tubenet: A special trumpetnet for explicit
solutions to inverse problems, International Journal of Computational Methods,
18(01), 2050030, 2021. https://doi.org/10.1142/S0219876220500309.
[5] Shuyong Duan, Zhiping Hou, G.R. Liu et al., A novel inverse procedure via creating
tubenet with constraint autoencoder for feature-space dimension-reduction, Interna-
tional Journal of Applied Mechanics, 13(08), 2150091, 2021.
 Chapter 4

Statistics and Probability-based Learning Model

This chapter discusses some topics of probability and statistics related to

machine learning models and the computation techniques using Python.
Referenced materials include codes from Numpy documentation (https://
numpy.org/doc/), Jupyter documentation (https://jupyter.org/), and
Wikipedia (https://en.wikipedia.org/wiki/Main Page). Codes from mxnet-
the-straight-dope (https://github.com/zackchase/mxnet-the-straight-dope)
are also used under the Apache-2.0 License.
Building a machine learning model is mostly for prediction, classification,
or identification, based on the data available and the knowledge about the
data. Predictions can be deterministic and probabilistic. We often want to
predict the probability of the occurrence of an event, which can be very
useful and more practical for some problems.
For example, for aircraft maintenance, the engineers might want to assess
how likely it is for the engine of the aircraft to get into an unhealthy state,
based on records and/or diagnostic data. For a doctor, he/she may want
to predict the possibility of a patient having a critical illness in the next
period of time, based on the patient’s health records, diagnostic data, and
the current health environment. Health care organizations want to predict
the likelihood of the occurrence of a pandemic. For all these types of tasks,
we need to resort to means of quantifying the probability of the occurrence of
the event. It can be a complicated topic of study and research, and machine
learning models may help.
This chapter focuses on some of the basic concepts, theories, formulations,
and computational techniques that we may need to build machine learning
models using probability and statistics. At the end of this chapter, we will
introduce a Naive Bayes classification model.

157
158 Machine Learning with Python: Theory and Applications

4.1 Analysis of Probability of an Event

4.1.1 Random sampling, controlled random sampling

In machine learning, one often needs to sample numbers in a random manner.

This can be done numerically. In Python, we import random module to do so.
The random() can then be used to generate “simulated” random numbers.
First, let us use a code to produce random integers. We shall use
random.randint() for this, which samples numbers uniformly in a given
range.

# help(random.randint) # check it out

import random # random module

na, nb, n = 1, 100, 5 # n integers in na~nb
for i in range(n): # First, generate n
print(random.randint(na,nb),' ',end ='') # random integers
# in na~nb
print('\n')
for i in range(n): # Generate again
print(random.randint(na,nb),' ',end ='') # n random integers

86 61 35 81 40

4 92 95 31 70

We generated 5 random integer numbers twice. These generated numbers

are “random”, because two of the same generations gave two sets of different
numbers. One can try to execute the above cell multiple times, and it should
be found that each time different sets of numbers are generated.
Now, let us redo the same, but this time we use random.seed() to specify
the same seed for each of the generations.

random.seed(1) # seed value 1 for random number generation

for i in range(n): # Generate n random integers
print(random.randint(na, nb),' ',end ='')
print('\n')
random.seed(1) # The same seed value (try also seed(2))
for i in range(n):
print(random.randint(na, nb),' ',end ='')
 Statistics and Probability-based Learning Model 159

18 73 98 9 33

18 73 98 9 33

We see now that the same set of numbers is generated, which is some kind
of controlled random sampling by a seed value. The use of random.seed()
may confuse many beginners, but the above example shall eliminate the
confusion. Function random.seed() is used just to ensure the repeatability
when one reruns the code again, which is important for code development
ensuring reproducibility. We will use it quite frequently.
Also, we see the fact that random numbers generated by a computer
are not entirely random and are controllable to a certain degree. Naturally,
it should be, because any (classic) computer is deterministic in nature.
This pseudo-random feature is useful: when we study a probability event,
we make use of the randomness of random.randint() or random.random().
When we want our study and code to be repeatable, we make use of
random.seed().
Note that the seed value of 1 can be changed to any other number,
and with a different seed value used, a different set of random numbers
is generated.
Let us now generate real numbers.

#random.seed(1) # seed for random number generation

n = 5
for i in range(n): #generates n random real numbers
print(random.random())

0.11791870367106105
0.7609624449125756
0.47224524357611664
0.37961522332372777
0.20995480637147712

It is seen that real numbers are generated in between 0 and 1. It is

produced by generating a random integer first using random.randint() and
then dividing it by its maximum range. The reader may switch on and off
random.seed(1) or change the seed value to see the difference.
160 Machine Learning with Python: Theory and Applications

4.1.2 Probability
Probability is a numerical measure on the likelihood of the occurrence of an
event or a prediction. Assume, for example, the probability of the failure of
a structure is 0.1. We can then denote it mathematically as

P r(failure=“yes”) = 0.1 (4.1)

In this case, there is only one random variable that takes two possible discrete
values: “yes” with probability of 0.1, and “no” with probability of 0.9. Such
a distribution of a random variable is known as the Bernoulli distribution.
For general events, there may be more possible discrete random variables
and random variables with continuous distributions. Statistics studies the
techniques for sampling, interpreting, and analyzing the data about an event.
Machine learning is based on a dataset available for an event, and thus
statistical analysis helps us to make sense of a dataset and hopefully produces
a prediction in terms of probability.
We use Python to perform statistical analysis to datasets. We first
import necessary packages, including the MXNet packages (https://gluon.
mxnet.io/).

import numpy as np # numpy package, give an alias np

import mxnet as mx # mxnet package, give an alias mx
from mxnet import nd # ndarray class from mxnet package
mx.random.seed(1) # seed for random number generation
# for repeatability of the code

Let us consider a simple event: tossing a die that has six identical surfaces,
each of which is marked with a unique digit number, from 1 to 6. In this
case, the random variable can take 6 possible discrete values. Assume that
such markings do not introduce any bias (fair die), and do not affect in any
way the outcome of a tossing. We want to know the probability of getting
a particular number on the top surface, after a number of tossings. One can
then perform “numerical” experiments: tossing the die the large number of
times virtually in a computer and counting the times that the number shows
on the top surface. We use the following code to do this:

pr = nd.ones(6)/6 # probability distribution for a

# number on top. A total of 6
# values. Assume they have the
# same Pr (uniform distribution)
 Statistics and Probability-based Learning Model 161

print(pr)
n_top_array = nd.sample_multinomial(pr,shape=(1))
# toss once using the
# sample_multinomial() function
print('The number on top surface =', n_top_array)

[0.16666667 0.16666667 0.16666667 0.16666667 0.16666667

0.16666667]
<NDArray 6 @cpu(0)>
The number on top surface =
[3]
<NDArray 1 @cpu(0)>

For this problem, we know (assumed) that the theoretical or the “true”
probability for a number showing on the top surface is 1/6 ≈ 0.1667.
The one-time toss above gives an nd-array with just one entry that is the
number on the top surface of the die. To obtain a probability, we shall toss
for many times for statistics to work. This is done by simply specifying the
length of the nd-array in the handy nd.sample multinomial() function.

n_surfaces = 6 # number of possible values

n_tosses = 18 # number of tosses
mx.random.seed(1)
toss_results = nd.sample_multinomial(pr, shape=(n_tosses))
# toss n_tosses times
print("Tossed", n_tosses,'times.')
print("Toss results", toss_results)

Tossed 18 times.
Toss results
[3 0 0 3 1 4 4 5 3 0 0 2 2 2 4 2 4 4]
<NDArray 18 @cpu(0)>

This time, we tossed 18 times, resulting to an nd-array with 18 entries.

Note that if mx.random.seed(1) is not used in the above cell, we would get a
different array each tossing, because of the random nature. For this controlled
tossing (that readers can repeat) with random.seed(1), we got 1 “5” in 18
times of tossing. We thus have Pr(die=“5”) = 1/18. We got 3 “3”s, which
gives Pr(die=“3”) = 3/18=1/6, and so on. The values of Pr(die=“5”) and
Pr(die=“3”) are quite apart. Let us toss some more times.
162 Machine Learning with Python: Theory and Applications

n_t = 20
print(nd.sample_multinomial(pr, shape=(n_t))) #toss n_t times

[2 3 5 4 0 0 1 1 0 2 3 0 0 0 2 4 5 2 0 4]
<NDArray 20 @cpu(0)>

It is difficult to count and calculate the probabilities manually. Let us use

the following code available at the mxnet site to do so:

# The code is modified from these at https://github.com/

# zackchase/mxnet-the-straight-dope/blob/master/chapter01_
# crashcourse/probability.ipynb; # Under Apache-2.0 License.
#
n_tosses = 2000 # number of tosses
toss_results = nd.sample_multinomial(pr, shape=(n_tosses))
# toss, record the results
record=nd.zeros((n_surfaces,n_tosses)) # count the event (tossing)
# results: times of each of 6 surfaces appearing on top

n_digit_number = nd.zeros(n_surfaces) # Initial with zeros for an

# array to hold the probability of on-top
# appearances for each of the 6 numbers
for i, digit_number in enumerate(toss_results):
n_digit_number[int(digit_number.asscalar())] += 1
# counts and put in the
# corresponding place.
record[:,i] = n_digit_number # records the results

n_digit_number[:]=n_digit_number/n_tosses # compute the Pr

print('Total number of tosses:',n_tosses)
print('Probability of each of the 6 digits:',n_digit_number)
print('Theoretical (true) probabilities:',pr)

Total number of tosses: 2000

Probability of each of the 6 digits:
[0.1675 0.1865 0.1705 0.1635 0.15 0.162 ]
<NDArray 6 @cpu(0)>
Theoretical (true) probabilities:
[0.16667 0.16667 0.16667 0.16667 0.16667 0. 16667]
<NDArray 6 @cpu(0)>

We see the probability values for all the 6 digits getting closer to the
theoretical or true probability.
 Statistics and Probability-based Learning Model 163

import numpy as np
np.set_printoptions(suppress=True)
print(record) # print out the records
[[ 0. 0. 0. ... 333. 334. 335.]
[ 0. 0. 1. ... 373. 373. 373.]
[ 0. 1. 1. ... 341. 341. 341.]
[ 1. 1. 1. ... 327. 327. 327.]
[ 0. 0. 0. ... 300. 300. 300.]
[ 0. 0. 0. ... 324. 324. 324.]]
<NDArray 6x2000 @cpu(0)>

We now normalized the data, which we often do in machine learning, by

the total number of tosses using the following codes:

x = nd.arange(n_tosses).reshape((1,n_tosses)) + 1
#print(x)
observations = record / x # Pr of 6 digits for all tosses
print(observations[:,0]) # observations for 1st toss
print(observations[:,10]) # for first 10 toss
print(observations[:,999]) # for first 1000 toss
[0. 0. 0. 1. 0. 0.]
<NDArray 6 @cpu(0)>

[0.181819 0.272728 0.090909 0.181819 0.090909 0. 181819]

<NDArray 6 @cpu(0)>

[0.175 0.185 0.16 0.164 0.144 0.172]

<NDArray 6 @cpu(0)>

This simple experiment gives us 1,000 observations for six possible values
of uniform distribution (any of the 6 digits has equal chance to land on
top). When the probability of the appearance of each of the six surfaces
of the die is computed after 1,000 times of tossing, we got roughly 0.14 to
0.19. These probabilities will change a little each time we do the experiment
because of the random nature. If we would do 10,000 times of tossing for each
experiment, we shall get all probabilities quite close to the theoretical value
of 1/6 ≈ 0.1667. Readers can try this very easily using the code given above.
Let us now plot the “numerical” experimental results. For this, we use
matplotlib library.
164 Machine Learning with Python: Theory and Applications

# The code is modified from these at https://github.com/

# zackchase/mxnet-the-straight-dope/blob/master/chapter01_
# crashcourse/probability.ipynb; Under Apache-2.0 License.
%matplotlib inline
from matplotlib import pyplot as plt
plt.plot(observations[0,:].asnumpy(),label="Observed P(die=1)")
plt.plot(observations[1,:].asnumpy(),label="Observed P(die=2)")
plt.plot(observations[2,:].asnumpy(),label="Observed P(die=3)")
plt.plot(observations[3,:].asnumpy(),label="Observed P(die=4)")
plt.plot(observations[4,:].asnumpy(),label="Observed P(die=5)")
plt.plot(observations[5,:].asnumpy(),label="Observed P(die=6)")
plt.axhline(y=0.166667, color='black', linestyle='dashed')
plt.legend()
plt.show()

Figure 4.1: Probabilities obtained via finite sampling from a uniform distribution of a
fair die.

It is clear that the more experiments we do, the probability gets closer
to the theoretical value of 1/6.
The above discussion is for very simple events of a die toss. It gives a
clear view on some of the basic issues and procedures related to the statistics
analysis and probability computation for complicated events.

4.2 Random Distributions

In machine learning, one often needs to sample numbers in a random manner.

Depending on types of problems, the distribution of the data of a variable
can have different types. Sampling of data numerically shall be based on a
given/assumed distribution type. We did so in the beginning for this chapter
using uniform distribution. We shall now examine this further.
 Statistics and Probability-based Learning Model 165

4.2.1 Uniform distribution

These numbers generated based on uniform distribution shall have an equal
chance to land anywhere in between the specified range. To check the
uniformity of the numbers generated using random.randint(), we can run
it for a large number of times, say 1 million, and see how these numbers are
distributed. We use the following code to do so:
# This code are modified from these at https://github.com/
# zackchase/mxnet-the-straight-dope/blob/master/chapter01_
# crashcourse/probability.ipynb; Under Apache-2.0 License.
import numpy as np
import matplotlib.pyplot as plt
import random
na, nb, n = 0, 99, 100
counts = np.zeros(n) # Array to hold counted numbers
fig, axes = plt.subplots(2,3,figsize=(15,8),sharex=True)
axes = axes.reshape(6)
n_samples = 1000001
for i in range(1, n_samples):
counts[random.randint(na, nb)]+=1 # Random integers
if i in [10, 100, 1000, 10000, 100000, 1000000]:
axes[int(np.log10(i))-1].bar(np.arange(na+1,nb+2),counts)
plt.show()

Figure 4.2: Finite samplings from a uniform distribution.

It is observed that with the increase of sampling, the uniformity increases.

4.2.2 Normal distribution (Gaussian distribution)

The normal distribution is also called Gaussian distribution. It is widely used
in statistics because many events in nature, science, and engineering obey
166 Machine Learning with Python: Theory and Applications

this distribution. It is defined using the following Gaussian density function

of variable x:
 2
1 − x−μ
√
p(x) = √ e 2σ (4.2)
σ 2π
where μ and σ are, respectively, the mean and the standard deviation of
the distribution. The normal distribution is often denoted as N (μ, σ 2 ).
In particular, when μ = 0 and σ = 1, we have the standard normal
distribution denoted as N (0, 1) and its density function becomes simply
−( √x )2
p(x) = √12π e 2 .
The gauss() in the numpy random module can be used to conveniently
generate normal distribution numbers.
from random import gauss
mu, sigma, n = 0., 0.1, 10
for i in range(n): # generates n random numbers
print(f'{gauss(mu, sigma):.4f} ', end ='')
# mu: mean; sigma: standard deviation

-0.1939 0.1794 0.0614 -0.1348 0.1020 0.0432 -0.2144 -0.0636

0.0502 -0.1377

Let us plot out the density function defined in Eq. (4.2). The bell shape of
the function may already be familiar to you.
x = np.arange(-0.5, 0.5, 0.001) # define variable x
def gf(mu,sigma,x): # define the Gauss function
return 1/(sigma*np.sqrt(2*np.pi))*np.exp(-.5*((x-mu)/sigma)**2)
mu, sigma = 0, 0.1 # mean 0, standard deviation 0.1
plt.figure(figsize=(6, 4))
plt.plot(x, gf(mu,sigma,x))
plt.show()

Figure 4.3: A typical normal distribution (Gaussian distribution).

 Statistics and Probability-based Learning Model 167

Let us now generate some random samples of Gauss distribution using

np.random.normal(). We then compare the sampled data with the “true”
Gaussian distribution.
n = 500
samples = np.random.normal(mu, sigma, n) #generate samples
count, bins, ignored = plt.hist(samples, 80, density=True)
# plot histogram of the samples
plt.plot(bins, gf(mu,sigma,bins),linewidth=2, color='r')
# plot true Gauss distribution
plt.show()

Figure 4.4: Sampling from a normal (Gaussian) distribution.

Numpy can generate samples of about 40 different types of distributions.

Readers are referred to the numpy documentation (https://numpy.org/
doc/1.16/reference/routines.random.html) for details when needed.

4.3 Entropy of Probability

For given probabilities of random variables of a statistics event, one can

evaluate the corresponding entropy. It is a measure for uncertainty of the
probability distribution for the event. It is the dot-product of the probability
vector (that holds the probability values of a random variable) with its
negative logarithm. The entropy Hp for an event with probability p is
expressed by  
Hp = − pi log pi = −p · log(p) (4.3)
i

where pi is the probability of the ith possible value of the variable and

i pi = 1. Vector p is the vector that holds these the probabilities. The
168 Machine Learning with Python: Theory and Applications

negative sign is needed, because entropy is positive and log(pi ) is always

negative for 0 ≤ pi ≤ 1. In the computation, we often have it normalized by
dividing it with the total number of possible values. Entropy is used very
often in machine learning, especially for constructing objective functions,
because it is a measure of the uncertainty that needs to be minimized.
Since the logarithm is used frequently, we shall first examine it in more
detail using numpy log() function.

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
p = np.arange(0.01, 1.0, .01) # generate variables
logp = -np.log(p)
# negative sign for positive value: log(p)<0 for 0<p<1
plt.plot(p,logp,color='blue')
plt.xlabel('Probability p')
plt.ylabel('Negative log(p)')
plt.title('Negative log Value of Probability')
plt.show()

Figure 4.5: A log function of probability.

Let us mention a few important features, which are the root reasons for
why the logarithm is used so often in machine learning.

• −log(p) is monotonically varying with argument p. The monotonous is

important because it does not affect the locations of the stationary points
of the original function when it is in logarithm. This is an excellent
property for optimization algorithms that are frequently used in machine
learning.
 Statistics and Probability-based Learning Model 169

• −log(p) is monotonically decaying with increasing probability. This

reverses the trend of the probability, and thus it is the proper behavior
for measuring the entropy for a high probability range. This is because
when the probability is high, the uncertainty level is low (because we are
quite certain that it is likely to happen) and so the value of entropy. When
the probability is low, the uncertainty should also be low (because we are
quite certain that it is unlikely to happen). In this case, we simply make
use of the probability itself in the entropy equation.
• Entropy Hp is a combination of both the probability and its negative
logarithm in the form of a product, as shown in Eq. (4.3). This combination
gives the needed behavior, and it is nicely defined to suit our purpose by
making use of the features of the logarithm function.

The following examples demonstrate how the entropy function works:

4.3.1 Example 1: Probability and its entropy

Consider an event with a variable that takes two values. We made an

observation first, which produces probability vector q1 with entries of these
two probabilities of the corresponding two variables. We then made another
observation, which produces probabilities q2 . We would like to evaluate the
entropy of the probability of these two observations.

q1 =np.array([0.999, 0.001]) # Pr. distribution with low

# uncertainty: quite sure whether the event is to
# happen, because the variable is with either a
# very high or low chance to be observed.
q2 =np.array([ 0.5,0.5 ]) # Distribution with high uncertainty:
# Not sure whether the event is to happen,
# because the variable is with neither a
# high or low chance to be observed.
# - sign for getting a positive value:
print('q1=',q1,' -log(q1)=',-np.log(q1)) # log(p): negative
# for 0<p<1
print('q2=',q2,' -log(q2)=',-np.log(q2))
H_q1 = -np.dot(q1,np.log(q1))/len(q1) #Entropy: Uncertainty
H_q2 = -np.dot(q2,np.log(q2))/len(q2)
print('H_q1=',H_q1,'H_q2=',H_q2)

q1= [0.999 0.001] -log(q1)= [0.0010005 6.90775528]

q2= [0.5 0.5] -log(q2)= [0.69314718 0.69314718]
H q1= 0.003953627556116044 H q2= 0.34657359027997264
170 Machine Learning with Python: Theory and Applications

In this example, we see the contradicting behavior of p and −log(p).

Vector q1 has either low or high probability values for its two variables,
meaning that the uncertainty is low, and hence the computed entropy is
low. On the other hand, q2 have two probabilities in the middle for both
variables, meaning that it is very uncertain. The computed entropy is high
as expected.

4.3.2 Example 2: Variation of entropy

To show this more clearly, we create artificial events q1 with a variable that
takes two possible values, and let the probabilities be v1 and v2 for these two
values, which change in the reverse manner while the sum of the probabilities
equals 1. We write the following code to compute the entropy changing with
the changes in the probability of v1 and v2 :

# An event with a variable that takes two values

v1 = np.arange(0.01, 1.0, .05)
gap = (v1[1]-v1[0])*len(v1)/3.
v1 /= (v1[0]+v1[-1]) # create an array that holds linearly
# changing probability values.
v2 = v1[::-1] # create the revise of v1.
print(v1,np.sum(v1)) # to check it out
print(v2,np.sum(v2))
print(v1+v2,np.sum(v1+v2)/2)
xtick = range(len(v1)) #[0,1,2,3,4]
plt.bar(range(len(v1)),v1,width=gap*1.2,alpha=.9,color='blue')
plt.bar(range(len(v2)),v2,width=gap,alpha=.9,color='red')
plt.xlabel('Event ID, blue: v1, red: v2')
plt.ylabel('Probability')
plt.xticks(xtick)
plt.show()

H_qf = np.array([]) # initialize the array for entropy

for q1 in list(zip(v1,v2)):
# create a pair of probability compute the entropy and append
H_qf = np.append(H_qf,-(np.dot(q1,np.log(q1)))/2)

plt.plot(v1,H_qf)
plt.xlabel('Probability, v1 (v2=1-v1)')
plt.ylabel('Entropy of events')
plt.title('Entropy of Events')
plt.show()
 Statistics and Probability-based Learning Model 171

[0.01030928 0.06185567 0.11340206 0.16494845 0.21649485 0.26804124

0.31958763 0.37113402 0.42268041 0.4742268 0.5257732 0.57731959
0.62886598 0.68041237 0.73195876 0.78350515 0.83505155 0.88659794
0.93814433 0.98969072] 10.0
[0.98969072 0.93814433 0.88659794 0.83505155 0.78350515 0.73195876
0.68041237 0.62886598 0.57731959 0.5257732 0.4742268 0.42268041
0.37113402 0.31958763 0.26804124 0.21649485 0.16494845 0.11340206
0.06185567 0.01030928] 9.999999999999998
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.] 10.0

Figure 4.6: Probabilities of 20 events each of which has two values.

Figure 4.7: Variation of entropy of the probabilities of the 20 events.

It is clear that when the probability of v1 and v2 is at 0.5, the entropy is

the largest. The entropy is smallest at two ends, as expected.
172 Machine Learning with Python: Theory and Applications

4.3.3 Example 3: Entropy for events with a variable that takes

different numbers of values of uniform distribution
Let us take a look at events with a variable that can take different numbers of
possible values. We assume that the probability distribution for the variable
is uniform for all these events. We want to find out how the entropy for the
probability distribution changes with the number of variables of the events.

# An event with a variable that can take many values of

→uniform probability

N = 0
max_v = 100 # Events with N variables
# capped at max_values.
Ni = np.array([]) # For the number of v
H_qf = np.array([]) # To hold the entropy
while N < max_v:
N += 1
Ni = np.append(Ni,N)
qf = np.ones(N)
qf = qf/np.sum(qf) # uniform sample generated
H_qf = np.append(H_qf,-np.dot(qf,np.log(qf))/len(qf))

print('Probability distribution:',qf[0:max_v:10])
print('H_qf=', H_qf[0:max_v:10])
import matplotlib.pyplot as plt
import numpy as np
%matplotlib inline
plt.plot(Ni,H_qf)
plt.xlabel('Number of variables, all with same probability')
plt.ylabel('Entropy')
plt.title('Events with variables of uniform distribution')
plt.show()

Probability distribution: [0.01 0.01 0.01 0.01 0.01 0.01 0.01

0.01 0.01 0.01]
H qf= [-0. 0.21799048 0.14497726 0.11077378 0.09057493
0.07709462 0.06739137 0.06003774 0.05425246 0.04956988]
 Statistics and Probability-based Learning Model 173

Figure 4.8: Entropy of probability of events with uniform distributions.

We find that (1) the entropy is zero when N = 1; (2) it peaks at N = 3;

and (3) when N get very big, the entropy becomes small, implying that when
an event has a very large number of variables, the entropy becomes small.
This is because the probability of each of the variables becomes very small
due to the uniform probability assumption for all these variables.

4.4 Cross-Entropy: Predicated and True Probability

Let us now look at the cross-entropy, which is an often used concept in

statistics. The cross-entropy of a distribution q relative to the distribution
p is defined as follows:
 
Hpq = − pi log qi = −p · log(q) (4.4)
i

In general, the cross-entropy is a measure of the similarity of two distribu-

tions (from the same space). In machine learning, we are interested in the
cross-entropy of the predicated probability q with respect to the true one, p.
In this case, Hpq can be a measure of the performance of a prediction model,
and hence often used as an objective or loss function in machine learning
models.
We note the following properties:

• Cross-entropy is not symmetric: Hpq = Hqp , if p = q, which is obvious

from Eq. (4.4).
• We shall have Hpq ≥ Hp , and Hqp ≥ Hq . The difference will be the
KL-divergence, which is always positive (see the following section).
174 Machine Learning with Python: Theory and Applications

• When these two distributions are the same, the cross-entropy becomes the
entropy studied in the previous section. All these three inequalities above
become equal.
• Therefore, in machine learning models, even if the prediction is perfect,
the cross-entropy will still not be zero, because the true distribution itself
may have an entropy. If p is the entropy of the true distribution, the cross-
entropy Hpq is bounded from below by Hp . It can only be zero if the true
distribution is without any uncertainty (probabilities of the variables are
all zero, except for one of them, which is 1).

We look at some simple examples.

4.4.1 Example 1: Cross-entropy of a quality prediction

We examine a simple event with a variable that can take two possible values.
Assuming we have a quality prediction, how is it measured in cross-entropy?

# A good prediction case

q_good = np.array([0.9,0.1]) # predicted Pr. of 2 values
y = np.array([0.99,0.01]) # true Pr. of 2 values
p = y # Truth
q = q_good # prediction
print('p=',p,' log(p)=',-np.log(p))
print('q=',q,' log(q)=',-np.log(q))
print(' Entropy: Hp=',-np.dot(p,np.log(p))/len(p),\
' Hq=',-np.dot(q,np.log(q))/len(q))
print('\nCross-entropy: Hpq=',-np.dot(p,np.log(q))/len(p),\
'Hqp=',-np.dot(q,np.log(p))/len(q))
#Cross-entropy: Hpq>Hp; Hqp>Hq should hold.

p= [0.99 0.01] log(p)= [0.01005034 4.60517019]

q= [0.9 0.1] log(q)= [0.10536052 2.30258509]
Entropy: Hp= 0.028000767177423672 Hq= 0.1625414866957241

Cross-entropy: Hpq= 0.06366638 Hqp= 0. 234781160

It is seen that the cross-entropy Hpq is low, indicating that the prediction
q is good. Notice that Hpq = Hqp , Hpq ≥ Hp , and Hqp ≥ Hq .
 Statistics and Probability-based Learning Model 175

4.4.2 Example 2: Cross-entropy of a poor prediction

Consider again a simple event with a variable that can take two possible
values. This time, we assume a poor-quality prediction, and we examine
how it is measured in cross-entropy.

# A totally-off prediction case

q_bad = np.array([0.1,0.9]) # predicted Pr. of 2 values
y = np.array([0.99,0.01]) # true Pr. of 2 values
p = y # Truth
q = q_bad # prediction
print('p=',p,' log(p)=',-np.log(p))
print('q=',q,' log(q)=',-np.log(q))

print(' Entropy: Hp=',-np.dot(p,np.log(p))/len(p),\

' Hq=',-np.dot(q,np.log(q))/len(p))
print('\nCross-entropy: Hpq=',-np.dot(p,np.log(q))/len(p),\
'Hqp=',-np.dot(q,np.log(p))/len(q))
#Cross-entropy: Hpq>Hp; Hqp>Hq should hold.

p= [0.99 0.01] log(p)= [0.01005034 4.60517019]

q= [0.1 0.9] log(q)= [2.30258509 0.10536052]
Entropy: Hp= 0.028000767177423672 Hq= 0.1625414866957241

Cross-entropy: Hpq= 1.1403064236103417 Hqp= 2.072829100487316

It is seen that the cross-entropy Hpq is high, indicating that the prediction
q is bad. Notice again that Hpq = Hqp , Hpq ≥ Hp , and Hqp ≥ Hq .
We are now ready to discuss the KL-divergence.

4.5 KL-Divergence

Kullback-Leibler Divergence or KL-divergence is a measure of the relative

entropy from one distribution to another. For the given two distributions p
and q, the KL-divergence from q to p is defined as

DKL (p||q) = pi · [log pi − log qi ] = p · [log(p) − log(q)] (4.5)
i

It is also referred to as the relative entropy of q with respect to p that can

be regarded as the true or reference distribution. Using the definitions for
176 Machine Learning with Python: Theory and Applications

the entropy and the cross-entropy, we shall have

DKL (p||q) = Hpq − Hp (4.6)

Note that the KL-divergence of q with respect to p is different from that of

p with respect to q. We have also

DKL (p||q) ≥ 0, equality holds only if p = q (4.7)

This is known as the Gibbs’ inequality (https://en.wikipedia.org/wiki/

Gibbs%27 inequality).
Two simple examples of KL-divergence are given below.

4.5.1 Example 1: KL-divergence of a distribution

of quality prediction
We examine a simple event with a variable that can take two possible values.
Assuming we have a quality prediction of a distribution in relation to the
true distribution, how is it measured in KL-divergence?

# Good prediction case

p = y # true or reference distribution
q = q_good # prediction
print('p=',p,' log(p)=',-np.log(p))
print('q=',q,' log(q)=',-np.log(q))
Dpq=np.sum(np.dot(p,(np.log(p)-np.log(q))))/len(p)
Dqp=np.sum(np.dot(q,(np.log(q)-np.log(p))))/len(p)
print('Dpq=',Dpq,' Dqp=',Dqp)

p= [0.99 0.01] log(p)= [0.01005034 4.60517019]

q= [0.9 0.1] log(q)= [0.10536052 2.30258509]
Dpq= 0.035665613538170556 Dqp= 0.07223967373775611

It is seen that the KL-divergences Dpq and Dqp are all positive. They
all have low values, indicating that the prediction q is good. Notice that
Dpq = Dqp .

4.5.2 Example 2: KL-divergence of a poorly

predicted distribution
Consider again a simple event with a variable that can take two possible
values. Assume that we have a poor prediction of a distribution in relation
 Statistics and Probability-based Learning Model 177

to the true or reference distribution. We examine how it is measured in

KL-divergence using the following code:

# A bad prediction case

p = y
q = q_bad
print('p=',p,' log(p)=',-np.log(p))
print('q=',q,' log(q)=',-np.log(q))
Dpq=np.dot(p,(np.log(p)-np.log(q)))/len(p)
Dqp=np.dot(q,(np.log(q)-np.log(p)))/len(p)
print('Dpq=',Dpq, ' Dqp=',Dqp)

p= [0.99 0.01] log(p)= [0.01005034 4.60517019]

q= [0.1 0.9] log(q)= [2.30258509 0.10536052]
Dpq= 1.112305656432918 Dqp= 1.910287613791592

It is seen again that the KL-divergences Dpq and Dqp are all positive.
They all have high values, indicating the prediction q is poor. Notice also
that Dpq = Dqp .

4.6 Binary Cross-Entropy

Let us finally look at the so-called binary cross-entropy used in machine

learning. For the given two distributions p and q, the binary cross-entropy
of q with respect to p is defined as
 

B
Hpq =− pi · log qi + (1 − pi ) · log (1 − qi )
i (4.8)
= −p · log(q) − (1 − p) · log(1 − q)

In machine learning models, we usually assume that p is the truth

distribution, which can have probability 0 or 1, and hence cannot be subject
to logarithm. The binary cross-entropy can be viewed as a measure of the
entropy of the predicated probability with respect to the true one. It takes
into account both the probability p and q and the converse probability
(1 − p) and (1 − q), and computes the entropy of both of them. It roughly
doubles the cross-entropy, a somewhat enhanced measure of the discrepancy
of the predicated distribution from the true distribution. It is often used to
measure the performance of a model and used as one type of loss function.
We look at some examples.
178 Machine Learning with Python: Theory and Applications

4.6.1 Example 1: Binary cross-entropy for a distribution

of quality prediction

Consider a simple event with a variable that can take four possible values.
Assume that we have a good prediction of a distribution in relation to the
true or reference distribution. We examine how it is measured in the binary
cross-entropy using the following code:

# A good prediction case

import numpy as np
q = np.array([0.9,0.04,0.03,0.03]) # prediction
p = np.array([1.0,0.0,0.,0.]) # truth
p_conv = 1.0 - p # converse side of prediction
q_conv = 1.0 - q # converse side of truth
print(p, q, ' converse:',p_conv,q_conv)
cHpq = -np.sum(np.dot(p,np.log(q)))/len(p)
bcHpq = -np.sum(np.dot(p,np.log(q))+np.dot(p_conv,
np.log(q_conv)))/len(p)
print('Cross-entropy cHpq:',cHpq)
print('Binary cross-entropy bcHpq:',bcHpq)

[1. 0. 0. 0.] [0.9 0.04 0.03 0.03] converse: [0. 1. 1. 1.]

[0.1 0.96 0.97 0.97]
Cross-entropy cHpq: 0.02634012891445657
Binary cross-entropy bcHpq: 0.05177523128687465

It is found that the binary cross-entropy roughly doubles the cross-

entropy value, as expected.

4.6.2 Example 2: Binary cross-entropy for a poorly

predicted distribution

Consider an event with a variable that can take four possible values. Assume
that we have a poor prediction of a distribution in relation to the true
distribution. We examine again how it is measured in the binary cross-
entropy using the following code:

# A bad prediction case

q = np.array([0.4,0.2,0.3,0.1]) # prediction
p = np.array([1.0,0.0,0.,0.]) # truth
 Statistics and Probability-based Learning Model 179

p_conv = 1.0 - p # converse side of prediction

q_conv = 1.0 - q # converse side of truth
print(p, q, ' converse:',p_conv,q_conv)
cHpq = -np.sum(np.dot(p,np.log(q)))/len(p)
bcHpq = -np.sum(np.dot(p,np.log(q))+np.dot(p_conv,
np.log(q_conv)))/len(p)
print('Cross-entropy cHpq:',cHpq)
print('Binary cross-entropy bcHpq:',bcHpq)

[1. 0. 0. 0.] [0.4 0.2 0.3 0.1] converse: [0. 1. 1. 1.]

[0.6 0.8 0.7 0.9]
Cross-entropy cHpq: 0.22907268296853875
Binary cross-entropy bcHpq: 0.4003674356962309

It is also found that the binary cross-entropy roughly doubles the cross-
entropy, leading to an enhanced discrepancy measure.

4.6.3 Example 3: Binary cross-entropy for more uniform

true distribution: A quality prediction

In the previous two examples, we studied two cases with the true distribution
at extreme: its probabilities are 1.0 and zeros. For both examples, we
observed an enhanced entropy measure using the binary cross-entropy. In
this example, we consider a more even true distribution and examine the
behavior of the binary cross-entropy.

# A good prediction case

q = np.array([0.4,0.2,0.3,0.1]) # prediction
p = np.array([0.3,0.3,0.2,0.2]) # truth, rather even distribution
p_conv = 1.0 - p # converse side of prediction
q_conv = 1.0 - q # converse side of truth
print(p, q, ' converse:',p_conv,q_conv)
cHpq =-np.sum(np.dot(p,np.log(q)))/len(p)
bcHpq=-np.sum(np.dot(p,np.log(q))+np.dot(p_conv,
np.log(q_conv)))/len(p)
print('Cross-entropy cHpq:',cHpq)
print('Binary cross-entropy bcHpq:',bcHpq)

[0.3 0.3 0.2 0.2] [0.4 0.2 0.3 0.1] converse: [0.7 0.7 0.8
0. 8] [0.6 0.8 0.7 0.9]
180 Machine Learning with Python: Theory and Applications

Cross-entropy cHpq: 0.36475754318911824

Binary cross-entropy bcHpq: 0.5856092407474651

In this case, it is also found that the binary cross-entropy does not give
enhancement.

4.6.4 Example 4: Binary cross-entropy for more uniform

true distribution: A poor prediction

Same as the previous example, but consider a case with poor prediction.

# A bad prediction case

q = np.array([0.4,0.05,0.05,0.5]) # prediction
p = np.array([0.1,0.3,0.2,0.2]) # truth, rather even distribution
p_conv = 1.0 - p # converse side of prediction
q_conv = 1.0 - q # converse side of truth
print(p, q, ' converse: ',p_conv,q_conv)
cHpq =-np.sum(np.dot(p,np.log(q)))/len(p)
bcHpq=-np.sum(np.dot(p,np.log(q))+np.dot(p_conv,
np.log(q_conv)))/len(p)
print('Cross-entropy cHpq:',cHpq)
print('Binary cross-entropy bcHpq:',bcHpq)

[0.1 0.3 0.2 0.2] [0.4 0.05 0.05 0.5] converse: [0.9 0.7 0.
8 0.8] [0.6 0.95 0.95 0.5]
Cross-entropy cHpq: 0.43203116151910004
Binary cross-entropy bcHpq: 0.7048313483737685

In this case, we found a similarity: the binary cross-entropy gives an

enhancement.
In conclusion, our study on simple vents above shows that the binary
cross-entropy enhances the discrepancy measure, by taking into consider-
ation both positive samples (with probability close to 1) and “negative”
samples (with probability close to zero).

4.7 Bayesian Statistics

Consider a statistics event with more than one random variable that occurs
jointly. When we deal with such multiple random variables, we may want to
know the joint probability Pr(A,B): the probability of both A = a and B =
b occurring simultaneously, for given elements a and b.
 Statistics and Probability-based Learning Model 181

It is clear that for any values a and b, Pr(A,B) ≤ Pr(A = a), because
Pr(A = a) is measured regardless of what happens for B. For A and B to
happen jointly, A has to happen and B also has to happen (and vice versa).
Thus, A,B cannot be more likely than A or B occurring individually.
Pr(A,B)Pr(A) is called conditional probability and is denoted by Pr(B|A),
which is the probability that B happens, under the condition that A has
happened. This leads to the important Bayes’ theorem.

• By construction, we have: Pr(A,B) = Pr(B|A)Pr(A).

• By symmetry, this also holds: Pr(A,B) = Pr(A|B)Pr(B).
• We thus have
P r(A|B) = P r(B|A)P r(A)/P r(B). (4.9)

4.8 Naive Bayes Classification: Statistics-based Learning

4.8.1 Formulation
Based on the Bayesian statistics, a popular algorithm has been developed
known as the Naive Bayes Classifier. Consider an event with p variable
x = {x1 , x2 , . . . , xp } ∈ Xp . We assume that any variable xi is independent
of another. For a given label y, the conditional probability for being x is
expressed as 
p(x|y) = p(xi |y) (4.10)
i

Based on Bayes’ Theorem, we have the following formula:

p(x|y)p(y) p(xi |y)p(y)
p(y|x) = = i (4.11)
p(x) p(x)
Although we may not know p(x) (that is the probability that x occurs in the
event), it may not be needed because it is only a matter of normalization in
computing p(y|x). Thus, we may just use the following formula instead:

p(y|x) ∝ p(x|y)p(y) = p(xi |y)p(y) (4.12)
i

4.8.2 Case study: Handwritten digits recognition

We can now use the code provided at mxnet-the-straight-dope (https://
github.com/zackchase/mxnet-the-straight-dope/blob/master/chapter01
crashcourse/probability.ipynb) to show how a Naive Bayes classifier is coded
to identify handwritten digits. We will use the well-known MNIST dataset
182 Machine Learning with Python: Theory and Applications

to train this classifier. The MNIST (https://en.wikipedia.org/wiki/MNIST

database) contains a total of 70,000 images (60,000 for training and 10,000 for
testing) of handwritten 10 digits from 0 to 9, and all these images are labeled.
These images have been taken from American Census Bureau employees and
American high school students.
The digit classification problem is then casted to compute the probability
of a given image x being digit y: p(y|x). Any image x contains p pixels
xi (i = 1, 2, . . . , p), and each pixel xi can take a value of 1 (being lighted on)
or 0 (being lighted off), and hence is a binary variable.
Equation (4.12) can then be used, in which we need to estimate p(y)
and p(xi |y). Both can be computed using the MNIST training dataset for
each digit y. For example, in the total of 60,000 images of digits of the
MNIST training dataset, digit 4 is found 5,800 times, and we then have
5800
p(y = 4) = 60000 . To estimate p(xi |y), we can estimate p(xi = 1|y), because
xi is binary and p(xi = 0|y) = 1 − p(xi = 1|y). Estimating p(xi = 1|y) can
be done by counting the times that pixel i is on for label digit y, and then
dividing it by the number of occurrences of label y in the dataset. In this
simple algorithm, all we need is to count over the MNIST training dataset.
It is quite a straightforward strategy, the training is just counting, and we
can use the following code to get this done:

4.8.3 Algorithm for the Naive Bayes classification

# The codes are modified from these at https://github.com/

# zackchase/mxnet-the-straight-dope/blob/master/chapter01_
# crashcourse/probability.ipynb; Under Apache-2.0 License.
# import all the necessary packages
import numpy as np
import mxnet as mx
from mxnet import nd

def transform(data, label): # define a function to transfer data

return (nd.floor(data/128)).astype(np.float32),
label. astype(np.float32)
# floor of 255/128 = 1 pixel value = 1 when it is on.
# Divide dataset to 2 sets: one for training and one for testing
mnist_train = mx.gluon.data.vision.MNIST(train=True,
transform=transform)
mnist_test = mx.gluon.data.vision.MNIST(train=False,
transform=transform)
 Statistics and Probability-based Learning Model 183

print('type:',type(mnist_train))

type: <class 'mxnet.gluon.data.vision.datasets.MNIST'>

import matplotlib.pyplot as plt

%matplotlib inline
image_index=8888 # Check one. Any integer <60,000
# 8888 is digit with label 3, as printed below
print(mnist_train[image_index][1]) #image in 0; label in 1
plt.imshow(mnist_train[image_index][0].reshape((28, 28)).\
asnumpy(), cmap='Greys') #image pixel: 28 by 28

3.0

<matplotlib.image.AxesImage at 0x1c5743ba630>

Figure 4.9: One sample image of handwritten digit from the MNIST dataset.

# Initialize arrays for counts for computing p(y), p(xi|y)

# We initialize all numbers with a count of 1 to avoid
# division by zero, known as Laplace smoothing.

ycount = nd.ones(shape=(10)) #10 possible digits

xcount = nd.ones(shape=(784, 10)) #784 (= 28*28) variables
184 Machine Learning with Python: Theory and Applications

# Aggregate the count of the labels in training dataset

# and number of its corresponding pixels being on (value=1)
for data, label in mnist_train: # loop over the dataset
x = data.reshape((784,))
y = int(label) # get the digit-number
ycount[y] += 1 # add 1 to (digit)th entry
xcount[:, y] += x # add the image data to
# the (digit)th column
# compute the probabilities p(xi|y) (divide per pixel counts
# by total count of the label in the training dataset)
for i in range(10):
xcount[:, i] = xcount[:, i]/ycount[i]
# Compute the probability p(y)
py = ycount / nd.sum(ycount)

The model has been trained using the training dataset. We now plot the
“trained” model.
import matplotlib.pyplot as plt
%matplotlib inline
fig, figarr = plt.subplots(1, 10, figsize=(15, 15))

for i in range(10):
figarr[i].imshow(xcount[:,i].reshape((28,28)).asnumpy(),
cmap='hot')
figarr[i].axes.get_xaxis().set_visible(False)
figarr[i].axes.get_yaxis().set_visible(False)

plt.show()
np.set_printoptions(formatter={'float': '{: 0.3f}'.format})
print(py.asnumpy(),nd.sum(py).asnumpy())

Figure 4.10: A kind of mean appearance of handwritten digits.

[0.099 0.112 0.099 0.102 0.097 0.090 0.099 0.104

0. 098 0.099] [1.000]
 Statistics and Probability-based Learning Model 185

These pictures show the estimated probability distributions of observing

a switched-on pixel for all these 10 digits. These are the mean appearance of
these digits, or what a digit shall look like on average based on the training
dataset.

4.8.4 Testing the Naive Bayes model

We now examine the performance of this statistics-based model using the
MNIST test dataset. The training (which is just simple counting over
the training dataset) completed above gives us p(xi = 1|y) and p(y).
For a given image x from the test dataset, we compute the likelihood of
the image corresponding to a label y, which is to compute p(y|x) using
Eq. (4.12) where p(x|y) is in turn computed using the trained model. To
avoid chain multiplication of small probability numbers, we compute the
following logarithms instead (known as “log-likelihood”):

log p(y|x) ∝ log p(x|y) + log p(y) = log p(xi |y) + log p(y) (4.13)
i

For the given image x, a feature xi is binary and takes values of either 1
or 0. Because we are using the train model to compute the probabilities, we
shall have
p(xi = 1|y) = p(xi = 1|y) for predicting xi is on
(4.14)
p(xi = 0|y) = 1 − p(xi = 1|y) for predicting xi is off

Equation (4.14) can be written in a single one using a mathematical trick:

p(xi |y) = p(xi = 1|y)xi (1 − p(xi = 1|y))1−xi (4.15)

This is a general equation of computing the probability of an event with

binary variables using a trained model for predicting the probability of the
positive variable. We finally have
 
log p(xi |y) = [xi log p(xi = 1|y)+(1−xi ) log (1 − p(xi = 1|y))] (4.16)
i i

It is clear now that the testing is essentially measuring the binary cross-
entropy of the distribution of a given image (true distribution) with that
of the average image of a labeled letter computed from the dataset (the
model distribution). Therefore, we can write out Eq. (4.16) directly using
the binary cross-entropy formula.
186 Machine Learning with Python: Theory and Applications

To avoid re-computing the logarithms repetitively, we pre-compute

log p(y) for all y, and also log p(xi |y) and log (1 − p(xi |y)) for all pixels.

logxcount = nd.log(xcount) # pre-computations

logxcountneg = nd.log(1-xcount)
logpy = nd.log(py)
fig, figarr = plt.subplots(2, 10, figsize=(15, 3))
# test and show 10 images
ctr = 0 # initialize the control iterator
y = []
pxm = np.array([])
xi = ()
for data, label in mnist_test: # for any image
x = data.reshape((784,))
y.append(int(label))
# Incorporate the prior probability p(y) since p(y|x) is
# proportional to p(x|y) p(y)
logpx = logpy.copy() #nd.zeros_like(logpy)
for i in range(10):
# compute the log probability for a digit
logpx[i]+=nd.dot(logxcount[:,i],x)+nd.
dot(logxcountneg[:,i],1-x)

# normalize to prevent overflow or underflow by

# subtracting
# the largest value
logpx -= nd.max(logpx)
# and compute the softmax using logpx
px = nd.exp(logpx).asnumpy()
px = px*py.asnumpy() # this proportional to P(y|x)
px /= np.sum(px)
pxm = np.append(pxm,max(px)) # use the one with max Pr.
xi = np.append(xi,np.where(px == np.amax(px)))
# bar chart and image of digit
figarr[1, ctr].bar(range(10), px)
figarr[1, ctr].axes.get_yaxis().set_visible(False)
figarr[0, ctr].imshow(x.reshape((28,28)).asnumpy(),
cmap='hot')
 Statistics and Probability-based Learning Model 187

figarr[0, ctr].axes.get_xaxis().set_visible(False)
figarr[0, ctr].axes.get_yaxis().set_visible(False)
ctr += 1
if ctr == 10:
break
np.set_printoptions(formatter={'float': '{: 0.0f}'.format})
plt.show()
print('True label: ',y)
xi = np.array(xi)
print('Predicted digits:',xi)
print('Correct?',np.equal(y,xi))
np.set_printoptions(formatter={'float': '{: 0.1f}'.format})
print('Maximum probability:',pxm)

Figure 4.11: Predicted digits (in probability) using images from the testing dataset of
MINST.

True label: [7, 2, 1, 0, 4, 1, 4, 9, 5, 9]

Predicted digits: [7 2 1 0 4 1 4 9 4 9]
Correct? [True True True True True True True True False True]
Maximum probability: [1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0]

4.8.5 Discussion
The test shows that this classifier made one wrong classification for the first
10 digits in the testing dataset. The 9th digit should be 5, but classified
as 4. For this wrongly classified digit, the confidence level is very close to 1.
The wrong prediction may be due to the incorrect assumptions: each pixel
is independently generated, depending only on the label. Clearly, a digit is
a very complicated function of images, and statistic information alone has
a limit. This type of Naive Bayes classifier was popular in the 1980s and
1990s for applications such as spam filtering. For image processing types of
problems, we now have effective classifies (such as CNN; for example, see
Chapter 15).
188 Machine Learning with Python: Theory and Applications

Alternatively, we can use the cross-entropy or the binary cross-entropy

concept to perform the prediction. Once we obtained p(xi |yj ), j = 1, 2, . . . , 9
using the training dataset, one can then compute the (binary) cross-entropy
between p(xi |yj ) and p(xi |ytest ), where ytest is any given image from the test
dataset (or any other handwritten digit). The yj that gives the least (binary)
cross-entropy should be regarded as the predicted digit.
This simple example does show how statistical analyses are useful in
classifications and machine learning in general. We also showed how statistics
can be computed for given datasets. Powerful Naive classifiers can be
conveniently trained using the existing module at Sklearn (https://scikit-
learn.org/stable/modules/naive bayes.html) for practical problems that are
heavily governed by statistics and where the physics laws are unknown, such
as medical applications, recommendation systems, text classification, and
real-time prediction and recommendation.
 Chapter 5

Prediction Function and

Universal Prediction Theory

To build an ML model for predictions, one needs to use some hypothesis,

which predefines the prediction function to connect the feature variables to
the learning parameters. Thus, a proper hypothesis may follow the function
approximation theory, which has been studied intensively in physics-law-
based models [1–4]. The most essential rule is that the prediction function
must be capable of predicting an arbitrary linear function in the feature space
by a chosen set of learning parameters. Therefore, the prediction function is
assumed as a complete linear function of the feature variables, and it is
one of the most basic hypotheses.
It turns out such a complete linear prediction function performs affine
transformations of patterns in the affine space. Such a transformation
preserves affinity, meaning that the ratios of distances (Euclidean)
between points lying on a straight line and the parallelism of parallel line
segments remain unchanged after the transformation. It may not preserve the
angles between line segments and the (Euclidean) distances between points
in the original pattern. Further discussion on more general issues with affine
transformations can be found in Wikipedia (https://en.wikipedia.org/wiki/
Affine transformation) and the links therein.
The affinity ensures a special unique point-to-point seamless and gapless
transformation, meaning that it does not merge two distinct points to one, or
split one point to two, when the learning parameters are varying smoothly.
The connection between the feature variables and the learning parameters
is also smooth. Because an affine transformation is a combination of a linear
transformation and a translation controlled by the learning parameters, any
function up to the first order in the feature space can be reproduced, which
is critically important for machine learning models to be predictive.

189
190 Machine Learning with Python: Theory and Applications

This chapter discusses first the formulation and predictability of predi-

cation functions, followed by discussion on detailed process, properties, and
behavior of affine transformations. We shall focus on two aspects: (1) capabil-
ity in predicting functions in the feature space, and (2) affine transformation
of patterns in the affine space.
Then, the concept of affine transformation unit (ATU) (or linear predic-
tion function unit) is introduced as a building block, and simple neural net-
work codes will be built to perform affine transformations and demonstrate
their behavior and property. Feature encodings by learning parameters and
the uniqueness of the encodings are then studied, demonstrating the concept
of data-parameter converter. Next, an extension of ATU to form an affine
transformation array (ATA) and further extensions of the activation function
wrapped ATA to form MLPs or deepnets will be studied, which shows how
the predictability of a high-order nonlinear function can be established with
a deepnet. Finally, a Universal Prediction Theory is presented offering the
fundamental basis of why a deepnet can be made predictive.

5.1 Linear Prediction Function and Affine Transformation

Figure 5.1 shows the mechanism of the transmission of neurotransmitters in a

synaptic cleft of sensory neurons. Molecules of a neurotransmitter are shown
in a pseudo-colored image from a scanning electron microscope. A terminal
button (green) has been opened to reveal the synaptic vesicles (orange and
blue) inside the neurotransmitter.

(a) (c) (b)

Figure 5.1: (a) Transmission of neurotransmitters; (b) a pseudo-colored image of a

neurotransmitter from a scanning electron microscope: a terminal button (green) has been
opened to reveal the synaptic vesicles (orange and blue) inside; (c) a typical affine trans-
formation in xw formulation in an artificial NN. (Modified based on the image given in the
Psychology book (2d) by Rose M. Spielman et al. under the CC BY 4.0 License). (Credit b:
modification of work by Tina Carvalho, NIH-NIGMS; scale-bar data from Matt Russell).
 Prediction Function and Universal Prediction Theory 191

In an artificial NN, we use a so-called affine transformation between the

data-points and the learning parameters in an artificial neuron to somehow
mimic the information transformation process in a neurotransmitter. This
transformation is of most fundamental importance in many ML models, and
thus the related theory is discussed in great detail in this chapter.

5.1.1 Linear prediction function: A basic hypothesis

In machine learning models, the basic hypothesis is that the prediction
function z is given by the following equation:
 
b
z(ŵ; x) = x ŵ = [1 x] = xw + b
   w
x 
ŵ (5.1)
p

= W 1 x1 + · · · + W p xp + b = W i xi + b

i=1

where z(ŵ; x) reads as “z is a function of ŵ for given x”, and all vectors are
defined as
x = [x1 , x2 , . . . , xp ] ∈ Xp ∈ Rp

p
x = [1 x] = x0 , x1 , x2 , . . . , xp ∈ X ∈ Rp+1

1
(5.2)
w = [W 1 , W 2 , . . . , W p ] ∈ Wp ∈ Rp


ŵ = [b w] = W 0 , W 1 , W 2 , . . . , W p ∈ Wp+1 ∈ Rp+1

b

in which xi (i = 1, 2, . . . , p) are feature variables, and here we use only the

linear basis functions; W i (i = 1, 2, . . . , p) ∈ R are called weights. Constant
b ∈ R is called bias, and is also often denoted as W 0 ∈ R. These real numbers
form vectors in corresponding spaces (defined in Chapter 1) and used in the
operations above. The hat over w stands for the basis on top of it. It absorbs
the bias and become new vector in the hypothesis space Wp+1 . Both weights
and bias are parameters that can be tuned to predict exactly a desired
arbitrary linear function in the feature space. The relations of these spaces
are discussed in Chapter 1. Notice the transpose we used for the learning
parameters, which implies that they are formed in a matrix (with only one
column in this case). The features are formed in row vectors. This is why
the learning parameter matrix is acting from the right on the feature vector.
We intentionally put together most used formulas for the prediction
function in Eq. (5.1) a single united form, so that the relationship between all
192 Machine Learning with Python: Theory and Applications

these variables can be made clear once for all. Readers may take a movement
to digest this formulation, so that the later formulations can be understood
more easily.
When we write z = xw+b, we call it a xw+b formulation. When we write
z = x ŵ, in which the bias b is absorbed by w, we call it an xw formulation.
Both formulations will be used in the book interchangeably, because they are
essentially the same. The xw+b formulation allows explicit viewing the roles
of weights and biases separately during analysis. The xw formulation is more
concise in derivation processes, and also allows explicit expressions of affine
transformations, which are most essential for major machine learning models.

5.1.2 Predictability for constants, the role of the bias

Note that the bias b is a must have. If we set b = 0 and used only w, the
hypothesis will not be able to even predict a function that is a constant.
This can be easily proven as follows.
Consider a given (label) function y(x) = c, where c is a given constant
in R independent of x. This means that at x = 0, y(x = 0) = c. In order to
predict c using Eq. (5.1) we must have z(w, b; x = 0) = c. Now, if we drop
b in Eq. (5.1), regardless of what we choose for w, the hypothesis always
predicts

z =0·w =0 (5.3)

This means that the constant c ∈ R will never be predicted by the hypothesis
without b. This means also that a pure linear transformation through w is
insufficient for proper prediction, because it cannot even predict constants.
On the other hand, when b is there, we simply choose b = c, and the
constant c is then produced by the hypothesis. This implies also that z must
p
live in an affine space X , an augmented feature space that lives within Xp+1 .

5.1.3 Predictability for linear functions: The role of the weights

Further, by proper choices of w and b, any linear function can be produced
using Eq. (5.1). This can also be easily proven as follows.
Consider any given (label) linear function y ∈ Y of variables x ∈ Xp :

y(x) = xk + c (5.4)

where c ∈ R is a given constant and k is a (column) vector in Wp . Note k

in general may be in Rp , however, we need to perform vector operations on
it, it must be confined in the vector space Wp .
 Prediction Function and Universal Prediction Theory 193

By simply choosing w∗ = k, and b∗ = c, we obtain

z(w∗ , b∗ ; x) = xk + c = y(x) (5.5)

The given linear function is predicted exactly, using such a particular choice
of w∗ and b∗ . This means that any given arbitrary linear function of x ∈ Xp
can be predicted using hypothesis Eq. (5.1).

5.1.4 Prediction of linear functions: A machine

learning procedure
The above process showed that to predict a label linear function using
Eq. (5.1) is straightforward because we can choose these learning parameters
by inspection. For more complicated problems, this is not possible, and we
would use a minimization process to find these learning parameters. Here,
we demonstrate such a process to find w∗ and b∗ . This time, let us use the
xw formulation. We rewrite Eq. (5.4) to

y(x) = x k̂ (5.6)

where k̂ = [c k] ∈ Wp+1 ∈ Rp+1 .

Step 1: define a loss function in terms of the learning parameters. The loss
function shall evaluate the error between the hypothesis Eq. (5.1) and the
label function Eq. (5.6). It can have various forms. The widely used one is
the L2 error function that is the error squared:

L(z(ŵ)) = z(ŵ; x) − y(x))2 = [z(ŵ; x) − y(x)]2

= (x ŵ − x k̂) (x ŵ − x k̂) (5.7)
  
= (ŵ − k̂ )[x x](ŵ − k̂)

It is clear that the loss function is L(z) is a scalar function of the prediction
function z(ŵ) that is in turn a function of the vector of learning parameter
ŵ. Therefore, L(z) is in fact a functional. It takes a vector ŵ ∈ W(p+1) and
produces a positive number in R. It is also quadratic in ŵ.
In the 2nd line of Eq. (5.7), we first moved the transpose into the first
pair of parentheses, and then factor out x and x from these two pairs of
parenthesis to form matrix that is the out-product of [x x], which is a p × p
symmetric matrix of rank 1. All these follow the matrix operation rules. Note
that (ŵ − k̂) is a vector. Therefore, Eq. (5.7) is a standard quadratic form.
If [x x] is SPD, L has a unique minimum at (ŵ − k̂) = 0 or ŵ∗ = k̂. This
would prove that the prediction function is capable of reproducing any linear
194 Machine Learning with Python: Theory and Applications

function uniquely in the feature space, and we are done. However, because
[x x] has only rank 1, we need to manipulate a little further for deeper
inside.

Step 2: minimize the loss function with respect to ŵ ∈ Wp+1 .

This allow us to find ŵ that is the stationary point of L, by setting the

gradient to zero.

∂L(ŵ)
= 2x x(ŵ − k̂) = 0 (5.8)
∂ ŵ

In above, we used again that x x is symmetric. Regardless its contents,

Eq. (5.8) is satisfied when we set ŵ = ŵ∗ = k̂, which gives w∗ = k,
and b∗ = c. This is exactly the same as the results obtained previously.
This proves that the prediction function is capable of reproducing any linear
function in the feature space. Because x x is rank deficient (rank =1), the
solution of ŵ∗ = k̂ is not unique, and there are other (may infinite number
of) solutions of (ŵ − k̂) = 0 that satisfy Eq. (5.8). These solutions live in the
null-space of x x. This implies that we need to have more data points to
make the null-space zero, for unique solution. All these issues relate to the
solution existence theory that demands sufficient quality data-points,
which will be discussed in Chapter 9.
We have now solved analytically an ML problem using a typical mini-
mization procedure to predict a continuous function. This problem we just
examined is simple, but our analysis reveals essential issues in an ML model
using predictive functions. For more complicated problems with datasets of
discrete data-points, we would usually need computational means to solve
it, but the essential concept is the same.

5.1.5 Affine transformation

On the other hand, Eq. (5.1) can be used to perform an affine transfor-
mation, where weights wi (i = 1, 2, . . . , p) are responsible for (pure) linear
transformation and bias b is responsible for translation. Both wi and b are
learning parameters in a machine learning model. To show how Eq. (5.1) is
explicitly used to perform an affine transformation, we perform the following
maneuver in matrix formulations:
First, using each ŵi (i = 1, 2, . . . , k) and Eq. (5.1) we obtain,

zi = x ŵi (5.9)
 Prediction Function and Universal Prediction Theory 195

Now, form the following vector,

      
b1 b2 bk
z = [z1 , z2 , . . . , zk ] = x , x ,..., x
w1 w2 wk
   
b W0 (5.10)
=x =x or simply z = x Ŵ
W W
  
Ŵ

where b = [b1 , b2 , . . . , bk ], W0 = [ W 01 , W 02 , . . . , W 0k ] = b, and W =

  
b1 b2 bk
[w1 , w2 , . . . , wk ].
Notice that Ŵ absorbs b as the hat of W, and hence collects all the
learning parameters for predicting z. Our notation Ŵ allows easy tracking
the variables. It is often used for prediction at the output layer of a neural
network, because z = x Ŵ, and affine transformation is no longer needed at
the output layer.
With Eq. (5.10), we can further construct the following matrix opera-
tion [6].        
1 1 0 1 1 0 1
=  = or
z b W x W0 W x
              
z W
 x W
 x
   
  1 b  1 W0
1 z = 1 x = 1 x or simply (5.11)
      0 W    0 W
z x    x   
W W

z = xW

where 0 = [0, 0, . . . , 0] with p zeros. Matrix W we derived is the affine

transformation matrix. It has a dimension of (p + 1) × (k + 1) and
performs affine transformations an affine transformation from space Xp to
space Xk for given x. It is used in the hidden layers in neural networks,
because transformations must occur in the affine space in these layers to
ensure proper connections.
Matrix W will be used in Chapter 13 when studying MLPs or deepnets.
Note that W can be written as [e1 Ŵ], in which e1 = [1 0] is the first base
vector of space Wp+1 . W contains a unit vector (constant) as its column, and
thus the trainable parameters are in Ŵ. In the output layer of an NN, we use
only Ŵ, because affine transformation is no longer needed there. For machine
learning models, W can be called affine transformation weight matrix.
196 Machine Learning with Python: Theory and Applications

Formulation given above reveals clearly how the affine transformation

weight matrix is derived, and how the traditional weights and biases are
involved in the affine transformation weight matrix. Finally, we note the
following points:
p
1. We know that [1, x] is in X . Now, after action of W (from the right)
k
on it, the resulting [1, z] that is clearly also in X . This is known as the
automorphism property of an affine transformation. A pattern in an
affine space stays in an affine space after affine transformation. We will
demonstrate this in the example section.
2. The affine transformation uses only b and w given in Eq. (5.1), except
that it needs to be arranged in a form of W to properly act on x. This
may be the reason for Eq. (5.1) being often called affine transformation
(although not exactly at least in concept). Equation (5.1) is currently used
in the actual computations of affine transformation, including many parts
in this book. Our Eq. (5.11), however, allows more concise formulation,
and shows explicitly the automorphism. The last equation in Eq. (5.11)
and Eq. (5.10) can and should, of course, also be used for computation.
If we do, the codes and data structure may be even neater, because we
have only xw operations, and not b is involved.
3. For given W 0 (or b) and w, W is a linear operator. Weights w is responsible
for stretching-compression (scaling) and rotation, and W 0 is responsible
for translation. This property is shown as the affinity mentioned earlier.
We will discuss this further in the following sections and demonstrate this
in the example section.
4. For conveniences, we will use both xw formulation and xw+b formulation.

Let us look at some special cases, when Eq. (5.11) is used to perform affine
transformations.

Case 1: if we set all learning parameters to zero, Ŵ = 0, we will obtain

[1 z] = [1 0], meaning that any data-point [1 x] in an affine space
collapsed to the same point [1 0] in another affine space.
Case 2: if we set b = 0 and W = I where I is the identity matrix, we shall
have [1 z] = [1 x]. This means that any original point in affine
space is unchanged (no transformation).
Case 3: if we set b = c where c is a constant vector, and W = I, we shall
have [1 z] = [1 c + x]. This means that any original point in affine
space is translated by c.
 Prediction Function and Universal Prediction Theory 197

Case 4: if we set where b = [c, 0, . . . , 0], W = [k, e2 , . . . , ep ] where ei is a

base vector of Wp (with all zero entries except 1 at the ith entry),
we obtain zi = xi (i = 2, . . . , p) and

z1 = x k + c (5.12)

which is Eq. (5.5). This means that the prediction of a linear function in the
feature space can be viewed as an affine transformation in the affine space.
Since k in W is the gradient of the function, it is responsible for rotation.

5.2 Affine Transformation Unit (ATU), A Simplest Network

A p → 1 neural network can be built for predicting arbitrary linear functions

in the feature space, or to perform an affine transformations on the affine
space. The typical architectures are shown in Fig. 5.2.
This net can be set to predict arbitrary linear functions, or to perform
affine transformation defined in Eq. (5.1) for a given data-point xi (i =
1, 2, . . . , p) using sets of learning parameters wi (i = 1, 2, . . . , p) and b. Clearly,
any change in learning parameters shall result in a different value in function
z, for a given same data-point xi (i = 1, 2, . . . , p).
Note that Fig. 5.2 is a basic unit or a building block that can be used
to form a complicated neural network. Therefore, let us write a code to

Figure 5.2: A p → 1 neural network with one layer neurons taking an input of p features,
and one output layer of just one single neuron that produces a single prediction function
z. This forms an affine transformation unit or ATU (or linear function prediction unit).
The net on the left is for xw+b formulation, and on the right is for xw formulation with
p + 1 neurons in the input layer in which the one at the top is fixed as 1. Both ATUs are
essentially identical.
198 Machine Learning with Python: Theory and Applications

study it in great detail using the following examples. Let us first discuss the
data structures. Different ML algorithms may use a different one, and the
following one is quite typical.

5.3 Typical Data Structures

p → 1 nets:
Equation (5.1) can be written in the matrix form with dimensionality clearly
specified as follows:

z(ŵ; x) = x w + b = x ŵ (5.13)
1×p p×1 1×1 1×(p+1) (p+1)×1

p
The prediction function z ∈ X is now clearly specified as a function of w
and b corresponding to any x ∈ Xp . For the ith data-point xi , we have

z(ŵ; xi ) = xi w + b = xi ŵ (5.14)
1×p p×1 1×1 1×(p+1) (p+1)×1

Note that z(ŵ; x) is still a scalar for one data-point. Also, because no further
transformation, z is not needed for one layer nets.
p → k nets:
In hyperspace cases, we would have many, say k, neurons in the output
of the current layer, each neuron performing (independently) an affine
transformation based on the same dataset (see, Fig. 5.13). Therefore, the
output should be an array with k entries. The data may be structured in
matrix form: ⎡ ⎤
W 11 W 12 . . . W 1k
⎢W 21 W 22 . . . W 2k ⎥
⎢ ⎥
[z1 z2 · · · zk ] = [xi1 xi2 · · · xip ]⎢ . .. .. .. ⎥
z(W,b; x )
i 1×k xi1×p
⎣ .
. . . . ⎦
W p1 W p2 ... W pk
W p×k
+[b1 b2 · · · bk ] (5.15)
b 1×k

The above matrix can be written in a concise matrix form as follows, with
all the dimensionality specified clearly:

z(Ŵ; xi ) = xi W + b = xi Ŵ (5.16)
1×k 1×p p×k 1×k 1×(p+1) (p+1)×k

Note that one can stack up as many neurons in a layer as needed because
these weights for each neuron are independent of the those for any other
 Prediction Function and Universal Prediction Theory 199

neuron in the stack. This stacking is powerful because it makes the well-
known universal approximation theory (see Chapter 7) workable.
p → k nets with m data-points:
For a dataset with m points, the data may be structured as matrix Xm×p ,
by vertically stacking xi . In this case, m predictions can be correspondingly
made, and the formulation in matrix form becomes
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
z1 (W, b; x1 ) x11 x12 . . . x1p W 11 W 12 . . . W 1k b
⎢ z2 (W, b; x2 ) ⎥ ⎢ x21 x22 . . . x2p ⎥ ⎢W 21 W 22 . . . W 2k ⎥ ⎢b⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ .. ⎥ = ⎢ .. .. . . .. ⎥ ⎢ .. .. . . .. ⎥ + ⎢ .. ⎥ (5.17)
⎣ . ⎦ ⎣ . . . . ⎦ ⎣ . . . . ⎦ ⎣.⎦
zm (W, b; xm ) xm1 xm2 . . . xmp W p1 W p2 ... W pk b
Z m×1 (W,B;X) X m×p W p×k Bm×1

where vector B has the same b for all entries. The above matrix can be
written in a concise matrix form as follows, with all the dimensionality
specified clearly:

Z(W, B) = X W + B = X Ŵ (5.18)
m×1 m×p p×1 m×k m×(p+1) (p+1)×k

Note that we do not actually form matrix Z in practical computations,

because when a loss function is constructed, it is a form of a summation
over m or a mini-batch size. We will see this frequently in later chapters on
actual machine learning models.

5.4 Demonstration Examples of Affine Transformation

We now present a number of examples of affine transformations. This is

performed as follows.
For a given geometric pattern defined with a set of multiple data-
2
points X ∈ X (an affine space), its ith row is xi = [1, xi1 , xi2 ], computed
using Eq. (5.1):
zI = XŵI (5.19)

where ŵI = [bI , wI ] in which bI and wI are a given set of learning

parameters in the hypothesis space W3 , and

zII = XŵII (5.20)

where ŵII = [bII , wII ] in which bII and wII are a changed set of learning
parameters in W3 .
200 Machine Learning with Python: Theory and Applications

2
This results in a transformed data-point Z = [1, zI , zII ] ∈ X .
The above procedure using affine transformation on the original dataset
2
X ∈ X by varying ŵ for 2 times results in transformed dataset Z that is in
2
the same affine space X , the automorphism.
We now write a code to demonstrate the affinity of the above transforma-
2
tion. Because X is also a 2D plane, we can conveniently plot both original
and transformed patterns together in space R2 using only zI and zII for
visualization and analysis. We first define some functions.

import numpy as np

def logistic0(z): # The sigmoid/logistic function

return 1. / (1. + np.exp(-z))

def net(x,w,b): # An affine transformation net

y = np.dot(x,w) + b
return y

def edge(k,dd): # Define a line pattern in 2D

# feature space (x1, x2)
x = np.arange(-1.0,1.0+dd,dd) # dd: interval
x1 = x # x1 value.
x2 = k * x # x2 value, k slope of the line
x1 = np.append(x1,0) # add the center
x2 = np.append(x2,0)
X = np.stack((x1, x2), axis=-1) # X has two components.
return x1,x2,X

def circle(r,dpi): # Define a circular pattern in 2D

# feature space (x1, x2)
x = np.arange(0.0,2*np.pi,dpi)
x1 = r*np.cos(x) # circle function, for x1 value.
# Radius r and scaling factor c.
x2 = r*np.sin(x) # for x2 value.
x1 = np.append(x1,0) # add the center
x2 = np.append(x2,0)
X = np.stack((x1, x2), axis=-1) # X has two components.
return x1,x2,X

def rectangle(MinX, MaxX, MinY, MaxY,dd,theta):

# data for rectangle pattern in 2D space (x1, x2(y))
x = np.arange(MinX,MaxX+dd,dd)
ymin= np.full(x.shape,MinY) #y=x2
ymax= np.full(x.shape,MaxY)
 Prediction Function and Universal Prediction Theory 201

y = np.arange(MinY,MaxY+dd,dd)
xmin= np.full(y.shape,MinX)
xmax= np.full(y.shape,MaxX)

x1 = np.append(np.append(np.append(x,xmax),np.flip(x)),xmin)
x2 = np.append(np.append(np.append(ymin,y),ymax),np.flip(y))
x1 = np.append(x1,(MaxX+MinX)/2) # add the center
x2 = np.append(x2,(MaxY+MinY)/2)
X1 = x1*np.cos(theta)+x2*np.sin(theta)
X2 = x2*np.cos(theta)-x1*np.sin(theta)

X = np.stack((X1, X2), axis=-1) # X has two components.

return X1,X2,X

def spiral(alpha,c):
# Define a spiral pattern in 2D space (x1, x2)
xleft,xright,xdelta = 0.0, 40.01, 0.1
x = np.arange(xleft,xright,xdelta)
x1 = np.exp(alpha*x)*np.cos(x)/c # logarithmic spiral
# function, x1 with decay rate alpha & scaling factor c.
x2 = np.exp(alpha*x)*np.sin(x)/c # x2 value.
x1 = np.append(x1,0) # add the center
x2 = np.append(x2,0)
X = np.stack((x1, x2), axis=-1) # X has two components.
return x1,x2,X

Let us now set up learning parameters w and b.

# Set w0 & b0,so that original [x1,x2] pattern is reproduced.

w0_I=np.array([1.,0.]) # Initial 2 weights (vector),2D space
b0_I=0 # Initial bias, so that z_I = x1
w0_II=np.array([0.,1.]) # Initial 2 weights (vector),2D space
b0_II=0 # Initial bias, so that z_II = x2

# Set wI=w0_I, b_I=b0_I; but w_II, b_II be arbitrary.

#w_I, b_I = w0_I, b0_I # readers may try this
w_I=np.array([.8,.2]) # Arbitrary values for the two weights
# to perform scaling and rotation
b_I = -0.6 # Arbitrary values for the bias to
# perform translation
w_II=np.array([.2,.5]) # Arbitrary values for the two weights
# to perform scaling and rotation
b_II = 0.6 # Arbitrary values for the bias to
# perform translation
202 Machine Learning with Python: Theory and Applications

Next, we define a function for plotting these patterns, initial and affine
transformed with wII and bII , and linear transformed with wII and bII = 0.

%matplotlib inline
import matplotlib.pyplot as plt
def affineplot(x1,x2,X,w0_I,b0_I,w_I,b_I,w_II,b_II):
plt.figure(figsize=(4.,4.),dpi=90)
plt.scatter(net(X,w0_I,b0_I),net(X,w0_II,b0_II),label=\
"Original: w0I=["+str(w0_I[0])+","+str(w0_I[1])+
"], b0I="+str(b0_I)+"\n w0II=["+str(w0_II[0])+","+
str(w0_II[1])+"], b0II="+str(b0_II),s=10,c='orange')
#plot the initial pattern
plt.scatter(net(X,w_I,b_I),net(X,w_II,b_II),label=\
"Affine: wI=["+str(w_I[0])+","+str(w_I[1])+
"], bI="+str(b_I)+"\n wII=["+str(w_II[0])+","+
str(w_II[1])+"], bII="+str(b_II),s=10,c='blue')
# plot the affine transformed pattern
plt.scatter(net(X,w_I,b_I),net(X,w_II,b0_II),label=\
"Linear: wI=["+str(w_I[0])+","+str(w_I[1])+
"], bI="+str(b_I)+"\n wII=["+str(w_II[0])+","+
str(w_II[1])+"], b0II="+str(b0_II),s=10,c='red')
#plot the linear transformed pattern
plt.xlabel('$z_{I}$')
plt.ylabel('$z_{II}$')
plt.title('linear and affine transformation')
plt.grid(color='r', linestyle=':', linewidth=0.3)
plt.legend(loc='center left', bbox_to_anchor=(1, 0.5))
plt.axis('scaled')
#plt.ylim(-5,9)
plt.show()

5.4.1 An edge, a rectangle under affine transformation

We first create an edge (straight line segment) and a rectangle pattern
represented by a set of orange points, perform the affine transformations
defined by Eq. (5.1), and plot out these patterns before and after the affine
(blue) and linear (red) transformations. Because a rectangle consists of
straight lines, it is easy for us to observe the affinity of the transformation.

x1,x2,X = edge(1.5,0.2)
affineplot(x1,x2,X,w0_I,b0_I,w_I,b_I,w_II,b_II)
 Prediction Function and Universal Prediction Theory 203

Figure 5.3: Affine transformations of a straight line/edge.

x1,x2,X = rectangle(-2.,2.,-1.,1.,0.2,np.pi/4)
affineplot(x1,x2,X,w0_I,b0_I,w_I,b_I,w_II,b_II)

Figure 5.4: Affine transformations of a rectangle.

From the above two figures, the following observations can be made:

1. After the affine transformation, the original (orange) rectangular pattern

is only rotated, scaled, sheared, and translated to a new one (blue).
Weights w is responsible for the linear transformation results in scaling
and rotations, and b is responsible for translation.
204 Machine Learning with Python: Theory and Applications

2. The transformation moves point to point, edge to edge, and quadrilateral

to quadrilateral.
3. The transformation preserves the ratio of the lengths of parallel line
segments. For example, the ratio of the two longer sides of the orange
rectangle is the same as the ratio of the two longer sides of the blue
quadrilateral.
4. Parallel line segments remain parallel after the affine transformation.
5. It does not preserve distances between points. It preserves only the ratios
of the distances between points lying on a straight line.
6. The affine transformation does not preserve angles between lines.

This simple demonstration helps one to imagine how an affine transformed

pattern covers the (same) space by changing w and b. The pure linear
transformation alone does not change the origin, and hence has a much
limited coverage.

5.4.2 A circle under affine transformation

Let us now take a look at the affine (and linear) transformation to a circle
using the same code.

x1,x2,X = circle(1.0,0.1)
affineplot(x1,x2,X,w0_I,b0_I,w_I,b_I,w_II,b_II)

Figure 5.5: Affine transformations of a circle.

 Prediction Function and Universal Prediction Theory 205

This time, it is clearly seen from Fig. 5.5 that after the transformation, the
original (orange) circular pattern is rotated, scaled, sheared, and translated
to an ellipse (blue). The points that we have observed for the rectangles are
still valid.

5.4.3 A spiral under affine transformation

Let us examine the affine (and linear) transformation to a more complicated
pattern, spiral.

x1,x2,X = spiral(0.1,10.)
affineplot(x1,x2,X,w0_I,b0_I,w_I,b_I,w_II,b_II)

Figure 5.6: Affine transformations of a spiral.

In this case, the original (orange) spiral is rotated, scaled, sheared, and
translated to a new one. The observations made for the above two examples
still hold.

5.4.4 Fern leaf under affine transformation

Figure 5.7 shows an excellent example of affine transformation available
in the public domain provided by António Miguel de Campos (https://
en.wikipedia.org/wiki/Affine transformation). It is an image of a fractal of
Barnsley’s fern (https://en.wikipedia.org/wiki/Barnsley fern). A leaf of the
fern is an affine transformation of another by a combination of rotation,
scaling, reflection, and translation. The red leaf, for example, is an affine
206 Machine Learning with Python: Theory and Applications

Figure 5.7: An image of a leaf of the fern-like fractal is an affine transformation of another.

transformation of the dark blue leaf, or any of the light blue leaves. The fern
seems to have this typical pattern coded as an ATU in its DNA. This implies
that ATU is as fundamental as DNA.

5.4.5 On linear prediction function with affine transformation

One should not confuse the linear prediction function with the affine
transformation. They are essentially the same hypothesis, but viewed from
different aspects. The former is on the predictability of an arbitrary linear
function in the feature space using hypothesis Eq. (5.1), and the latter is on
affinity when Eq. (5.1) is used for pattern transformation on the affine space.
The predictability of the arbitrary linear function enables affine transforma-
tion. The predictability for constants allows translational transformation,
and the predictability for the linear function enables variable-wise scaling
and rotation, while maintaining the affinity. The prediction of a linear
function in the feature space can be viewed as an affine transformation in
the affine space. Because of this, we use these two terms interchangeably,
knowing this subtle difference.

5.4.6 Affine transformation wrapped with activation function

When an affine transformation z(w, b; x) is wrapped with a nonlinear activa-
tion function (see Chapter 7), the output φ(z) is confined by the activation
function, and the affinity is destroyed. However, φ(z) shall now have some
 Prediction Function and Universal Prediction Theory 207

capability to predict nonlinear functions, because the activation functions

used in ML are continuous, smooth (at least piecewise differentiable), and
vary monotonically with z.
As an example, we wrap the affine transformed pattern with the sigmoid
function. In this case, φ(z) is confined in (0, 1). We write the following code
to demonstrate some examples of affine mapping:

# Code for Affine transformation wrapped with sigmoid function

def sigmoidaffine(x1,x2,X,w0_I,b0_I,w0_II,b0_II,w_I,b_I,w_II,b_II):
# We do the same affine transformation and put the
# results to a sigmoid.
plt.figure(figsize=(4.5, 3.0),dpi=100)
plt.scatter(logistic0(net(X,w0_I,b0_I)),
logistic0(net(X,w0_II,b0_II)),label=\
"Original: w0I=["+str(w0_I[0])+","+str(w0_I[1])+
"], b0I="+str(b0_I)+"\n w0II=["+str(w0_II[0])+","+
str(w0_II[1])+"], b0II="+str(b0_II),s=10,c='orange')
#plot the initial pattern
plt.scatter(logistic0(net(X,w_I,b_I)),
logistic0(net(X,w_II,b_II)),label=\
"Affine: wI=["+str(w_I[0])+","+str(w_I[1])+
"], bI="+str(b_I)+"\n wII=["+str(w_II[0])+","+
str(w_II[1])+"], bII="+str(b_II),s=10,c='blue')
# plot the affine transformed pattern
plt.scatter(logistic0(net(X,w_I,b_I)),
logistic0(net(X,w_II,b0_II)),label=\
"Linear: wI=["+str(w_I[0])+","+str(w_I[1])+
"], bI="+str(b_I)+"\n wII=["+str(w_II[0])+","+
str(w_II[1])+"], b0II="+str(b0_II),s=10,c='red')
#plot the linear transformed pattern
plt.xlabel('$\sigma (z_{I})$')
plt.ylabel('$\sigma (z_{II})$')
plt.title('Affine transformation wrapped with sigmoid')
plt.grid(color='r', linestyle=':', linewidth=0.3)

plt.legend(loc='center left', bbox_to_anchor=(1, 0.5))

#plt.axis('scaled')
plt.ylim(-.05,1.1)
plt.show()

x1,x2,X = edge(5.0,0.1)
sigmoidaffine(x1,x2,X,w0_I,b0_I,w0_II,b0_II,w_I,b_I,w_II,b_II)
208 Machine Learning with Python: Theory and Applications

Figure 5.8: Nonlinear activation function wrapped affine transformations of a straight

line/edge.

x1,x2,X = rectangle(-2.,2.,-1.,1.,0.2,np.pi/4)
sigmoidaffine(x1,x2,X,w0_I,b0_I,w0_II,b0_II,w_I,b_I,w_II,b_II)

Figure 5.9: Nonlinear activation function wrapped affine transformations of a rectangle.

For this sigmoid wrapped affine transformation, the following observa-

tions can be made:
1. The transformation still sends a point to a point, an edge to an edge
uniquely. The use of the nonlinear activation function does not change
the uniqueness of the point-to-point transformation. This is because
the activation functions are continuous, smooth (at least piecewise
differentiable), and vary monotonically.
 Prediction Function and Universal Prediction Theory 209

2. The affinity is, however, destroyed: the ratios of distances between points
lying on a straight line are changed. Not all the parallel line segments
remain parallel after the sigmoid transformation. The use of the sigmoid
function clearly brings nonlinearity. This gives the net the following
capabilities:
• The output φ(z(ŵ; x)) is now nonlinearly dependent on the features x.
One can now use it for logistic regression for labels given 0, or 1, by
training ŵ.
• φ(z(ŵ; x)) is linearly independent of the features x in the input. This
allows further affine transformations to be carried out in a chain to the
next layer if needed.
• φ(z(ŵ; x)) is also linearly independent of the features ŵ used in this
layer. When we need more layers in the net, fresh ŵs can now be used
for the next layers independently. This enables the creation of deepnets.
Let us now take a look at the wrapped affine transformation to a circle
function, using the same code.
x1,x2,X = circle(2.5,0.1)
sigmoidaffine(x1,x2,X,w0_I,b0_I,w0_II,b0_II,w_I,b_I,w_II,b_II)

Figure 5.10: Nonlinear activation function wrapped affine transformations of a circle.

This case shows more server shape destroyed. The uniqueness of the point-
to-point transformation is still preserved. The following is for the wrapped
affine transformation for the spiral pattern.
x1,x2,X = spiral(0.1,10.)
sigmoidaffine(x1,x2,X,w0_I,b0_I,w0_II,b0_II,w_I,b_I,w_II,b_II)
210 Machine Learning with Python: Theory and Applications

Figure 5.11: Nonlinear activation function wrapped affine transformations of a spiral.

We see server distortions, due again to the nonlinearity of the nonlinear

activation function, sigmoid. The point-to-point transformation is still
observed, but when z is near 0.0 and 1.0, the original points and the
transformed points are “squashed” closer due to the sigmoid actions. Hence,
information near there is getting transmitted much less through updating of
these learning parameters w and b. In other words, the gradient gets close
to zero, due again to the properties of the sigmoid function.

5.5 Parameter Encoding and the Essential Mechanism

of Learning

5.5.1 The x to ŵ encoding, a data-parameter converter unit

Based on Eq. (5.1), we now see a situation where a dataset (x, z) can
be encoded in a set of learning parameters ŵ in the hypothesis space.
Figure 5.12 shows schematically such an encoded state.
The straight lines in Fig. 5.12 are encoded with a point in the hypothesis
space W2 . For example, the red line is encoded by a red dot at w0 = 1 and
w1 = 1. In other words, using w0 = 1 and w1 = 1, we can reproduce the red
line. The blue line is encoded by a blue dot at w0 = 2 and w1 = −0.5, with
which the blue line can be reproduced. The same applies to the black line. A
machine learning process is to produce an optimal set of dots using a dataset.
After that, one can then produce lines. In essence, a machine learning model
converts or encodes the data to wi . This implies that the size and quality of
 Prediction Function and Universal Prediction Theory 211

Figure 5.12: Data (on relations of x-z or x-y for given labels) encoded in model
parameters ŵ in the hypothesis space. In essence, a ML model converts data to ŵ during
training.

the dataset are directly related to the dimension of the affine spaces used in
the model.
On the other hand, if one tunes wi , different prediction functions can be
produced in the label space. Therefore, it is possible to find such a set of
wi that makes the prediction match the given label in the dataset for given
data-points. Finding such a set of wi is the process of learning. Real machine
models are a lot more complicated, but this gives the essential mechanism.

5.5.2 Uniqueness of the encoding

We state that the encoding of a line in the X-z space to a point in the
hypothesis space is unique. It is very easy to prove as follows.
Assume any arbitrary line in the X-z space that has two distinct points
in the hypothesis space ŵ(1) and ŵ(2) ; using Eq. (5.9), this line can be
expressed as
z = xŵ(1) (5.21)

which holds for arbitrary x. This same line can also be expressed as

z = xŵ(2) (5.22)

which holds also for arbitrary x. Using now Eqs. (5.21) and (5.22), we obtain

0 = x[ŵ(2) − ŵ(1) ] (5.23)

212 Machine Learning with Python: Theory and Applications

Equation (5.23) must also hold for arbitrary x. Therefore, we shall have

ŵ(1) = ŵ(2) (5.24)

which completes the proof.

On the other hand, we also state that a point in the hypothesis space
gives a unique line. It can be easily proven as follows.
Given any arbitrary point in the hypothesis space ŵ, assume we can
construct two lines in the X-z space. Using Eq. (5.9), the first line can be
expressed as
z (1) = xŵ (5.25)

The second line can be expressed as

z (2) = xŵ (5.26)

Using Eqs. (5.25) and (5.26), we obtain

z (2) − z (1) = 0 (5.27)

This means that these two lines are the same, which completes the proof.
In fact, the uniqueness can be clearly observed from Fig. 5.12, because a
line is uniquely determined by its slope and bias, and both are given by ŵ.
The uniqueness is one of the most fundamental reasons for a quality
dataset to be properly encoded with the learning parameters based on the
hypothesis of affine transformations, or for a machine learning model to be
capable of reliably learning from data.

5.5.3 Uniqueness of the encoding: Not affected

by activation function

It is observed that the uniqueness of the encoding of an affine transformation

wrapped with an activation function is not affected by the activation
function. This is because activation functions are all strictly monotonic
functions (as will be shown in Chapter 7), which does not change the
uniqueness of its argument. The proof is thus essentially the same as that
given above.
On the other hand, this property implies that the activation function
must be monotonic. This is true for all the activation functions discussed in
Chapter 7.
 Prediction Function and Universal Prediction Theory 213

5.6 The Gradient of the Prediction Function

The gradient of the prediction function with respect to the learning

parameters has the following simple forms:
∇w z = x
(5.28)
∇b z = 1
which shows that the gradient with respect to weights relates the feature
variable (the data). The gradient with respect to bias is, however, a unit.
This is because the data corresponding to bias are x0 = 1. This may suggest
that when a regularization technique (see Chapter 14) is used, one may
choose different regularization parameters for weight and bias.
When the xw formulation is used, we have

∇ŵ z = x (5.29)

which can be used in an autograd in machine learning processes.

5.7 Affine Transformation Array (ATA)

We can now duplicate vertically the single neuron on the right in Fig. 5.2
for k times and let all neurons become densely connected, meaning that
each neuron in the output is connected with each of the neurons in the input
(also known as fully connected). This forms a mapping of p → k network.
The prediction functions z from an ATA can be expressed as

z(Ŵ; x) = xW + b = x Ŵ (5.30)

where z is a vector of prediction functions given by

z(Ŵ; x) = [z1 (ŵ1 ; x), z2 (ŵ2 ; x), . . . , k (ŵk ; x)], with
ŵj = [W 0j , W 1j , W 2j , . . . , W pj ] with W 0j = bj are the (p + 1) learning
parameters for the jth neuron in the output layer,
Ŵ = [b W] is a matrix of (p + 1)×k containing all learning parameters
for all neurons in the output layer,
W = [W ij ], (i = 1, 2, . . . , p; j = 1, 2, . . . , k) is a matrix of weights (part of
the learning parameters), and
b = [b1 , b2 , . . . , bk ] is a vector of biases (part of the learning parameters).
The vector ŵ of total learning parameters of a p → k ATA becomes,

ŵ = [ŵ1 , ŵ2 , . . . , ŵk ]

214 Machine Learning with Python: Theory and Applications

Figure 5.13: A p → k neural network with one input layer of p neurons and one output
layer of k neurons that produces k prediction functions zi (i = 1, 2, . . . , k). Each neuron
at the output connects to all the neurons in the input with its own weights. This stack
of ATU forms an affine transformation array or ATA. In other words, the p → k net
has the predictability of k functions in the feature space with p dimensions. Left: xw+b
formulation; Right: xw formulation.

which can also be regarded as the flattened Ŵ. The total number of the
learning parameters is

P = (p + 1) × k

It is clear that the hypothesis space grows fast in multiples for an ATA.
Equation (5.30) is the matrix form of a set of affine transformations.
It is important to note that each zj (wj , bj ) is computed using Eq. (5.1),
using its own weights wj and bias bj . This enables all zj (wj , bj ), j =
1, 2, . . . , k being independent of each other. Therefore, the ATA given in
Fig. 5.13 creates the simplest mapping that can be used for k-dimensional
regression problems using a dataset with p features. Note also that when
k = p, it can perform the p → p affine transformation.

5.8 Predictability of High-Order Functions of a Deepnet

5.8.1 A role of activation functions

Now, we wrap the stack of prediction functions with nonlinear activation
functions. This leads to a vector of
⎡ ⎤
⎢ ⎥
⎢φ(z1 (w1 , b1 )), φ(z2 (w2 , b2 )), . . . , φ(zk (wk , bk ))⎥ (5.31)
⎣        ⎦
(new) (new) (new)
x1 x2 xk
 Prediction Function and Universal Prediction Theory 215

(new)
It becomes a set of new features xi , i = 1, 2, . . . , k that are linearly
independent of the original features xi , i = 1, 2, . . . , p. These new features
can then be used as inputs to the next layer. This allows the use of a new
set of learning parameters for the next layer.
It is clear that a role of the activation function is to force the outputs from
an ATA linearly independent of that of the previous ATA, enabling further
affine transformations leading to a chain of ATAs, a deepnet. To fulfill this
important role, the activation function must be nonlinear.
Such a chain of stacks of prediction functions (affine transformations)
wrapped with nonlinear activation functions gives a very complex deepnet,
resulting in a complex prediction function. Further, when affine transfor-
mation Eq. (5.1) is replaced with spatial filters, one can build a CNN
(see Chapter 15) for object detection, and when replaced with temporal
filters, we may have an RNN (see Chapter 16) for time sequential models,
and so on.

5.8.2 Formation of a deepnet by chaining ATA

These new features given in Eq. (5.31) can now be used as the inputs for the
next layer to form a deepnet. To illustrate this more clearly, we consider a
simplified deepnet with 4 − 2 − 3 neurons shown in Fig. 5.14.

Figure 5.14: Schematic drawing of a chain of stacked affine transformations wrapped with
activation functions in a deepnet for approximation of high-order nonlinear functions of
high dimensions. This case is an xw+b formulation. A deepnet using xw formulation will
be given in Section 13.1.4.
216 Machine Learning with Python: Theory and Applications

Here, let us use the number in parentheses to indicate the layer number:
(1)
1. Based on 4 (independent input) features xi (i = 1 ∼ 4) to the first layer,
(1)
a stack of 2 affine transformations zi (i = 1 ∼ 2) takes place, using a
(1)
4 × 2 weight matrix W(1) and biases bi (i = 1 ∼ 2). Affine transformation
(1) (1) (1) (1) (1) (1)
z1 uses wi1 (i = 1 ∼ 4) and b1 , and z2 uses wi2 (i = 1 ∼ 4) and b2 .
Clearly, these are carried out independently using different sets of weights
and biases.
(1) (1)
2. Next, z1 and z2 are, respectively, subjected to a nonlinear activation
(2)
function φ, producing 2 new features xi (i = 1 ∼ 2). Because of the
(2)
nonlinearity of φ, xi will no longer linearly depend on the original
(1)
features xi (i = 1 ∼ 4).
(2)
3. Therefore, xi (i = 1 ∼ 2) can now be used as independent inputs for
the 2nd layer of affine transformations, using a 2 × 3 weight matrix W(2)
(2)
and biases bi (i = 1 ∼ 3), in the same manner. This results in a stack
(3)
of 3 affine transformations zi (i = 1 ∼ 3), which can then be wrapped
again with nonlinear activation functions. This completes the 2nd layer
of 3 stacked affine transformations in a chain.

The above process can continue as desired to increase the depth of the neural
network. Note also that the number of neurons in each layer can be arbitrary
in theory. Because of the stacking and chaining, the hypothesis space is
greatly increased. The stacking causes the increase in multiples, and the
chaining in additions. The prediction functions may live in an extremely high
dimensional space WP for deepnets. For this simple deepnet of 4 − 2 − 3, the
dimension of the hypothesis space becomes P = (4× 2+ 2)+ (2× 3+ 3) = 19.
In general, for a net of p − q − r − k, for example, the formulation should be

P = (p × q + q) + (q × r + r) + (r × k + k) (5.32)
        
layer 1 layer 2 layer 3

The vector of all trainable parameters in an MLP becomes,


ŵ = Ŵ(1) .f latten(), Ŵ(2) .f latten(), . . . , Ŵ(NL ) .f latten()

where NL is the total number of hidden layers in the MLP. Note that we
may not really perform the foregoing flattening in actual ML models. It is
just for demonstrating the growth of the dimension of the hypothesis space.
In actual computations, we may simply group them in a Python list, and
use an important autograd algorithm to automatically perform the needed
 Prediction Function and Universal Prediction Theory 217

Figure 5.15: 1 → 1 → 1 net with sigmoid activation at the hidden and last layers.

forward and backward computations in training an MLP. The computations

over such a high dimension is achieved numerically. This will be discussed
in detail in Chapter 8.
The prediction functions are now functions of ŵ in a high dimensional
hypothesis space WP for MLPs. The formulation on calculating P for more
general MLPs will be given in Chapter 13. The Neurons-Samples Theory
that gives the relationship between the number of neurons and the number
of data-points will be discussed in Section 13.2.1.

5.8.3 Example: A 1 → 1 → 1 network

Consider the simplest 1 → 1 → 1 neural network shown in Fig. 5.15. Let us
use the same linear prediction function Eq. (5.1) and a sigmoid activation
function for the hidden and last layers. In this case, the output at the last
layer x(3) can be obtained as follows:
    1
x(3) = σ z (2) = σ w(2) x(2) + b(2) = (2) (2) (2)
(5.33)
1 + e−(w x +b )

where x(2) is the output from the hidden layer:

    1
x(2) = σ z (1) = σ w(1) x(1) + b(1) = (1) x+b(1) )
(5.34)
1+e−(w

in which x(1) = x, which is the inputs that can be normalized to be in

(−1, 1). The number in the parenthesis in the superscripts stands for the
layer number. Because x(2) is in (0, 1), we next use the Taylor expansion
consecutively twice to approximate the sigmoid function; we obtain

x(3) = c0 + c1 x + c2 x2 + c3 x3 + · · · (5.35)

where these constants are given, through a lengthy but simple derivation, by
 
1 (1) (2) (1) (1) (2) (1)

1 (2) (1)  (2) 2  (1) 2
c0 = − b b w b −b w − b w b w − 12
16 48
 
1 (1)  (1) 2
− b b − 12
48
218 Machine Learning with Python: Theory and Applications

 
1 (1) (2)  (1) 2 (1) (2) (1)
 2 
(2) (1)
2
c1 = − w w b + 2b b w + b w −4 (5.36)
16
1  (1) 2  (2) 2  (1) 
c2 = − w w b + b(2) w(1)
16
1  (1) 3  (2) 3
c3 = − w w
48

It is clear from Eq. (5.36) that by properly setting (training using a

dataset) the weights and biases for the neurons in the hidden and output
layers, all these constants ci , i = 1, 2, 3, 4 can be determined. The prediction
function at the output x(3) becomes a 3rd-order polynomial of the feature
x as given in Eq. (5.35). This means that our 1 → 1 → 1 net has the
predictability for 3rd-order functions approximately, in contrast to a 1 → 1
net that is only capable for 1st-order functions.
Note that the same analysis can be performed for other types of activation
functions. Also, if we cut off higher-order terms in the Taylor series, we can
approximate even higher-order functions. The point of our discussion here
is not about how well to approximate the sigmoid function, but to show the
capability of the simple 1 → 1 → 1 net. This simple analysis supports a
very important fact that adding layers gives the net the capacity to predict
higher-order nonlinear latent behavior. This is the reason why a deepnet is
powerful if it can be effectively trained.
We note, without further elaboration, that increasing the depth of a net is
equivalent to increasing the order of the shape functions in the physics-law-
based models such as the FEM [1] or the meshfree methods [3]. In contrast,
increasing the number of neurons in a layer is equivalent to increasing the
number of the elements or nodes [5].

5.9 Universal Prediction Theory

A deepnet can be established with the following important properties:

1. The capability of the linear prediction function (affine transformation)

in predicting exactly any function up to the first order, and one-to-one
unique transformation (with or without activation function) in each ATU.
2. The independence of the function approximation (affine transformation)
of a neuron to another in an ATA (due to its independent connections).
3. New independent features are produced in each layer using nonlinear
activation functions for each ATA.
 Prediction Function and Universal Prediction Theory 219

4. Chaining of ATA wrapped with nonlinear activation functions provides

the capability for predicting complex nonlinear functions to an arbitrarily
higher order.

In the opinion of the author, these are the fundamental reasons for various
types of deepnets being capable of creating p → k mappings for extremely
complicated problems from p inputs of features to k labels (targeted features)
existing in the dataset. We now summarize our discussion to a Universal
Prediction Theory [6].

Universal Prediction Theory: A deepnet with sufficient layers of suffi-

cient neurons wrapped with nonlinear activation functions can be established
for predictions of latent features existing in a dataset when properly trained.

This theory claims only the capability of a deepnet in terms of creating

giant prediction functions based on the dataset. How to realize its capability
requires a number of techniques, including how to set up a proper structure
of the deepnet for given types of problems and how to find these optimal
learning parameters reliably and effectively. In addition, the applicability of
the trained MLP model depends on the quality of the dataset, as mentioned
in Section 1.5.6. The dataset quality is defined as its representativeness
to the underlaying problem to be modeled, including correctness, size, data-
point distribution over the features space, and noise level.

5.10 Nonlinear Affine Transformations

Note that in the above formulations, the features xi , i = 1, 2, . . . , p are used

in an affine transformation as linear basis functions. However, the basis func-
tions do not have to be linear. Take a one-dimensional problem, for example;
when linear approximation is used, the vector of the features should be

x = [1, x] (5.37)

If one would like to use 2nd-order approximation (often times called

nonlinear regression), the vector of the features simply becomes

x = 1, x, x2 (5.38)

If one knows the dataset well and believes that a particular function can be
used as a basis function, one may simply add it as an additional feature. For
220 Machine Learning with Python: Theory and Applications

example, one can include sin(x) as a feature in the following form:

x = [1, x, sin(x)] (5.39)

The use of nonlinear functions as bases for features is also related to the
so-called support vector machine (SVM) models that we will discuss in
Chapter 6, where we use kernel functions for linearly un-separable classes.
This kind of nonlinear feature basis or kernel is sometimes called feature
functions.
In our neural network models, higher-order and enrichment basis func-
tions can also be used in higher dimensions. For example, for two-dimensional
spaces, we may have features like

x = 1, x1 , x2 , x1 x2 , x21 , x22 , sin(x) (5.40)

This has completed 2nd-order polynomial bases functions, and is enriched

with a sin(x) function (with proper scaling on x). The dataset X shall also
be arranged in the order of all these features. The affine transformation
using the nonlinear basis functions can be exactly the same as using linear
basis discussed in this chapter. In fact, we have already done the affine
transformations for the circle and spiral. Note that for high-dimensional
problems, the feature space with nonlinear bases can be in extremely higher
dimensions. For such cases, the so-called kernel trick may apply to avoid
dimension increase.

5.11 Feature Functions in Physics-Law-based Models

This concept of using higher-order or special feature functions is essentially

the same as in the physics-law-based computational methods. In these
physics-law-based methods, the feature functions are called basis func-
tions. These basis functions are used to approximate the field variables
(displacements, stress, velocity, pressure, etc.) that are governed by a physic-
law in either strong or weak forms.
For example, in the finite element approximation [1], and the smoothed
finite element methods [2], we frequently use higher-order polynomial bases
called higher-order elements. Special basis functions are also used but called
enrichment functions. For example, when we would like to capture the
√
singular stress field in the domain, we add in r in the bases [2]. In the
meshfree methods [3], one can also use distance basis functions, such as
the radial basis functions (RBFs).
 Prediction Function and Universal Prediction Theory 221

In methods used to solve linear mechanics problems governed by physics

laws, we often use high-order and special basis functions. The resulting
system equations will still be linear in the field variables. The essential
concept is the same: to capture necessary features in the system (governed by
law or hidden in data), one shall use feature or basis functions of necessary
complexity. The resulting models may still be linear in field variables.

References
[1] G.R. Liu and S.S. Quek, The Finite Element Method: A Practical Course, Butterworth-
Heinemann, London, 2013.
[2] G.R. Liu and T.T. Nguyen, Smoothed Finite Element Methods, Taylor and Francis
Group, New York, 2010.
[3] G.R. Liu, Mesh Free Methods: Moving Beyond the Finite Element Method, Taylor and
Francis Group, New York, 2010.
[4] G.R. Liu and Gui-Yong Zhang, Smoothed Point Interpolation Methods: G Space Theory
and Weakened Weak Forms, World Scientific, London, 2013.
[5] G.R. Liu, A Neural Element Method, International Journal of Computational Methods,
17(07), 2050021, 2020.
[6] G.R. Liu, A thorough study on affine transformations and a novel Universal Prediction
Theory, International Journal of Computational Methods, 19(10), in-printing, 2022.
MACHINE LEARNING
WITH PYTHON
 Chapter 6

The Perceptron and SVM

This chapter discusses two fundamentally important supervised machine

learning algorithms, the Perceptron and the Support Vector Machine (SVM),
for classification. Both are conceptually related, but very much different in
formulation and algorithm. In the opinion of the author, there are a number
of key ideas used in both classifiers in terms of computational methods. These
ideas and the resulting formulations are very inspiring and can be used for
many other machine learning algorithms. We hope the presentation in this
chapter can help readers to appreciate these ideas. The referenced materials
and much of the codes are from the Numpy documentations (https://
numpy.org/doc/), Scikit-learn documentations (https://scikit-learn.org/
stable/), mxnet-the-straight-dope (https://github.com/zackchase/mxnet-
the-straight-dope), Jupyter Notebook (https://jupyter.org/), and Wikipedia
(https://en.wikipedia.org/wiki/Main Page).
Our discussion starts naturally from the Perceptron. It is one of the
earliest machine learning algorithms by Frank Roseblatt in 1957 [1, 2].
It is used for problems of binary classification and was one of the well-studied
problems of classification in 1960s. Here, we first introduce the mathe-
matical model of a typical classification problem with the related formulation
and its connection to affine transformation. We then examine in detail
Coli’s Perceptron algorithm that is made available at mxnet-the-straight-
dope (https://github.com/zackchase/mxnet-the-straight-dope) (under the
Apache License 2).
The discussion on the Perceptron is naturally followed by that on SVM
[3, 4, 9]. A complete description of the SVM formulation is provided,
including detailed process leading to a quadratic programming problem, the
kernel trick for linearly inseparable datasets, as well as the concept of affine
transformation used in SVM.

223
224 Machine Learning with Python: Theory and Applications

6.1 Linearly Separable Classification Problems

Let us consider the following problem. Given a set of m data-points with p

features, the ith data-point is noted as xi = {xi1 , xi2 , . . . , xip , } ∈ Xp with
its corresponding label yi ∈ {±1} (meaning the label can be either +1 or −1
for a data-point). Such labels distinguish these data-points in two classes:
positive points and negative ones. We assume that this set of data-points
is linearly separable, meaning that these data-points can be separated into
these two distinct classes using a hyperplane (a line in higher-dimensional
space, Xp ).
Consider a simple problem with only 2 features, x1 and x2 , in a two-
dimensional (2D) space x = {x1 , x2 } ∈ X2 , so that we can have a good
visualization. A set of m data-points is scattered in 2D space shown in
Fig. 6.1.
All these data-points are labeled into two classes: positive points marked
with “+” symbols, and each of the points is labeled with y = +1; negative
points marked with “−” symbols, and each of them is labeled with y = −1.
These two classes of points can be separated by straight lines, such as the
red-dashed line and the red-dotted line in Fig. 6.1. For datasets in the real
world, there are infinite numbers of such lines forming a street. Our goal
is to develop a computer algorithm to find such a line with a given labeled
dataset (the data-points with corresponding labels). This is simple but quite
a typical classification problem.
Assume for the moment that we know the orientation of one such red
line, say the middle red-dashed line for easy discussion. Hence, we know its
normal direction vector w = [w1 , w2 ] ∈ W2 , although we do not yet know

Figure 6.1: Linearly separable data-points in 2D space.

 The Perceptron and SVM 225

the translational location of the red-dashed line along its normal. We then
w
have the unit normal vector as w with a length of 1. For any point (not
necessarily the data-point) in the 2D space (marked with a small cross in
Fig. 6.1), we can form a vector x starting at the origin. Now, the dot-product
w
x· (6.1)
w
w
becomes the length of the projection of x on the unit normal w . Therefore,
it is the measure that we need to determine how far point x is away from the
w
origin in the direction of w , which is a useful piece of information. Because
we do not yet know the translational location of the red line in relation to x,
b
we thus introduce a parameter w , where b ∈ W1 is an adjustable parameter
to allow the red line move up and down along w.
Notice in Eq. (6.1) that we used dot-product (the inner product). This
is the same as the matrix-product we used in the Python implementation,
because their shapes match: x is a (row) vector, and w is column vector (a
matrix with single column) with the same length. Therefore, we use both
interchangeably in this book.
The equation for an arbitrary line in relation to given point x shall have
this form:
x·w+b (6.2)

Our task now is to find the red-dashed line by choosing a particular

set of w and b in W3 that separates the data-points, which is the affine
transformation discussed in Chapter 5. The conditions should be

(x · w + b) > 0, for points in upper-right side of the red-dash line: y = +1

(6.3)
(x · w + b) < 0, for points in lower-left side of the red-dash line: y = −1
(6.4)

Note that Eqs. (6.3) and (6.4) are for ideal situations where these two sets of
points might be infinitely close. In practical applications, we often find that
these points are in two distinct classes, and they are separated by a street
with a finite width w (that may be very small). The formulation can now be
modified as follows:

(x · w + b) > w/2, for points in upper-right side of the red-dash line: y = + 1

(6.5)
(x · w + b) < w/2, for points in lower-left side of the red-dash line: y = −1
(6.6)
226 Machine Learning with Python: Theory and Applications

This type of equation is also known as the decision rule: when the condition
is satisfied by an arbitrary point x, it then belongs to a labeled class (y = 1
or y = −1), when the parameters w and b are known. We made excellent
progress.
It is obvious that Eqs. (6.5) and (6.6) can be magically written in a single
equation by putting these two conditions together with their corresponding
labels y.

y(x · w + b) > w/2 or y(x · ŵ) > w/2 or mg > w/2 (6.7)

This is a simplified single-equation decision rule: when the condition is

satisfied by an arbitrary point, it belongs to the labeled domain, and is not
within the street. This single equation is a lot more convenient for developing
the algorithm called a classifier to do the task. Note that mg is called
margin, which will be formally discussed in detail in Chapter 11.
We need to now bring in a labeled dataset to find w and b using the
above decision rule. For any data-point (the ith, for example, and regardless
of which class it belongs to), it must satisfy the following equation:

yi (xi · w + b) > w/2 or yi (xi · ŵ) > w/2 or mg(i) > w/2 (6.8)

where mg(i) is the margin for the ith data-point. Because there are an
infinite number of lines (such as the red-dashed and red-dotted lines shown in
Fig. 6.1) for such a separation, there exist multiple solutions to our problem.
We just want to find one of them that satisfies Eq. (6.8) for all data-points
in the dataset. This process is called training. Because labels are used, it is
a supervised training. The trained model can be used to predict the class of
a given data-point (which may not be from the training dataset), known as
classification or prediction in general. The following is an algorithm to
perform all those: training as well as prediction.

6.2 A Python Code for the Perceptron

The following is an easy-to-follow code available at mxnet-the-straight-

dope (https://github.com/zackchase/mxnet-the-straight-dope), under the
Apache-2.0 License. We have modified it a little and added in some detailed
descriptions as comment lines with the code.
Let us examine the details in this algorithm. As usual, we import the
needed libraries.
 The Perceptron and SVM 227

import mxnet as mx
from mxnet import nd, autograd
import matplotlib.pyplot as plt
import numpy as np
# We now generate a synthetic dataset for this examination.
mx.random.seed(1) # for repeatable output of this code
# define a function to generate the dataset that is
# separable with a margin strt_w
def getfake(samples, dimensions, domain_size, strt_w):
wfake = nd.random_normal(shape=(dimensions)) # weights
bfake = nd.random_normal(shape=(1)) # bias
wfake = wfake / nd.norm(wfake) # normalization
# generate linearly separable data, with labels
X = nd.zeros(shape=(samples, dimensions)) # initialization
Y = nd.zeros(shape=(samples))
i = 0
while (i < samples):
tmp = nd.random_normal(shape=(1,dimensions))
margin = nd.dot(tmp, wfake) + bfake
if (nd.norm(tmp).asscalar()<domain_size) & \
(abs(margin.asscalar())>strt_w):
X[i,:] = tmp[0]
Y[i] = 1 if margin.asscalar() > 0 else -1
i += 1
return X, Y, wfake, bfake
# Plot the data with colors according to the labels
def plotdata(X,Y):
for (x,y) in zip(X,Y):
if (y.asscalar() == 1):
plt.scatter(x[0].asscalar(),x[1]. asscalar(),color='r')
else:
plt.scatter(x[0].asscalar(),x[1]. asscalar(),color='b')

# define a function to plot contour plots over [-3,3] by [-3,3]

def plotscore(w,d):
xgrid = np.arange(-3, 3, 0.02) # generating grids
ygrid = np.arange(-3, 3, 0.02)
xx, yy = np.meshgrid(xgrid, ygrid)
zz = nd.zeros(shape=(xgrid.size, ygrid.size, 2))
zz[:,:,0] = nd.array(xx)
zz[:,:,1] = nd.array(yy)
vv = nd.dot(zz,w) + d
CS = plt.contour(xgrid,ygrid,vv.asnumpy())
plt.clabel(CS, inline=1, fontsize=10)
228 Machine Learning with Python: Theory and Applications

street_w = 0.1
ndim = 2
X,Y,wfake,bfake = getfake(50,ndim,3,street_w)
#generates 50 points, in 2D space with a margin of street_w
plotdata(X,Y)
plt.show()

Figure 6.2: Computer-generated data-points that are separable with a straight line.

We now see a dataset with 50 scattered points in 2D space separated by

a street with width street w. These data-points can be clearly separable by
at least one line in between the blue and red data-points.
Let us first take a look at the points after their vectors are all projected
on separable with an arbitrary vector w and with an arbitrary bias b (an
arbitrary affine transformation). We do the same projection using also the
true w and bias b used to generate all these points for comparison. We write
the following code to do so:

wa = nd.array([0.5,0.5]) # a given vector

cs = (wa [0]/nd.norm(wa)).asnumpy() # cosine value
si = (wa [1]/nd.norm(wa)).asnumpy() # sine value
ba = 0.5 * nd.norm(wa) # bias (one may change it)
Xa = nd.dot(X,wa)/nd.norm(wa) + ba # projection (affine mapping)

plt.plot(Xa.asnumpy()*cs,Xa.asnumpy()*si,zorder=1) # results
for (x,y) in zip(Xa,Y):
if (y.asscalar() == 1):
plt.scatter(x.asscalar()*cs,x. asscalar()*si,color='r')
else:
plt.scatter(x.asscalar()*cs,x.asscalar()*si, color='b')
cs = (wfake [0]/nd.norm(wfake)).asnumpy() # projection on true norm
 The Perceptron and SVM 229

si = (wfake [1]/nd.norm(wfake)).asnumpy()
Xa = nd.dot(X,wfake) + bfake # with true bias

Figure 6.3: Data-points projected on an arbitrary straight line.

plt.plot(Xa.asnumpy()*cs,Xa.asnumpy()*si,zorder=1)
for (x,y) in zip(Xa,Y):
if (y.asscalar() == 1):
plt.scatter(x.asscalar()*cs,x. asscalar()*si,color='r')
else:
plt.scatter(x.asscalar()*cs,x.asscalar()*si, color='b')
plt.show()

Figure 6.4: Data-points projected on a straight line that is perpendicular to a line that
separates these data-points.

It is seen that

• When these points are projected on a vector that is not the true normal
(along the blue line) direction, the blue and red points are mixed along
the blue line.
230 Machine Learning with Python: Theory and Applications

• When the true normal direction is used, all these points are distinctly
separated into two classes, blue and red, along the orange line. This dataset
is linearly separable.

Let us see how the Perceptron algorithm finds a direction and the bias,
and hence the red line that separates these two classes. We use again the algo-
rithm available at mxnet-the-straight-dope (https://github.com/zackchase/
mxnet-the-straight-dope). The algorithm is based on the following encour-
agement rule: positive events should be encouraged and negative ones should
be discouraged. This rule is used with the decision rule discussed earlier for
each data-point in a given dataset.

# The Perceptron algorithm

def Perceptron(w,b,x,y,strt_w):
if (y*(nd.dot(w,x)+b)).asscalar()<=strt_w/2:
# Decision rule to check whether the line with
# the current parameters in the street
w += y * x # In the street, update w
b += y # update b
update = 1
else: # Otherwise (outside the street)
update = 0 # No action
return update

The above Perception algorithm is used in an iteration to update the

learning parameters: the weights in vector w and bias b, by looping over the
dataset.

w = nd.zeros(shape=(ndim)) #stars with zero (worst case)

b = nd.zeros(shape=(1))
t = 0
print('w:',w.shape,' b:',b.shape,' X:',X.shape,' Y:',Y.shape)
for (x,y) in zip(X,Y):
update = Perceptron(w,b,x,y,street_w)
if (update == 1):
t += 1
print('In the street: update the parameters')
print('data{}, label{}'.format(x.asnumpy(),y. asscalar()))
print('weight{}, bias{}'.format(w.asnumpy(),b. asscalar()))
 The Perceptron and SVM 231

plotscore(w,b) # The plane with updated w and b

plotdata(X,Y) # data-points
plt.scatter(x[0].asscalar(),x[1]. asscalar(),color='g')
# currently updated data-point
plt.show()

w: (2,) b: (1,) X: (50, 2) Y: (50,)

In the street: update the parameters
data [ 0.03751943 -0.7298465 ], label -1.0
weight [-0.03751943 0.7298465 ], bias -1.0

Figure 6.5: Results at the first iteration.

In the street: update the parameters

data [-2.0401056 1.4821309], label -1.0
weight [ 2.0025861 -0.7522844], bias -2.0

Figure 6.6: Results at the second iteration.

232 Machine Learning with Python: Theory and Applications

In the street: update the parameters

data [ 1.040828 -0.45256865], label -1.0
weight [ 0.96175814 -0.29971576], bias -3.0

Figure 6.7: Results at the third iteration.

In the street: update the parameters

data [-0.934901 -1.5937568], label 1.0
weight [ 0.02685714 -1.8934726 ], bias -2.0

Figure 6.8: Results at the final iteration.

print('Total number of points:',len(Y),'; times of updates:',t)

print('weight {}, bias {}'.format(w.asnumpy(),b.asscalar()))
print('wfake {}, bfake {}'.format(wfake.asnumpy(),bfake. asscalar()))

Total number of points: 50 ; times of updates: 4

weight [ 0.02685714 -1.8934726 ], bias -2.0
wfake [ 0.0738321 -0.99727064], bfake -0.9501792788505554
 The Perceptron and SVM 233

It is seen that all the red dots are on the positive side of the straight line
of x · w + b = 0 with the learned parameters of the weight vector w∗ and
bias b∗ . All the data marked with blue dots are on the negative side of the
line. In the entire process, all these points stay still, and the updates are
done only on the weight vector w and bias b. We shall now examine the
fundamental reasons for this simple algorithm to work.

6.3 The Perceptron Convergence Theorem

Theorem: Consider a dataset with a finite number of data-points. The ith

data-point is paired with its label as [xi , yi ]. Any data-point xi is bounded
by xi  ≤ R < ∞, and its label yi ∈ {±1}.

• If the data-points are linearly separable, meaning that there exists at least
one pair of parameters (w∗ , b∗ ) with w∗  ≤ 1 and b2 ≤ 1, such that
yi (xi · w∗ + b∗ ) ≥ w/2 > 0 for all data pairs, where w is a given scalar of
the street width,
• then the Perceptron algorithm converges after at most t = 2(R2 +1)/w2 ∝
 R 2
w iterations, with a pair of parameters (wt , bt ) forming a line xi ·wt +bt
that separates the data-points in two classes.

We now prove this Theorem largely following the procedure with codes
at mxnet-the-straight-dope (https://github.com/zackchase/mxnet-the-
straight-dope/blob/master/chapter01 crashcourse/probability.ipynb), under
the Apache-2.0 License. We first check the convergence behavior numerically
(this may take minutes to run).
ws = np.arange(0.025,0.45,0.025) #generate a set of street widths
number_iterations = np.zeros(shape=(ws.size))
number_tests = 10
for j in range(number_tests): #set number of tests to do
for (i,wi) in enumerate(ws):
X,Y,_,_=getfake(1000,2,3,wi) #generate dataset
for (x,y) in zip(X,Y):
number_iterations[i] += Perceptron(w,b,x,y,wi)
#for each test, record the number of updates
number_iterations = number_iterations / 10.0
plt.plot(ws,number_iterations,label='Average number of iterations')
plt.legend()
plt.show()

The test results are plotted in Fig. 6.9. It shows that the number of
iterations needs to increase with the decrease of the street width w, and
234 Machine Learning with Python: Theory and Applications

Figure 6.9: Convergence behavior examined numerically.

the rate is roughly quadratic (inversely). This test supports the convergence
theorem. Let us now prove this in a more rigorous mathematical manner.
The proof assumes that the data are linearly separable. Therefore, there
exists a pair of parameters (w∗ , b∗ ) with w∗  ≤ 1 and b2 ≤ 1. Let us
examine the inner product of the current set of parameters ŵ with the
assumed existing ŵ∗ at each iteration. What we would like the iteration
to do is to update the current ŵ to approach ŵ∗ iteration by iteration, so
that their inner product can get bigger and bigger. Eventually, they can be
parallel with each other. Let us see whether this is really what is happening
in the Perceptron algorithm given above. Our examination is also iteration
by iteration but considers only the iterations when an update is made by the
algorithm, because the algorithm does nothing otherwise. This means that
we perform update only when yt (xt · ŵt ) ≤ w/2 at at the tth step.
At the initial setting in the algorithm, t = 0, we have no idea on what ŵ
should be, and thus set ŵ0 = 0. Here for neat formulation using dot-product.
We assume that column vector ŵ0 and ŵ∗ are flatten to (row) vectors, so
that ŵ can have a dot-product directly with any other (flattened) ŵ resulted
in the iteration process. This can be done easily in numpy using the flatten()
function. Thus we have at the initial setting,

ŵ0 · ŵ∗ = 0

At t = 1, following the algorithm, we bring in arbitrarily a data-point, say

x1 with y1 . We shall find y1 (x1 · ŵ1 ) = 0 ≤ (w/2), because at this point,
the current ŵ1 are still all zero. Therefore, data-point x1 with y1 is in the
 The Perceptron and SVM 235

street defined by the line with current ŵ1 . Next, we perform the following
updates:

ŵ1 = ŵ0 + y1 x1

We thus have,

ŵ1 · ŵ∗ = ŵ0 · ŵ∗ + y1 (x1 · ŵ∗ ) ≥ w/2

This is because yi (xi · ŵ∗ ) ≥ w/2 given by the Theorem as a condition. We

see the direction of vector ŵ1 approaches to that of ŵ∗ by w/2. They are
w/2 more aligned.
Similarly, at t = 2, following the algorithm, we have the following update.

ŵ2 = ŵ1 + y2 x2

We now have,

ŵ2 · ŵ∗ = ŵ1 · ŵ∗ + y2 (x2 · ŵ∗ ) ≥ 2(w/2)

This is because of the result obtained at t = 1 and the new addition of

yi (ŵ∗ · xi ) ≥ w/2 condition given by the Theorem. We see the direction of
vector ŵ1 approaches to that of ŵ∗ by 2(w/2). They are now 2(w/2) more
aligned.
It is clear that inner product adds one (w/2) in each iteration. Now at
tth update, we shall have

ŵt · ŵ∗ ≥ t(w/2) (6.9)

We see the direction of vector ŵt approaches to that of ŵ∗ by t(w/2). It is

clear that the algorithm drives ŵ more and more in alignment with ŵ∗ , at
a linear rate of (w/2). The wider the street, the faster the convergence.
We next examine the evolution of the length (amplitude) of vector ŵt+1 .

ŵt+1 2 = ŵt+1 · ŵt+1 = (ŵt + yt xt ) · (ŵt + yt xt )

= ŵt · ŵt + 2yt xt · ŵt + yt xt · yt xt (6.10)
2
= ŵt  + 2yt xt · ŵt + yt2 xt 2

Using the conditions given by the Theorem: xi  = xi  + 1 ≤ R + 1, and

yi ∈ {±1}, and yt (xt · ŵt ) ≤ w/2 (this is the condition for starting the tth
update), we shall have

ŵt+1 2 ≤ ŵt 2 + R2 + 1 + w
236 Machine Learning with Python: Theory and Applications

When t = 0, we have ŵ0  = 0, and hence,

ŵ1 2 ≤ ŵ0 2 + R2 + 1 + w = R2 + 1 + w

When t = 1, we shall have,

ŵ2 2 ≤ ŵ1 2 + R2 + 1 + w ≤ 2(R2 + 1 + w)

This means that each iteration gives (R2 + 1 + w).

At the t = T iteration, we shall have

ŵT  ≤ T (R2 + 1 + w) (6.11)

Using the Cauchy-Schwartz inequality, i.e., a·b ≥ a·b, and then Eq. (6.9),
we obtain,

ŵT ŵ∗  ≥ ŵT · ŵ∗ ≥ T (w/2)

Using the conditions given by the Theorem: ŵ∗  = w∗  + b ≤ 1 + 1 = 2,

we have
√
ŵT  2 ≥ T (w/2)

Combining this with the inequality Eq. (6.11) yields,


2T (R2 + 1 + w) ≥ T (w/2) (6.12)

Let us examine this equation. The number of iterations T is linear on the

left and in a square root on the right side of Eq. (6.12), and all others are
constants. Thus, this inequality cannot hold for large T . Therefore, T must
be limited to satisfy Eq. (6.12), which means that the Perceptron algorithm
2
will converge in a finite number of iterations, and T ≤ 8(R w+1+w)
2 ∝ (R 2
w) .
This can also be written as
R √
∝ T (6.13)
w
This means that it takes a square root of times for the Perceptron algorithm
to converge for a given relative street width.
We now give the following remarks:

1. The Perceptron convergence proof requires that the data-points be

separable with a line.
 The Perceptron and SVM 237

2. The convergence is independent of the dimensionality of the data. This

is not a condition used in the proof. It is also independent of the number
of observations.
3. The number of iterations increases with the decrease of the street width
w, and the rate is (inversely) quadratic. This echoes the numerical test
conducted earlier.
4. The number of iterations decreases with the increase of the upper bound
of the data R, and the rate is also (inversely) quadratic. This means that if
the data-points are generally farther apart, it is easier to separate, which
is intuitively understandable.
5. The algorithm updates only for the data-points that are not in alignment.
It simply skips the data-points that are already in alignment.

6.4 Support Vector Machine

6.4.1 Problem statement

In our discussions above on the Perceptron, it is seen that there are in fact
an infinite number of solutions of parameters w and b in the hypothesis
space to form straight lines for a linearly separable dataset, as long as the
street width w has a finite value. Readers may observe the different straight
lines obtained by simply starting the classifier with different initial weights
and biases. One may naturally ask what the best solution is among all these
possible solutions. One answer is that the line that separates these two classes
of data with largest “street” width and sits in the middle line of the street
may be the best. To obtain the widest street optimal solution, one can
formulate the problem as an optimization one. In fact, one can simply bring
in an optimization algorithm to control the Perceptron to try for multiple
times to find an optimal solution if the efficiency is not a concern.
Here, we introduce the well-known algorithm known as the support vector
machine or SVM. The initial idea was invented in the early 1960s in Vapnik’s
PhD thesis, and became popular in the 1990s when it (with kernel trick) was
applied for handwritten digit recognition [3, 4].
The SVM is an effective algorithm that is constructed using a systematic
formulation and the Lagrangian multiply approach. Given below is the
detailed description and formulation on SVM. A good reference is the excel-
lent lecture by Prof. Partick Winston at MIT(https://www.youtube.com/
watch?v= PwhiWxHK8o), and also some workings available online (https://
towardsdatascience.com/support-vector-machine-python-example-d67d9b
63f1c8).
238 Machine Learning with Python: Theory and Applications

6.4.2 Formulation of objective function and constraints

Our formulation continues from the formulation we derived for the Percep-
tron. The difference here is that the street width is no longer given in this
SVM setting. We just assume it is there. We must find a way to formulate
the street width, and then maximize it for a given dataset. We first derive
the formula for the street width.
Because we assume that the dataset is linearly separable, there must be
a street width of some finite value w. From the decision rule we derived,
we know the equation for the middle line of the street can be written in an
affine transformation form as
x·w+b=0 (6.14)

where b ∈ W is a learning parameter called bias, x ∈ Xp is a position

vector of an arbitrary point in the feature space with linear polynomial
bases (features), x1 , x2 , . . . , xp ,

x = [x1 , x2 , . . . , xp ] (6.15)

and w ∈ Wp is a vector of weights that are also learning parameters

w1 , w2 , . . . , wp :
w = [w1 , w2 , . . . , wp ] (6.16)

This middle line in a 2D feature space X2 is shown in Fig. 6.10 with red dash-
dot line, which is approximated using the weights w and bias b in hypothesis
space W3 .

Figure 6.10: Linearly separable data-points in 2D space, the width of the street is to be
maximized using SVM.
 The Perceptron and SVM 239

On the upper-right gutter, there should be at least one positive data-

point, say x1 , right on it. Its vector x1 is the blue arrow, a support vector
to the gutter. Because x1 belongs to the positive class, its label is +1. The
equation for this gutter line of the street can be given as

x1 · w + b = +1 (6.17)

This is the decision border for data-point x1 . Similarly, on the lower-left

gutter, there is at least one negative data-point, say x2 , right on it. Its vector
x2 supports the gutter. Because x2 belongs to the negative class, its label is
−1. The equation for this gutter line can be given as

x2 · w + b = −1 (6.18)

Consider now the projections of these two support vectors, x1 and x2 on the
w
unit normal of the middle line for the street, w . The projection of x1 gives
the (Euclidean) distance of the upper-right gutter to the origin along the
normalized w. Similarly, the projection of x2 gives the distance of the lower-
left gutter to the origin along the normalized w. Therefore, their difference
gives the width of the street:
w w x1 · w − x2 · w
w = x1 · + b − x2 · −b= (6.19)
w w w

Substituting Eqs. (6.17) and (6.18) to (6.19), we obtain

2
w= (6.20)
w
A very simple formula. Now, we got the equation for the street width, and
that depends only on training parameter weights w! This is not so difficult
to understand because these weights determine the orientation of the gutter
of the street and hence the direction of the street. When the weights change,
the street turns accordingly while remaining in touch with both data-points
x1 and x2 , which results in a change in the street width. The bias b affects
only the translational location of a line. Because the street width is the
difference of the two gutter lines, the bias is thus canceled. Therefore, the
bias b should not affect the width. Here, we observe a fact that the affine
space is not a vector space (As mentioned in Chapter 1), and the difference
of two vectors in the affine space comes back to the feature space (because
of the cancellation of the augment 1).
However, why is the street weight inversely related to the norm of the
w? This may be seemly counterintuitive, but it is really the mathematics at
240 Machine Learning with Python: Theory and Applications

Figure 6.11: Change of width of the street when the street is turned with respect to w.

work. To examine and view what is really happening here, let us look at a
simpler setting where both x1 and x2 (that are on the gutters) are sitting
on the x2 -axis, as shown in the Fig. 6.11.
When the gutters are at the horizontal direction, the equation for the
upper gutter is
0 · x1 + 1 · x2 = +1 + b (6.21)

The street width is w0 . The normal vector w and its norm are given as
follows:
  
w = 0 1 , w = 02 + 12 = 1 (6.22)

When these gutters rotate to have a slope k while remaining supported by

both x1 and x2 , the equation for the upper gutter becomes

k · x1 + 1 · x2 = +1 + b (6.23)

The new normal vector wk and its norm are given as follows:
  
wk = k 1 , wk  = k 2 + 12 (6.24)

It is obvious that the street width after the rotation, wk , is clearly smaller
than the original street width before the rotation, w0 , while the norm of w
has increased. This is also true for any value of k. The street width is at the
maximum when it is along the horizontal direction which is perpendicular
to the vector x1 − x2 . We write the following code to plot this relationship:
 The Perceptron and SVM 241

import matplotlib.pyplot as plt

import numpy as np
%matplotlib inline
fig, figarr = plt.subplots(1, 2, figsize=(11, 4))
def w_norm(k): # norm of vector w for a line of slope k
return np.sqrt(1. + k**2)
k = np.arange(-10, 10, .1)
y = 2/w_norm(k) # compute the width of the street
figarr[0].plot(k,y,c='r')
figarr[0].set_xlabel('Slope of the street gutter, $k$')
figarr[0].set_ylabel('Width of the street')
figarr[1].plot(w_norm(k),y)
figarr[1].set_xlabel('Norm of the weight vector')
figarr[1].set_ylabel('Width of the street')
#figarr[1].ax1.set_title('ax1 title')
plt.show()

Figure 6.12: Variation of the street width with the slope of the street gutter (left) and
with the norm of the weight vector (right).

It is clear from Fig. 6.12 that the street width is inversely related to the
norm of the w.
Most importantly, this analysis shows that if the street width is maxi-
mized, w must be perpendicular to the gutters (decision boundaries). This
conclusion is true for the arbitrary pair of data-points x1 and x2 on these
two gutters.
Now, Eq. (6.19) can be rewritten as
w
w= [x1 − x2 ] · (6.25)
 
w
Data pair in linear polynomial bases  

weights or optimization parameters

242 Machine Learning with Python: Theory and Applications

This means that when w is maximized, the inner production of [x1 − x2 ] and
w
w is maximized (these two vectors are parallel), where [x1 −x2 ] is a vector of
pairs of data-points in the linear polynomial bases x1 , x2 , . . . , to approximate
w
a line in the feature space, and w is the vector of the normalized weights
or tuning/optimization parameters. The use of linear polynomial bases here
is because we assume the data-points are linearly separable by a hyperplane.
Remember our original goal is to find maximum street width. Based on
Eq. (6.20), this is equivalent to minimizing the norm of w, which in turn
is the same as minimizing 12 w2 . The benefit of such simple conversions
will soon be evidenced. We now have our objective function:
1 1
L= w2 = w w (6.26)
2 2
The above function needs to be minimized. We see a nice property of the
above formulation: the objective function is quadratic, and its Hessian matrix
is a unity matrix that is clearly SPD. Therefore, it has one and only one
minimal. The local minimal is the global one. This is the fundamental reason
why local minimal/optimal is not a concern when an SVM model is used.
Notice that our minimization has to be done under the following condition
or constraint:

[yi (xi · w + b) − 1] ≥ 0 (6.27)

for all the data-points. Assume that there are a total of m data-points for
our p dimensional problem. In matrix form, this set of constraint equations
has the following form:

[y(X · w + 1b) − 1] ≥ 0 (6.28)

where y is an m × m diagonal matrix with all the labels on the diagonal

terms, X is an m × p rectangular matrix with xi sitting on the ith row, w
is a vector of p unknown weight parameters, b is the unknown parameter of
bias, and 1 is a vector of m × 1 with value 1 for all the entries in it.

6.4.3 Modified objective function with constraints:

Multipliers method
Because these constraint equations need to be satisfied, our minimization
becomes a constrained optimization problem. The objective function is nice
(quadratic), but with many constraints. These constraints are linear but in
an inequality form.
 The Perceptron and SVM 243

A very well-known technique for this type of optimization problem is

the so-called Lagrangian multipliers method (invented in 1902!), which
is widely used in many fields including the finite element methods for
imposing the so-called multi-point constraints [5]. The essential idea is that
when the parameter in the objective function has to satisfy some equality
constraints/conditions, the objective function should be modified. The
amount of modification depends on how well these conditions will be satisfied
by these parameters. However, before we obtain the solution for these
parameters that minimize the objective function, we do not really know.
The natural way to take this into consideration in the objective function is
to add a cost to it. Such a cost should be tunable by a set of additional
parameters, and the number of the additional parameters should correspond
to the number of constraints we have. For our case, we have m constraints,
and hence we shall have m additional parameters (although some of them
will be zero eventually after minimization). These parameters are called
Lagrangian multipliers. The modified objective function can be written as
1
L = w w − λ  [y(X · w + 1b) − 1] (6.29)
2
where λ is a vector of m × 1 with m Lagrangian multipliers λi that are the
additional unknown parameters.
From Eq. (6.29), we can see that when any of these m equations (equality
part) in the brackets is fully satisfied, the added cost is zero. If not, the
dissatisfaction amount multiplied with λ is added into the objective function
as an additional cost. This means that any dissatisfaction to the constraints
is accounted for a cost. In addition, we purposely put these Lagrangian
multipliers in a linear form, so that if we perform the differentiation to
the modified objective function with respect to the Lagrangian multipliers
to derive the minimization conditions, these constrained equations will be
fully recovered. This means that when our modified objective function is
minimized, these constraints will be fully satisfied.
Moreover, because of our original constraints, the terms in the square
brackets in Eq. (6.29) must be larger than or equal to zero. Therefore, any
dissatisfaction will be negative. For the added cost to be positive, we must
impose the following constrains to these multipliers (Note the minus sign on
the front of λ ):
λi ≥ 0 (6.30)
Therefore, Eq. (6.30) will be the new constraints on λi for our minimization
problem using the modified objective function, to ensure the inequality of
244 Machine Learning with Python: Theory and Applications

the original constraints on w. In the algorithm to be discussed later, this set

of new constraints will be implemented.
Note that the Lagrangian multiplier method shall be applicable only to
equality constraints. However, Eq. (6.29) includes data-points subject to
inequality constraints. The justification is given as follows.
When [yi (xi · w+b)−1] < 0, meaning the condition/constraint is violated,
λi will be driven by the minimization to zero, when the loss is minimized,
because λi is set nonnegative.
When [yi (xi · w + b) − 1] > 0, the inequality parts of the constraints are
satisfied. These constraints can be changed to equality ones by introducing
independent slack variables s2i :

[yi (xi · w + b) − 1] − s2i = 0

The modified loss function with only equality constraints can now be
written as,
m

1   
L= w w− λi [yi (xi · w + b) − 1] + s2i
2
i

When L is minimized, it is stationary at the minimizer. Setting the partial

differentiation of L with respect to each si to zero gives,
∂L
= −2λi si = 0
∂si
This means that the modified loss function with the slack variables is the
same as Eq. (6.29) without the slacks. In addition, the foregoing equation can
be satisfied, if either λi or si is zero. Also, when si = 0, [yi (xi ·w +b)−1] = 0,
hence we have

λi [yi (xi · w + b) − 1] = 0

Therefore, for all the data-points that satisfies [yi (xi · w + b) − 1] > 0 (data-
points away from the street), the corresponding λi must be zero, meaning
that it is not active. This proves that Eq. (6.29) can be used, because it
added only zeros. We will use Eq. (6.29) to derive the equations for our dual
problem.
When [yi (xi · w + b) − 1] = 0, the data-points are on the gutters, and
the added cost is zero. Thus, λi can be positive nonzero. Thus, for it to be
bounded, some additional condition on λi is needed, which will be naturally
obtained through minimization conditions.
 The Perceptron and SVM 245

When a data-point satisfies the inequality in Eq. (6.27), the corresponding

λ should be automatically become zero, when the minimization process
is completed. The positive nonzero λs correspond to these data-points
satisfying the equality part of Eq. (6.27), or zero margin data-points that are
right on the gutters. These vectors of these data-points is called supporting
vectors. This will be demonstrated in the example given in Section 6.4.6.
Finally, we note that the constrained optimization problem now has a
total of p + 1 + m parameters.
Note that there is another possible method to modify the objective
function, known as the penalty method, which does not bring in new
parameters. Interested readers may refer to the literature [5].

6.4.4 Converting to a standard quadratic programming problem

Equation (6.29) is still quadratic in w (for fixed λi s), but now lives in
higher-dimensional space. We need to search for the stationary point for the
modified objective function by performing its differentiations with respect,
first, to w and then force it to zero, which gives

∂L(w, b, λ)
= w − X yλ = 0 (6.31)
∂w
Noticing that L is a scalar, we seek for its partial derivative with respect to
a vector w. Therefore, the outcome is a vector, as expected. Equation (6.31)
leads to
n
w = X yλ = λi yi xi (6.32)
i=1

Equation (6.32) implies that the weight parameter w is some kind of

linear combination of the data vectors attached with their corresponding
labels. The coefficients for the combination are the Lagrangian multipliers.
Observing from Fig. 6.10, this finding does make a lot of sense: vector w is
indeed surrounded by all the data vectors xi . Their combination should be
able to produce w. Because the Lagrangian multipliers are “watching” how
these decision rules are obeyed by each of the data vectors, they should have
the “right” to decide on the coefficients of the contributions of xi to w.
We next take differentiations with respect to b and then force it to zero,
which gives
∂L(w, b, λ)
= λ y1 = 0 (6.33)
∂b
246 Machine Learning with Python: Theory and Applications

which leads to
n

 
λ y1 = 1 yλ = λi yi = 0 (6.34)
i=1

This equation implies that the sum of the label-attached Lagrangian

multipliers vanishes. This is somehow interesting: the label of the data-points
and the multipliers has come into some kind of “neutral” agreement!
With the above relations obtained, the objective function can be rewritten
as follows:
1
L = λ yXX yλ − λ yXX yλ − λ y1b + 1 λ (6.35)
2
which is reduced to the objective function for the so-called dual problem.
 n
 n n
1
L = 1 λ − λ yXX yλ = λi − λi yi λj yj (xi · xj ) (6.36)
2
i=1 i=1 j=1

Note that b is now out of the picture, which echoes Eq. (6.20): the width
of the street should not depend on b. This is expected because b affects
only the translational location of the street. In addition, the data-points (in
the feature space) used in the SVM formulation are in the form of inner
product: (xi · xj ). This means that what matters is the interrelationship of
these data-points. This special formulation of SVM allows the effective use
of the so-called nonlinear kernel functions that evaluate the interrelationship
of these data-point for problems with nonlinear decision boundaries. We will
discuss this further in Section 6.4.10.
In practical computation, we may compute first Xy = yX, whose ith row
is simply yi xi , and can be computed efficiently. Equation (6.36) can now be
further simplified as
1
L = 1 λ − λ Pλ (6.37)
2
where

P = [Pij ] = Xy X
y , where Pij = yi yj (xi · xj ) (6.38)

It is clear that matrix P is symmetric (and of course real). It is non-negative

definite. In this case, our objective function is also quadratic in parameter λ,
which is the preferable property for our optimization algorithm. However, we
see that the sign on the quadratic term of λ is negative. Hence, its curvature
is negative with respect to λ, which is the reverse of the sign in Eq. (6.26).
Therefore, our original minimization problem is converted to an optimization
 The Perceptron and SVM 247

problem. In addition, our objective function given in Eq. (6.29) must have a
saddle shape: it has a positive curvature along w, and a negative curvature
along λ.
In the theory of Lagrangian multipliers, when the objective function is
altered with λ, the partial differentiations of the modified function can in
general only give stationary points where the extrema reside. These points
are saddle points, which means that along some parameter axes these points
may be minimums, and along other parameter axes they may be maximums.
Therefore, whether the extremum is an optimum or minimum needs to be
examined carefully. To see the issue more clearly, let us take a look at a
simplified problem with a loss function of only two parameters of w and λ.
The simplified function still has the same essential behavior, so that we can
write a Python code and plot a 3D figure to view what is really going on.

import numpy as np
import matplotlib as mpl
mpl.rcParams['figure.dpi']= 220
import matplotlib.pyplot as plt
from mpl_toolkits import mplot3d
from mpl_toolkits.mplot3d import axes3d
%matplotlib inline
#%matplotlib tk

w = np.linspace(-2.0,2.0,100) # just one weight

l = np.linspace(-4.,4.0,100) # just one lambda
W, L = np.meshgrid(w, l) # create grids for 3D plot
c1, c2 = 10.0, 2.0 # assign some constants
Loss = c1*(W*W) - c2*(L*L) # simplified loss function
fig_s = plt.figure(figsize=(8,5)) # define sizes for 3D plot
ax = fig_s.add_subplot(1,1,1,projection='3d')
ax.set_xlabel('$w$', fontsize=12, rotation=0)
ax.set_ylabel('$\lambda$', fontsize=12)
ax.set_zlabel('$\mathcal{L}$', fontsize=12, rotation=0)
ax.yaxis._axinfo['label']['space_factor'] = 3.0
#ax.plot_wireframe(X1, X2, Z, color='b',rstride=5, cstride=5)
ax.plot_surface(W,L,Loss,color='b',rstride=2,cstride=2,
shade=False, cmap="jet", linewidth=1)
plt.title('A simplified loss function with a saddle point,
+str(c1)+', $c_2$=' +str(c2))
248 Machine Learning with Python: Theory and Applications

#plt.legend()
fig_s.tight_layout() # otherwise y-label will be clipped
plt.show()

Figure 6.13: An idealized loss function with a saddle point.

We see a saddle surface and a saddle point at the middle point on the
surface. It is clear now that the function has minima along w (for any fixed
λ) and maxima along λ (for any fixed w).
Back to our general objective functions L given in Eq. (6.37), let us look
at the 2nd differentiation with respect to λ:
∂ 2 L(λ)
= −P (6.39)
∂λ2
which is the Hessian matrix (−P, a measure of curvature), which is a negative
definite matrix; because P is non-negative definite, we can clearly again see
that our problem has now become an optimization problem with respect to
our newly introduced Lagrangian multiplier λ.
Finally, noticing that our objective function is with constraints given in
Eqs. (6.31) and (6.34), our SVM problem can be written as
1
max − λ Pλ + 1 λ
λ 2
subject to λ ≥ 0 (6.40)

1 yλ = 0
 The Perceptron and SVM 249

Note that 1 y is simply a row vector that collects all the labels of the data-
points, which can be made use of in the algorithm to avoid unnecessary
computation.
Of course, we can easily convert our optimization problem into a
minimization problem as follows, by simply revising the sign of the objective
function:
1 
min λ Pλ − 1 λ
λ 2
subject to λ ≥ 0 (6.41)
1 yλ = 0

This simple change allows us to plug our problem to a CVXOPT standard

solver to obtain the solution.
Let us now take a look at the standard solver at the CVXOPT library
(use help(cvxopt.solvers.qp) to check it out). It solves a quadratic program
defined as follows:
1 
minimize x Px + q x
2
subject to Gx ≤ h (6.42)

Ax = b

Compared with Eq. (6.41), it is seen clearly that our problem is finally
formulated as a typical and standard problem with a quadratic (and hence
convex) objective function and a set of linear inequality constraints, known as
the quadratic programming (QP) problem. Well-established QP techniques
can readily be used as standard routines to solve this problem, including
the CVXOPT (https://cvxopt.org/), which is a free software package for
convex optimization based on Python. Because of the excellent property of
convexity, one does not need to worry about the local optimums. It has one
and only one optimum (or minimum) and it is the global optimum.

6.4.5 Prediction in SVM

After an SVM model is trained, we shall know w and b. For any given test
data-point xj , we can then make a prediction using
n

ŷj = xj · w + b = λi yi (xi · xj ) (6.43)
i=1
250 Machine Learning with Python: Theory and Applications

When ŷ ≥ 1, xj belongs to the positive class, ŷ ≤ −1, it belongs to the

negative class, and in the street otherwise.
Note that the prediction also needs only the inner product of the feature
variables of the dataset: (xi · xj ).

6.4.6 Example: A Python code for SVM

We now use (with permission) the Python code given by Cory Maklin at htt
ps://github.com/corymaklin/svm to examine the SVM approach. The code
uses the QP tools and solvers.

#Importing the necessary libraries

import numpy as np
import cvxopt # installation needed, for QP tools & solvers
from sklearn.datasets import make_blobs
from sklearn.model_selection import train_test_split
from matplotlib import pyplot as plt
from sklearn.svm import LinearSVC
from sklearn.metrics import confusion_matrix
print('cvxopt version',cvxopt.__version__)
cvxopt version 1.2.6

Define a class called SVM.

class SVM:
def fit(self, X, y):
n_samples, n_features = X.shape
# P = X X^T
K = np.zeros((n_samples, n_samples))
for i in range(n_samples):
for j in range(n_samples):
K[i,j] = np.dot(X[j], X[i])
P = cvxopt.matrix(np.outer(y, y) * K) #use cvxopt
print('K.shape',K.shape,' y.shape',y.shape)
# q = -1 (1xN)
q = cvxopt.matrix(-1*np.ones(n_samples))
# A = y^T
A = cvxopt.matrix(y, (1, n_samples))
# b = 0
b = cvxopt.matrix(1.0)
# -1 (NxN)
 The Perceptron and SVM 251

G = cvxopt.matrix(-1*np.diag(np.ones(n_samples)))
# 0 (1xN)
h = cvxopt.matrix(np.zeros(n_samples))
solution = cvxopt.solvers.qp(P, q, G, h, A, b) # QP
# Lagrange multipliers
a = np.ravel(solution['x']) #flatten to a 1D array
# Lagrange have nonzero Lagrange multipliers
sv = a > 1e-5
ind = np.arange(len(a))[sv]
self.a = a[sv]
self.sv = X[sv]
self.sv_y = y[sv]
# Intercept
self.b = 0
for n in range(len(self.a)):
self.b += self.sv_y[n]
self.b -= np.sum(self.aself.sv_yK[ind[n], sv])
self.b /= len(self.a)
# Weights
self.w = np.zeros(n_features)
for n in range(len(self.a)):
self.w += self.a[n] * self.sv_y[n] * self.sv[n]

def project(self, X):

return np.dot(X, self.w) + self.b
def predict(self, X):
return np.sign(self.project(X))

We next use the scikit-learn library to generate a set of linearly separable

data. We need to label the negative samples as −1 instead of 0 in order to
use cvxopt.

X, y = make_blobs(n_samples=255, centers=2,
random_state=0, cluster_std=0.65)
y[y == 0] = -1
tmp = np.ones(len(X))
y = tmp * y
# Let us take a look at the data.
print('X.shape',X.shape)
plt.scatter(X[:, 0], X[:, 1], c=y, cmap='winter');

X.shape (255, 2)
252 Machine Learning with Python: Theory and Applications

Figure 6.14: Computer-generated data-points that can be separated via a straight line.

Caution: If for some reason a “math domain error” is encountered in

running the following cell, try to restart Kernel and re-run the code from
the beginning of this subsection.

# We split the data into training and testing sets.

X_train,X_test,y_train,y_test=train_test_split(X,y, random_state=0)
print(X_train.shape,X_test.shape,y_train.shape, y_test.shape)
# Create an instance, and then train the instance of our SVM class.
svm = SVM()
svm.fit(X_train, y_train)
#Plot the decision boundaries and support vectors.
def f(x, w, b, c=0):
return (-w[0] * x - b + c) / w[1] #x1.w1 + x2.w2 + b = c

plt.scatter(X_train[:,0],X_train[: ,1],c=y_train, cmap='winter')

# w.x + b = 0
a0 = -1; a1 = f(a0, svm.w, svm.b)
b0 = 4; b1 = f(b0, svm.w, svm.b)
plt.plot([a0,b0], [a1,b1], 'k')
# w.x + b = 1
a0 = -1; a1 = f(a0, svm.w, svm.b, 1)
b0 = 4; b1 = f(b0, svm.w, svm.b, 1)
plt.plot([a0,b0], [a1,b1], 'k--')
# w.x + b = -1
a0 = -1; a1 = f(a0, svm.w, svm.b, -1)
b0 = 4; b1 = f(b0, svm.w, svm.b, -1)
plt.plot([a0,b0], [a1,b1], 'k--')
w_street = 2/np.linalg.norm(svm.w) # compute the street width
print('The width of the street:',w_street)
 The Perceptron and SVM 253

(191, 2) (64, 2) (191,) (64,)

K.shape (191, 191) y.shape (191,)
pcost dcost gap pres dres
0: -2.1259e+01 -4.3138e+01 7e+02 2e+01 2e+00
1: -3.3107e+01 -3.1685e+01 3e+02 1e+01 8e-01
2: -5.5227e+01 -3.9785e+01 3e+02 8e+00 6e-01
3: -1.0725e+02 -3.9083e+01 1e+02 3e+00 3e-01
4: -1.9273e+01 -1.4357e+01 2e+01 4e-01 3e-02
5: -1.3042e+01 -1.3209e+01 3e+00 5e-02 4e-03
6: -1.2960e+01 -1.2963e+01 4e-02 6e-04 5e-05
7: -1.2961e+01 -1.2961e+01 4e-04 6e-06 5e-07
8: -1.2961e+01 -1.2961e+01 4e-06 6e-08 5e-09
Optimal solution found.
The width of the street: 0.6301007705116851

Figure 6.15: SVM results of a widest street that separates these data-points.

# Finally, the trained model is used for prediction

# using the test dataset.
print('X_test.shape',X_test.shape)
print('y_test.shape',y_test.shape)
y_proj = svm.project(X_test)
y_pred = svm.predict(X_test)
print('Projected:', y_proj[:5])
print('Predicted:', y_pred[:5])
#use a confusion matrix showing its accuracy
confusion_matrix(y_test, y_pred)
254 Machine Learning with Python: Theory and Applications

X test.shape (64, 2)
y test.shape (64,)
Projected: [ 8.4316 -2.7710 -5.9351 7.9451 5.31072511]
Predicted: [ 1. -1. -1. 1. 1.]

array([[31, 0],
[ 0, 33]], dtype=int64)

6.4.7 Confusion matrix

A confusion matrix C is generally defined as follows: Cij : the number of
observations that are in group i and predicted to be in group j.
For binary classification problems, the count of true negative classes is
C00 and false negative is C10 . The count of true positive classes is C11 and
false positives is C01 .
For our above problem, we had all 31 negative counts and 33 positive
counts. There was no false prediction for both classes.

6.4.8 Example: A Sickit-learn class for SVM

The above code allows one to modify when it is needed for his/her problems.
As an alternative, we also use the existing powerful SVM library to solve
the same problem, just for comparison of various approaches. The following
codes are from the Scikit-learn [6].

# Import the necessary libraries

import numpy as np
from sklearn.datasets import make_blobs
from sklearn.model_selection import train_test_split
from matplotlib import pyplot as plt
from sklearn.svm import LinearSVC
from sklearn.metrics import confusion_matrix
# Create the dataset
X, y = make_blobs(n_samples=255, centers=2, random_state=0,
cluster_std=0.65)
y[y==0] = -1
y = np.ones(len(X)) * y
print('X.shape',X.shape) # a look at the dataset
plt.scatter(X[:, 0], X[:, 1], c=y, cmap='winter');
# Split the dataset into training and testing sets.
X_train,X_test,y_train,y_test=train_test_split(X,y, random_state=0)
print(X_train.shape,X_test.shape,y_train.shape, y_test.shape)
 The Perceptron and SVM 255

X.shape (255, 2)
(191, 2) (64, 2) (191,) (64,)

Figure 6.16: Computer-generated data-points that can be separated via a straight line
for sciket-learn SVM classifier.

# scikit-learn SVM class as SVC (support vector classifier)

svc = LinearSVC(tol=0.0001, C=400.0,max_iter=3000) # instance
svc.fit(X_train, y_train) # train the model
#Plot the decision boundaries and the support vectors.
plt.scatter(X_train[:,0],X_train[:,1],c=y_train,
cmap='winter');
ax = plt.gca()
xlim = ax.get_xlim()
w = svc.coef_[0]
a = -w[0] / w[1]
w_street = 2/np.linalg.norm(w) # compute the street width
print('The width of the street:',w_street)
xx = np.linspace(xlim[0], xlim[1])
yy = a * xx - svc.intercept_[0] / w[1]
plt.plot(xx, yy)
yy = a * xx - (svc.intercept_[0] - 1) / w[1]
plt.plot(xx, yy, 'k--')
yy = a * xx - (svc.intercept_[0] + 1) / w[1]
plt.plot(xx, yy, 'k--');

The width of the street: 0.6549977642924647

256 Machine Learning with Python: Theory and Applications

Figure 6.17: Sciket-learn SVM results of a widest street that separates these data-points.

Notice that the width of the street computed here is a little narrower than
that obtained using the previous code. This is because LinearSVC allows one
to compute the so-called soft decision boundary, by setting the C arguments
when creating the svc instance. One may try to change the C value and run
the code again to see what happens.
We now make the prediction again using these test samples, and then
evaluate the prediction quality in the form of a confusion matrix.

y_pred = svc.predict(X_test)
confusion_matrix(y_test, y_pred)
array([[31, 0],
[ 0, 33]], dtype=int64)

For this example, we also had all 31 negative counts and 33 positive
counts. There was no false prediction for both classes.

6.4.9 SVM for datasets not separable with hyperplanes

The above examples classify data-points that are all separable by straight
lines (or hyperplanes). There are datasets that can only be separated
properly by nonlinear curves. The SVM is also an effective approach for
such datasets. It is done through expansion of the feature space to higher
dimensions, by adding nonlinear basis functions in Eq. (6.14), similar to
what we do in the FEM [5]. For example, for a problem with p = 3, the
preliminary feature variables are

x = {x1 , x2 , x3 } ∈ X3 (6.44)
 The Perceptron and SVM 257

We can use the Pascal pyramid to include more nonlinear (NL) basis to form
features of
 √ √ √
xN L = 1, 2x1 , 2x2 , 2x3 , x1 x1 , x1 x2 , x1 x3 ,
(6.45)
x2 x1 , x2 x2 , x2 x3 , x3 x1 , x3 x2 , x3 x3 ∈ X13

The dimension of the expanded feature space is calculated using p × p + p +

1 = 13. If we use xN L instead of x, all the formulation for SVM presented
above shall hold, but the inner product (xi · xj ), i, j = 1, 2, 3 is replaced
by the inner product (xN I
L · xN L ), I, J = 1, 2, . . . , 13. Data-points that are
J
separable by a hyperplane may become separable when mapped to a higher-
dimensional space. An excellent video demonstrating this can be found at
“SVM with polynomial kernel visualization” (https://www.youtube.com/
watch?v=OdlNM96sHio&t=0s).
However, when p is a big number, the dimension of the expanded feature
space grows drastically, and it can be too costly in computation. Therefore,
we do not actually form the expanded features. Instead, we use a kernel
trick to achieve effects of using expanded features.

6.4.10 Kernel trick

The kernel trick places the inner product (xN
I
L · xN L ) with a proper kernel
J
function that is much cheaper to compute. For the bases in Eq. (6.45), the
kernel function is found as

k(x, xj ) = (x · xj + 1)2 (6.46)

It is not too difficult for readers to verify:

 L N L
(xi · xj + 1)2 = xN I · xJ (6.47)

Notice that the order of computation for the left is p, but that for the right
would be p2 . The saving by using a kernel is significant. The generalized
nonlinear polynomial kernels can be given as follows:

k(x, xj ) = (x · xj + c)d (6.48)

where c and d are predefined constants. Often used constants are c = 0

or 1, and d = 2 ∼ 5. In general cases, we may not know precisely what are
the expanded features xN L . Fortunately, we do not need to know. Because of
this, many types of kernel functions can be used in SVM without the need to
258 Machine Learning with Python: Theory and Applications

find xN L , including the radial basis functions (RBFs). For example, when the
Gaussian basis function (a typical RBF) is used, we have the Gaussian kernel:
 
x − xj 2
k(x, xj ) = exp − (6.49)
2σ 2

where σ is a predefined parameter which controls the width of the kernel

function. Such a kernel is a function of (Euclidean) distance measured
from xj to x, and hence it is also called a distance function. It is
sometimes called similarity function, because when x approaches xj ,
k(x, xj ) approaches 1 and is thus similar. When x departs from xj , k(x, xj )
approaches zero exponentially, indicating dissimilarity.
With such a simple replacement of the inner product by a kernel, the
rest of the operations to find out the optimized parameters for curved
decision boundaries remain basically the same. Because of this convenience
in using kernels, machine learning packages, such as SK-learn (https://scikit-
learn.org/stable/modules/svm.html), allow users to define their own custom
kernels that may fit better to their applications.
All the similarity functions can be used as kernel for SVM, by simply
replacing Pij in Eq. (6.47) with Pij = yi yj k(xi xj ), as long as the kernel is
SPD based on the [Mercer’s theorem]. RBFs are symmetric, because they
are distance functions and hence swapping x and xj gives the same results.
It is also positive-definite, hence SPD, as proven by Powell [7]. Care may be
needed on the conditioning and computation complexity, when large number
of data-points are used.
The concept of kernel is quite generally applicable. It is not just used in
SVM, but also possibly in other machine learning techniques, such as the
nonlinear PCA, as long as the formulation uses the data-points in an inner
product manner. It is also frequently used in physics-law-based methods,
such as the meshfree methods [8].

6.4.11 Example: SVM classification with curves

The following code from Scikit learn [6] performs nonlinear SVM, which
classifies data-points that are separated by curves.

import numpy as np
import matplotlib.pyplot as plt
from sklearn import svm
 The Perceptron and SVM 259

xx, yy = np.meshgrid(np.linspace(-3, 3, 500),

np.linspace(-3, 3, 500))
np.random.seed(0)
X = np.random.randn(300, 2)
Y = np.logical_xor(X[:, 0] > 0, X[:, 1] > 0)
# fit the model
clf = svm.NuSVC()
clf.fit(X, Y)

# plot the decision function for each datapoint on the grid

Z = clf.decision_function(np.c_[xx.ravel(), yy.ravel()])
Z = Z.reshape(xx.shape)
plt.imshow(Z, interpolation='nearest', extent=(xx.min(),
xx.max(), yy.min(), yy.max()), aspect='auto',
origin='lower', cmap=plt.cm.PuOr_r)
contours = plt.contour(xx, yy, Z, levels=[0], linewidths=2)
plt.scatter(X[:, 0], X[:, 1], s=30, c=Y, cmap=plt.cm.Paired)
plt.xticks(())
plt.yticks(())
plt.axis([-3, 3, -3, 3])
plt.show()

Figure 6.18: SVM classification of data-points with curves via an Sklearn classifier.

It is seen that the Sklearn classifier SVC has also quite correctly classified
all these samples in this test dataset.
260 Machine Learning with Python: Theory and Applications

6.4.12 Multiclass classification via SVM

In many machine learning applications, there are datasets of multiple classes.
SVM can also be used for such classification tasks. Multiclass reduction
techniques are used in multi-classifications, including One-vs-All (also known
as One-vs-Rest) and One-vs-One. Let us take a look at the following
comparison case study.

6.4.13 Example: Use of SVM classifiers for iris dataset

This section introduces a comparison study performed by the Sklearn team.
It compares different linear SVM classifiers on a 2D projection of the iris
dataset, considering the first two features in the dataset:

• Sepal length.
• Sepal width.

Decision surfaces are produced using SVM classifiers with four different
kernels. The linear models LinearSVC() and SVC(kernel='linear') give
slightly different decision boundaries. This is because

• LinearSVC minimizes the squared hinge loss, while SVC minimizes the
regular hinge loss.
• LinearSVC uses the One-vs-All (also known as One-vs-Rest) multiclass
reduction, while SVC uses the One-vs-One multiclass reduction.

%matplotlib inline
print(__doc__)
import numpy as np
import matplotlib.pyplot as plt
from sklearn import svm, datasets
def make_meshgrid(x, y, h=.02):
"""Create a mesh of points to plot in
Parameters
----------
x: data to base x-axis meshgrid on
y: data to base y-axis meshgrid on
h: stepsize for meshgrid, optional
Returns
-------
 The Perceptron and SVM 261

xx, yy : ndarray
"""
x_min, x_max = x.min() - 1, x.max() + 1
y_min, y_max = y.min() - 1, y.max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
np.arange(y_min, y_max, h))
return xx, yy

def plot_contours(ax, clf, xx, yy, **params):

"""Plot the decision boundaries for a classifier.
Parameters
----------
ax: matplotlib axes object
clf: a classifier
xx: meshgrid ndarray
yy: meshgrid ndarray
params: dictionary of params to pass to contourf, optional
"""
Z = clf.predict(np.c_[xx.ravel(), yy.ravel()])
Z = Z.reshape(xx.shape)
out = ax.contourf(xx, yy, Z, **params)
return out

# import some data to play with

iris = datasets.load_iris()
# Take the first two features. We could avoid this by using a
# two-dim dataset
X = iris.data[:, :2]
y = iris.target
# Create an instance of SVM and fit out data. We do not scale
# data since we want to plot the support vectors
C = 1.0 # SVM regularization parameter
models = (svm.SVC(kernel='linear', C=C),
svm.LinearSVC(C=C, max_iter=10000),
svm.SVC(kernel='rbf', gamma=0.7, C=C),
svm.SVC(kernel='poly', degree=3, gamma='auto', C=C))
models = (clf.fit(X, y) for clf in models)
# title for the plots
titles = ('SVC with linear kernel',
'LinearSVC (linear kernel)',
'SVC with RBF kernel',
'SVC with polynomial (degree 3) kernel')
262 Machine Learning with Python: Theory and Applications

# Set-up 2×2 grid for plotting.

fig, sub = plt.subplots(2, 2, figsize=(10, 8))
plt.subplots_adjust(wspace=0.2, hspace=0.2)
X0, X1 = X[:, 0], X[:, 1]
xx, yy = make_meshgrid(X0, X1)
for clf, title, ax in zip(models, titles, sub.flatten()):
plot_contours(ax, clf, xx, yy,
cmap=plt.cm.coolwarm, alpha=0.8)
ax.scatter(X0,X1,c=y,cmap=plt.cm.coolwarm,s=40,edgecolors='k')
ax.set_xlim(xx.min(), xx.max())
ax.set_ylim(yy.min(), yy.max())
ax.set_xlabel('Sepal length')
ax.set_ylabel('Sepal width')
ax.set_xticks(())
ax.set_yticks(())
ax.set_title(title)
plt.show()

Automatically created module for IPython interactive environment

Figure 6.19: Results of Sklearn classifiers of SVM on the iris dataset.

 The Perceptron and SVM 263

The results show that both linear models have linear decision boundaries
of intersecting hyperplanes. The nonlinear kernel models (polynomial or
Gaussian RBF) produce more flexible curved decision boundaries. The
shapes of the boundaries depend on the kernel type and parameters used.
Readers are referred to Sklearn’s case studies for more details (https://scikit-
learn.org/stable/auto examples/svm/plot rbf parameters.html).

References
[1] F. Rosenblatt, The perceptron: A perceiving and recognizing automation, Report 85-
60-1, Cornell Aeronautical Laboratory, 1957.
[2] F. Rosenblatt, Principles of Neurodynamics: Perceptrons and the Theory of Brain
Mechanisms, 1962. https://books.google.com/books?id=7FhRAAAAMAAJ.
[3] C. Corinna and V.N. Vapnik, Support-vector networks, Machine Learning, 20(3), 273–
297, 1995.
[4] B.-H. Asa, H. David, S. Hava et al., Support vector clustering, Journal of Machine
Learning Research, 2, 125–137, 2001.
[5] G.R. Liu and S.S. Quek, The Finite Element Method: A Practical Course, Butterworth-
Heinemann, London, 2013.
[6] P. Fabian, V. Gae, G. Alexandre et al., Scikit-learn: Machine learning in Python,
Journal of Machine Learning Research, 12(85), 2825–2830, 2011. http://jmlr.org/pap
ers/v12/pedregosa11a.html.
[7] M.J.D. Powell, The theory of radial basis function approximation in 1990, Advances
in Numerical Analysis, II, 105–210, 1992. https://ci.nii.ac.jp/naid/10008965650/en/.
[8] G.R. Liu, Mesh Free Methods: Moving Beyond the Finite Element Method, Taylor and
Francis Group, New York, 2010.
[9] Y. Xiaowei, S. Lei, H. Zhifeng et al., An extended Lagrangian support vector machine
for classifications, Progress in Natural Science, 14(6), 519–523, 2004.
MACHINE LEARNING
WITH PYTHON
 Chapter 7

Activation Functions and

Universal Approximation Theory

Activation functions are some of the most essential elements in machine

learning models based on artificial neural networks. Activation functions are
used for the purpose of bringing in must-have nonlinearity to the neural
network to enable it for complex problems. We have discussed about this in
Chapter 5. This chapter takes a close look at this topic. In this book, the
activation function is generally denoted using φ(z) with an argument of z
that is often an input to a neuron in a neural network and often an affine
transformed output from the previous layer of neurons.
There are many types of activation functions that have been used, some
are still in active use, and more are being developed. Our discussion will focus
on the typical ones. After establishing a clear understanding of these acti-
vation functions in this chapter, it should be reasonably straightforward to
understand other functions. A quite complete list of activation functions can
be found at Wikipedia (https://en.wikipedia.org/wiki/Activation function).
With a good understanding of the roles of the weights and bias of the
neuron on the output of the activation function, we shall then be able
to appreciate the basic capability of a neuron. Such capability offers a
pictorial presentation of the universal approximation theorem [1, 2] of neural
networks.
It is worth noting that in training a neural network, we would usually
need the derivatives of the loss function with respect to all its training
parameters, which is often part of the argument of the activation function.
Therefore, the activation function should be differentiable or at least
piecewise differentiable. All activation functions discussed in this chapter
satisfy this requirement.
We first examine the famous sigmoid activation function.

265
266 Machine Learning with Python: Theory and Applications

7.1 Sigmoid Function (σ(z))

The sigmoid function is the earliest activation function used in training

neural networks. It is still widely used today, and we will use it in multiple
chapters. It has a coined symbol σ and is defined as follows:

1
φ(z) = σ(z) = (7.1)
1 + e−z

It is clear that for any given argument of a real number z in (−∞, ∞), it
returns a positive number within (0,1). In machine learning, the argument
is often an array, and we use element-wise operations. Hence, σ(z) is also
an array, given an array input. The sigmoid activation function σ maps,
or practically squashes, any argument value into (0,1). Hence, it is called
squashing function. It is also called logistic function, because it is often used
in logistic regression, giving a result in a kind of probability.
The derivative of the sigmoid function has a simple form of

φ (z) = φ(z)(1 − φ(z)) (7.2)

We now write the following codes of the activation function and to demon-
strate the behavior of the sigmoid function using mxnet and matplotlib.

%matplotlib inline
import mxnet as mx
from mxnet import nd, autograd, gluon
import matplotlib.pyplot as plt
import numpy as np
from matplotlib import rc
def logistic0(z): # also known as the sigmoid function
return 1. / (1. + nd.exp(-z))

plt.figure(figsize=(4.0, 2.5),dpi=100)
z = nd.arange(-10, 10, .1)
y = logistic0(z)
plt.plot(z.asnumpy(),y.asnumpy())
plt.xlabel('z')
plt.ylabel(r'$\sigma(z) $')
plt.title('Sigmoid Function')
plt.grid(color='r',which='both',linestyle=':',linewidth=0.5)
plt.show()
 Activation Functions and Universal Approximation Theory 267

l0 = int(len (z)/2) #take look at its value at the center

print('z=',z[l0].asscalar(),'sigmoid(0)=',
logistic0(z[l0]). asscalar())

Figure 7.1: Variation of the sigmoid function: positive, monotonic, smooth, and
differentiable.

z= 0.0 sigmoid(0)= 0.5

It is clear that the sigmoid outputs are between 0 and 1, and hence
qualified for probability outputs. It has a value of 0.5 at z = 0. Therefore,
if we set its output as positive when the probability is greater than 0.5, and
negative whenever the output is less than 0.5, we have an adequate predictor
for samples with ±1 labels.
The derivative of the sigmoid function can be plotted using the following
simple code:

yg = logistic0(z)*(1.-logistic0(z))
plt.figure(figsize=(4.0, 2.5),dpi=100)
plt.plot(z.asnumpy(),yg.asnumpy())
plt.xlabel('z')
plt.ylabel(r"$\sigma \ '(z) $")
plt.title('Derivative of the sigmoid Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.show()

As shown in Figs. 7.1 and 7.2, the sigmoid activation function is

monotonic, confined in (0, 1), smooth, and differentiable. Its derivative is
positive, smooth and further differentiable. These are all important and
268 Machine Learning with Python: Theory and Applications

Figure 7.2: Variation of the derivative of the sigmoid function: positive, smooth, and
further differentiable.

useful properties for ML models. The higher-order differentiability is useful

for ML models using loss functions containing higher-order differentiations.
It is also observed that for large arguments (positive or negative), the
derivative of sigmoid function vanishes. This has implications when it is
used in backward propagation processing when training a deepnet, where a
gradient is needed.

7.2 Sigmoid Function of an Affine Transformation Function

In neural network-based machine learning, the inputs to the first layer take
the direct inputs of samples or data-points. Therefore, no activation function
is needed for the input layer. For all the subsequent layers, the inputs to a
layer of neurons are an affine transformation of the output of the previous
layer, as discussed in Chapter 5. Let us consider here one neuron at an output
layer. We shall have
z =x·w+b (7.3)

where x is the input from the previous layer, w is the weight matrix that
connects the previous layer to the current layer, and b is the bias for the
neurons of the current layer. Both w and b are to be updated through
training, and hence often referred to as training (or learning) parameters. In
this chapter, we do not study how to update these parameter, but focus on
the effects of them on the output of the neuron. When the sigmoid function
is applied to produce the output of the current layer, we shall have the
following general expression.
1
xout = −(x·w+b)
(7.4)
1+e
 Activation Functions and Universal Approximation Theory 269

7.3 Neural-Pulse-Unite (NPU)

Let us write a Python code to examine how the outputs of the activation
function are influenced by the weights and biases. We also use the mxnet
model to do the number crunching and plot out the results in graphics.
We first let w vary in a range, and b be fixed at 0.

def logistic(w,x,b):
return 1. / (1. + nd.exp(-(x*w + b)))

x = nd.arange(-10, 10, .1)

w = nd.arange(1.0, 50, 5.0)
b = nd.arange(0, 50, 10)
b0=b[0]
plt.figure(figsize=(4.0, 3.5),dpi=100.)
for wi in w:
y = logistic(x,wi,b0) # b is fixed
plt.plot(x.asnumpy(),y.asnumpy(),
label="wi="+str(wi. asnumpy()))
plt.xlabel('x')
plt.ylabel(r"$\sigma(z=xw+b)$")
plt.title('Sigmoid Function with b='+str(b0.asscalar()))
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.show()

Figure 7.3: Effects of weights on the sigmoid function.

270 Machine Learning with Python: Theory and Applications

Figure 7.3 shows clearly that a larger weight makes the sigmoid curve
steeper near x = 0. We can practically make a function that is as closer as
we want to the step function, by using a very large weight.
Let us now examine the influence of the bias on the shape of the activation
function with a fixed weight.

x = nd.arange(-10, 10, .1)

w = nd.arange(0.0, 50, 5.0)
b = nd.arange(-30, 30, 10)
w1=w[1].asscalar()
plt.figure(figsize=(4.0, 3.0),dpi=100.)
for bi in b:
y = logistic(x,w1,bi)
plt.plot(x.asnumpy(),y.asnumpy(),\
label="b="+str(bi[0].asnumpy()))
plt.xlabel('x')
plt.ylabel(r"$\sigma(z=xw+b) $")
plt.title('Sigmoid Function with w='+str(w1))
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend(loc='center right', bbox_to_anchor=(1, 0.5))
plt.show()

Figure 7.4: Effects of the bias.

Figure 7.4 shows clearly that the bias makes the sigmoid curve shift: A
positive bias value shifts the sigmoid curve leftwards and a negative value
 Activation Functions and Universal Approximation Theory 271

Figure 7.5: NPU: A pair of neurons is capable of producing a close-to-rectangular pulse

when a large w value is used. The pulse location and its width are controlled by the base
bias b0 , width of the bias bw , and w.

shifts the curve rightwards. It practically controls where in space the output
should appear. If we make use of these effects of both weight and bias, we
can produce a close-to-rectangular pulse function using a pair of activation
functions. We call it Neural-Pulse-Unit (NPU). It is schematically drawn in
Fig. 7.5.
The width (in the x coordinate) at the middle point of the pulse, pw , is
controlled by the difference of the biases used in the two paired neurons and
the weight. The formula can be derived in the following procedure.
First, the z value at the middle point of the sigmoid function should
always be zero; we shall then have for the neuron (on the top)

z = x+ · w + b = 0 (7.5)

where b is the base bias of the neuron pair, as shown in Fig. 7.5 above. For
the neuron below,

z = x− · w + b − bw = 0 (7.6)

where bw is the negative offset of the bias at the neuron below. The width
at the middle point of the pulse, pw , becomes

pw = x+ − x− = bw /w (7.7)

It depends on bw and w.
By altering the weight and bias values, one can also create such a pulse
at any x. The center location of the pulse, xc , can be calculated using
x+ + x− 2b − bw
xc = =− (7.8)
2 2w
It depends on b, bw , and w.
It is seen that the width and the location of the pulse depend only on
the weight and bias, which are trainable parameters. This implies that once
the neurons are trained with a set of fixed parameters, it will be capable of
producing the function value for any given input x.
272 Machine Learning with Python: Theory and Applications

We now write the following code to generate a pulse:

x = nd.arange(-8, 8, .1)
w = nd.arange(0.0, 30, 5.0)
w1=w[-1].asscalar()
b0=10.0 # the base bias
bw=20.0 # negative shift of b at the 2nd neuron
bc = 2*b0 - bw # compute the center value of b
xc = -0.5* bc/w1 # compute x of the center of the pulse
pw = bw/w1 # width of the pulse (at the middle)

y = logistic(x,w1,b0)-logistic(x,w1,b0-bw)
plt.figure(figsize=(4.5, 3.0),dpi=100.)
plt.plot(x.asnumpy(),y.asnumpy())
plt.xlabel('x')
plt.ylabel(r"$\sigma(x*w+b0)-\sigma(x*w+b0-bw) $")
plt.title('Neural Pulse, Sigmoid, w='
+str(w1)+' b='+str(b0)+' bw='+str(bw))
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.show()
print('b0=',b0,'bc=',bc,'xc=',xc,'pw=',pw)

Figure 7.6: A pulse generated using a pair of sigmoid functions.

b0= 10.0 bc= 0.0 xc= -0.0 pw= 0.8

We have successfully created a rectangular pulse function with unit height

at x = 0, as shown in Fig. 7.6.
The above neural network has just one input neuron (the green filled
circle), one hidden layer with two neurons (the two vertical orange filled ell-
ipses in the middle), and one output layer with one neuron (orange filled
ellipse on the right). To produce the pulse, we have set the weights for both
 Activation Functions and Universal Approximation Theory 273

hidden neurons as the same large value and let the bias shift. The outputs
of these two neurons will all be close to step functions, but in opposite
orientation and at different stepping-up locations. The final output neuron
sums these two step functions and delivers a pulse. It functions as an NPU.
Using multiple such NPUs, one can generate as many NPUs located
at different xc as one desires. Below is a code for producing a lined-up
rectangular unit pulse.
plt.figure(figsize=(6.5, 3.0),dpi=200)
x = nd.arange(-8, 8, .1)
w = nd.arange(0.0, 51, 5.0)
b = nd.arange(-120, 190, 40)
w1=40. #w[-1].asscalar() #We use only one but large weights
#for all NPUs
db=20.0 #=bw #define the width of the pulse
plt.figure(figsize=(6.5, 3.0),dpi=100.)
for bi in b:
y = logistic(x,w1,bi)-logistic(x,w1,bi-db)
plt.plot(x.asnumpy(),y.asnumpy(),
label="b="+str(bi[0]. asnumpy()))

plt.xlabel('x')
plt.ylabel(r"$\sigma(x*w+b)-\sigma(x*w+b-bw) $")
plt.title('Subtraction of two Sigmoid Functions with w='+str(w1)
+' bw='+str(db))
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.show()

<Figure size 1300x600 with 0 Axes>

Figure 7.7: Eight pulses generated using sigmoid functions.

274 Machine Learning with Python: Theory and Applications

Figure 7.7 shows eight pulses lined up created using eight NPUs (eight
pairs of neurons in the hidden layer). Each NPU is responsible for a pulse
centered at its xc value. As long as we have enough number of pair neurons
in the hidden layer, one can create as many pulses as needed to cover the
interest domain of x.

7.4 Universal Approximation Theorem

Universal approximation theorem [1] states that a neural network

with sufficiently large number of neurons in one hidden layer is capable of
approximating any continuous function.

7.4.1 Function approximation using NPUs

Now, when the outputs of these pairs of neurons are fed to the final output
neuron, a linear combination using the following formula can be used to
approximate an arbitrary given function:

Nx

f (x) = f (xi ) · P (xi ) (7.9)
i=1

where f (x) is the given function, f (x) denotes the approximated function,
f (xi ) is the given function value at xi , P (xi ) is the unit rectangular pulse
created by the ith neuron pair for xi , and Nx is the number of xi sampled
in the domain of x.
Let us write another code for this single hidden layer neuron network to
get this done.

plt.figure(figsize=(6.5, 3.0),dpi=100)
x = nd.arange(-10, 10, .1)
w = nd.arange(0.0, 50, 5.0)
b = nd.arange(-400, 470, 40)
w1=w[-1].asscalar() # use one large weights for all NPUs
bw=40.0 # negative shift (width) of the bias
for bi in b:
bc = 2*bi - bw # center value of b
xc = -0.5* bc/w1 # x at the center of pulse
fcos = nd.cos(xc/2) # value of cosine function
pulse = (logistic(x,w1,bi)-logistic(x,w1,bi-bw))
 Activation Functions and Universal Approximation Theory 275

y = fcos * pulse #linear combination of pulse

plt.plot(x.asnumpy(),y.asnumpy(),
label="b="+str(bi[0]. asnumpy()))

xi = -0.5*(2*b - bw)/w1
plt.plot(xi.asnumpy(),np.cos(xi.asnumpy()/2),c='b',
linewidth=2.5)
plt.xlabel('x')
plt.ylabel("cos(x/2)")
plt.title('Approximating cos(x) using neuron pulses, w='
+str(w1)+' bw='+str(bw))
plt.grid(color='r', linestyle=':', linewidth=0.5)
#plt.legend()
plt.show()

Figure 7.8: A cosine function is approximated using a series of neuron pulses.

The above demonstrations, as plotted in Figs. 7.6–7.8, show that at most

two neurons in the hidden layer can approximate one point in a curve.
The number of neurons required in approximating a curve depends on the
complexity of the curve varying with x. If 200 points needed to approximate
the curve with desired accuracy, one would need (at most) to use 400 neurons
in the hidden layer.

7.4.2 Function approximations using neuron basis functions

We now change our way to do the approximation using the node-based

approximation idea in the meshfree methods [3]. In this case, we treat all the
center points of the pulses as nodes, and all the pulse functions as continuous
276 Machine Learning with Python: Theory and Applications

nodal shape functions. Our approximation of the function will become as

follows:
Nx

f (x) = f (xi ) · P (x) (7.10)
i=1

where f (xi ) is the nodal value of the function at xi , and P (x) is now a
continuous neuron basis function of pulse-like. In this case, the approximated
function f (x) becomes a continuous function, and the neuron pair become
a basis function.
Using a series of NPUs, this universal approximation theorem (UAT) can
be schematically drawn in Fig. 7.9. A UAT-Net can be constructed with
one input layer, one hidden layer, and one output layer. If the hidden layer
has sufficient number of neuron pairs, each pair can form one NPU unit,
which is responsible to produce one pulse function at a location. When all

Figure 7.9: Schematic description of the Universal Approximation Theorem. This UAT-
Net consists of one hidden layer with sufficient number of neuron pairs. Each pair forms
one NPU unit responsible to produce one pulse at a location. When all these weights at the
input and output layers and all these biases at the hidden layer neurons are set (trained)
properly, any continuous function can be approximated for any input value of x.
 Activation Functions and Universal Approximation Theory 277

these weights at input and output layers and all these biases at the hidden
layer neurons are set (or trained) properly, the UAT-Net is capable of
approximating any continuous function for any input value of x.
Let us plot this function to see what it looks like.

plt.figure(figsize=(6.5, 3.0),dpi=200)
x = nd.arange(-10, 10, .05) #we refined the points
w = nd.arange(0.0, 50., 5.0)
b = nd.arange(-400, 470, 40)

w1=w[-1].asscalar() # Use one large weights for all NPUs

bw= b[1]-b[0] #20.0 # negative shift of bias
sx = int(x.size)
yfem = np.zeros(sx)
for bi in b:
bc = 2*bi - bw # center value of b
xc = -0.5* bc/w1 # x at the center of the pulse
fcos = nd.cos(xc/2) # nodal function value of cosine
pulse = (logistic(x,w1,bi)-logistic(x,w1,bi-bw))
y = fcos * pulse # linear combination of pulse
plt.plot(x.asnumpy(),y.asnumpy(),
label="b="+str(bi[0]. asnumpy()))
yfem += y.asnumpy()

plt.plot(x.asnumpy(),yfem,c='r')
#plot the approximated function
xi = -0.5*(2*b - bw)/w1
plt.plot(xi.asnumpy(),np.cos(xi.asnumpy()/2),c='b',
linewidth=0.5)
plt.xlabel('x')
plt.ylabel("cos(x/2)")
plt.title('Approximating cos(x) using neuron pulses, w='
+str(w1)+' bw='+str(bw.asscalar()))
plt.grid(color='r', linestyle=':', linewidth=0.5)
#plt.legend()
plt.show()
278 Machine Learning with Python: Theory and Applications

Figure 7.10: A cosine function is approximated using a series of neuron basis functions
that are pulse-like.

The approximated function using pulse-like neural basis functions

becomes the thinker red line, as shown in Fig. 7.10. It is clearly continuous.
One may also use the FEM method do to a similar approximation [4]. In
this case, we still treat all the center points of the pulses as nodes, and our
elements can now be defined as the segments between the center point of a
pulse and the center of its next pulse. Each of the pulse functions becomes
a nodal shape function. The FEM approximation can then be done using
only two pieces of pulse functions that cover the element. The outcome of
the approximated results will be quite similar.
The above analysis clearly shows the capability of a neuron and a neural
network. Let us end this section by mentioning the following points:

1. We found one possible way to approximate a function. It is not necessarily

the best possible way.
2. In actual applications, if the weights and biases of a neural network are
trained in a systematic optimal way, it should perform much better than
the hand-designed neural network above.
3. It is not really necessary to create pulse functions to perform approxima-
tion in reality. Doing this here is for the purpose of showing the concept
in an easy-to-understand manner.
4. The width of each pulse does not have to be the same. In a trained neural
network, the training parameters in each of the neurons are individually
tuned for an overall performance measured by a loss function.
5. All the neurons on the hidden layer in an actual trained neural network
are working together in a system, not necessarily in pairs. Hence, it is
expected to perform much better, because of this flexibility.
 Activation Functions and Universal Approximation Theory 279

6. Our above analysis is for approximating a curve in one dimension. A

similar idea is applicable to approximation of surfaces and even higher-
dimensional hyper-surfaces.
7. When more hidden layers are used, one can significantly reduce
the number of neurons in the hidden layers. Theoretical analysis has
already shown that the total number of neurons used should be much
less, provided such a deepnet can be properly trained. It is an issue
of trainability. Deepnets have become popular and practical to may
applications because we now have increasingly better ways to train them
well with big datasets.
8. We rarely see people using the neural network to approximate a function.
Our tasks in real-life problems are often a lot more complicated and
practical.
9. The significance of the universal approximation theorem reveals the
capability of a neural network in terms of function approximation.

The affine transformation applied in the neural network setting has

created a connection between the data-prediction relationship (x, ŷ) and
the model parameters (w, b). In other words, (x, ŷ) or the features in the
data can be corded in (w, b). In addition, when activation functions are
used for all the neurons in a layer, the outputs of all these neurons become
linearly independent, and can be treated as new inputs to be fed further to
the next layers, forming an MLP, leading to the Universal Prediction Theory
for ML model creation. This forms the foundation of neural network-based
machining learning models, as discussed in Chapter 5.
To support item (3), we can use a smaller weight to create the nodal
shape functions and then approximate the cosine function as below.

plt.figure(figsize=(6.5, 3.0),dpi=100.)
x = nd.arange(-10, 10, .05) # refined points
#w = nd.arange(0.0, 50., 5.0)
#b = nd.arange(-400, 470, 40)
w = nd.arange(0.0, 10., 5.0)
b = nd.arange(-400, 470, 5)
w1=w[-1].asscalar() # Use one but large weights for all NPUs
bw= b[1]-b[0] #20.0 # define the width of the bias
sx = int(x.size)
yfem = np.zeros(sx)
pu = np.zeros(sx)
280 Machine Learning with Python: Theory and Applications

for bi in b:
bc = 2*bi - bw # center value of b
xc = -0.5* bc/w1 # x at the center of the pulse
fcos = nd.cos(xc/2) # nodal function value of cosine
pulse = (logistic(x,w1,bi)-logistic(x,w1,bi-bw))
y = fcos * pulse # linear combination of pulse
plt.plot(x.asnumpy(),y.asnumpy(),
label="b="+str(bi[0]. asnumpy()))
yfem += y.asnumpy()
pu += pulse.asnumpy() # Check on whether the partitions
# of unity (PU) is satisfied.
plt.plot(x.asnumpy(),yfem,c='r')
plt.plot(x.asnumpy(),pu,c='b')
print(pu[:60:10])
xi = -0.5*(2*b - bw)/w1
plt.xlabel('x')
plt.ylabel("cos(x/2)")
plt.title('Approximating cos(x) using neuron pulses, w='
+str(w1)+' bw='+str(bw.asscalar()))
plt.grid(color='r', linestyle=':', linewidth=0.5)
#plt.legend()
plt.show()

[0.99999999 1. 0.99999999 1. 0.99999999 1. ]

Figure 7.11: A cosine function is approximated using a series of neuron basis functions
as “nodal shape functions”.

Figure 7.11 shows that the approximated function (the red line) is
smoother and hence has better quality in approximating the cosine function.
 Activation Functions and Universal Approximation Theory 281

This is because the neural pulse is used as a nodal shape function that is
smooth and not necessarily being a sharp pulse.
Note that the partition of unity (PU) property/condition is one of the
most important properties of nodal shape functions used in a numerical
method for a mechanics problem [3]. The PU condition states that
Nx

P (x) = 1 (7.11)
i=1

This needs to be satisfied at any x. In the code above, we added one more
computation to check whether this property is satisfied if we would use
the neuron pair as a nodal shape function. As shown in Fig. 7.11, the PU
condition is indeed naturally satisfied by these neuron basis.
If we further reduce the pulse width, we shall obtain a very accurate
approximation of the function, as shown in Fig. 7.12.

Figure 7.12: Neuron Basis: A pair of neurons can be used as nodal basis functions for
function approximation. In this case, one does not have to use very large weight. The
resolution can be controlled by the width of the bias used in the neuron pairs.

In the study above, we used the sigmoid activation function. Many other
activation functions to be discussed in the following sections may also be
used. In Ref. [2], triangular pulses are constructed and used in addition to
the sigmoid pulses to approximate functions, and they worked better in some
ways.

7.4.3 Remarks
Note that Eq. (7.9) uses NPUs to support the Universal Approximation The-
ory, which provide a nice intuition for understanding. However, Eq. (7.10)
is more general. This is because the neuron basis function P (x) used in
282 Machine Learning with Python: Theory and Applications

Eq. (7.10) do not have to be a pulse for function approximation, as shown

in the previous two figures. Therefore, any activation function can be used
for constructing P (x), as long as it is monotonic for uniqueness (see Section
5.5.3) and nonlinear for linearly independence (see Section 5.8.2).

7.5 Hyperbolic Tangent Function (tanh)

The tanh function is another activation function widely used in neural

networks to bring in nonlinearities. It is defined as follows:
ez − e−z
φ(z) = tanh(z) = (7.12)
ez + e−z
It is clear that for any given argument of a real number z in (−∞, ∞), it
returns a real number within (−1, 1). Again, in machine learning models,
the argument is often an array. Hence, we use element-wise operations,
and tanh(z) will also be an array. The tanh activation function maps any
argument value into (−1, 1), in comparison with the sigmoid function that
maps to (0, 1). It comes handy for many machine learning models that need
outputs in real numbers within (−1, 1).
The derivative of the tanh function has a simple form of

φ (z) = 1 − φ(z)2 (7.13)

def tanhf(z):
return nd.tanh(z)

x = nd.arange(-10, 10, .1)

y = tanhf(x)
#plt.figure(figsize=(4.5, 3.0),dpi=100.)
plt.plot(x.asnumpy(),y.asnumpy())
plt.xlabel('z')
plt.ylabel('$tanh(z)$')
plt.title('Hyperbolic Tangent Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.show()
yg = 1.-tanhf(x)**2
plt.plot(x.asnumpy(),yg.asnumpy())
plt.xlabel('x')
plt.ylabel("$tanh(z)\ '(z) $")
plt.title('Derivative of Hyperbolic Tangent Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.show()
 Activation Functions and Universal Approximation Theory 283

Figure 7.13: Hyperbolic tangent (tanh) activation function (monotonic, smooth, and
differentiable) and its derivative (positive, smooth, and further differentiable).

l0 = int(len (x)/2)
print('x=',x[l0].asscalar(),'tanh(0)=',tanhf(x[l0]). asscalar())

x= 0.0 tanh(0)= 0.0

As shown in Fig. 7.13, the hyperbolic tangent (tanh) activation function

is monotonic, confined in (−1, 1), smooth, and differentiable. Its derivative
is positive, smooth and further differentiable.

7.6 Relu Functions

Relu stands for “rectified linear unit” [5]. It is widely used in deep neural
nets in applications in computer vision, speech recognitions, etc. It is the
most popular activation function for training deep neural networks, due to
its effectiveness in mitigating the so-called gradient vanishing issues. It has
the simple form of
284 Machine Learning with Python: Theory and Applications


0 z<0
φ(z) = z + = max(0, z) = (7.14)
z z≥0
Relu function is also known as a ramp function. It is an analogy for half-
wave rectification in electrical engineering. It simply wipes out the effects
from the inputs that are negative, but keeps the non-negative inputs as it
is, resulting in nonlinearity. Such a drastic discount to the inputs might be
one of the reasons why it worked well for densely connected deepnets.
Relu function is piecewise differentiable. The derivative of the Relu
function also has a very simple form of

 0 z<0
φ (z) = (7.15)
1 z≥0
The unit gradient is helpful in mitigating the gradient vanishing in a deepnet,
because it remains unchanged with the depth of the net.
Relu has a number of variations, including leaky Relu and parametric
Relu. A general formulation can be given as follows:

kn z z < 0
φ(z) = (7.16)
kp z z ≥ 0
where kn and kp are all positive constants, in which kn represents the
derivative of the negative portion and kp represents the derivative of the
positive portion. For the leaky Relus, one often uses kn = 0.01 and kp = 1.0.
In parametric Relus, we make these parameters trainable.
The derivative of the generalized Relu function has the form of

 kn z < 0
φ (z) = (7.17)
kp z ≥ 0
It is seen that for a deep net, the negative part effect can vanish quickly with
depth because of the small value of kn . In such cases, the leaky Relu may
behave like the Relu.
We now write a simple code to demonstrate the generalized Relu function
and its derivatives, with tunable parameters kn and kp .

def reluf(z,kn,kp):
f = [kn*zi if zi < 0.0 else kp*zi for zi in z]
return f

def relug(z,kn,kp):
g = [kn if zi < 0.0 else kp for zi in z]
return g
 Activation Functions and Universal Approximation Theory 285

x = np.arange(-10, 10, .1)

y = reluf(x,0.1,1.0)
plt.plot(x,y)
plt.xlabel('x')
plt.ylabel("$relu(z)$")
plt.title('Relu Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.show()
yg = relug(x,0.1,1.0)
plt.plot(x,yg)
plt.xlabel('x')
plt.ylabel("$relu(z)\ '(z) $")
plt.title('Derivative of the Relu Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.show()

Figure 7.14: Relu activation function (monotonic, continuous, piecewise smooth, and
differentiable), and its derivative (positive, discontinuous at the origin).
286 Machine Learning with Python: Theory and Applications

Note that Relu or leaky Relu function is not smooth, only piecewise
differentiable, and not differentiable at the origin. Its gradient jumps at the
origin as shown in Fig. 7.14. One would need to use the so-called sub-gradient
in the optimization process (to be discussed in Chapter 9). The sub-gradient
can be simply the average of the gradients on the both sides of the origin.
The non-smooth continuity of Relu function at the origin is similar to
domain discretization via elements in the FEM [4] where the approximated
function is continuous but not smooth on the elemental interfaces. The
treatment of the non-smoothness in FEM is using the so-called weak form
formulation.

7.7 Softplus Function

The softplus function has a similar shape as the rule, but is differentiable in
the entire 1D space. It is defined as follows:

φ(z) = ln(1 + ez ) (7.18)

It is clear that for any given argument of a real number z in (−∞, ∞), it
returns a positive number within (0, ∞).
The derivative of the softplus function is
1
φ(z) = σ(z) = (7.19)
1 + e−z
It is the sigmoid function (and hence it is also smooth and differentiable).
We now plot the function and its derivative using matplotlib.

import mxnet as mx
from mxnet import nd, autograd, gluon
import matplotlib.pyplot as plt
import numpy as np
from matplotlib import rc
%matplotlib inline

def softplus(z):
return np.log(1. + np.exp(z))

def softplusg(z):
return 1./(1. + np.exp(-z))

x = np.arange(-10, 10, .1)

y = softplus(x)
plt.plot(x,y)
plt.xlabel('x')
 Activation Functions and Universal Approximation Theory 287

plt.ylabel('Softplus')
plt.title('Softplus Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.show()
plt.plot(x,softplusg(x))
plt.xlabel('x')
plt.ylabel('Derivative of Softplus function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.title('Derivative')
plt.show()

Figure 7.15: The softplus activation function (monotonic, continuous, smooth, and
differentiable), and its derivative (positive, smooth, and further differentiable).

As shown Fig. 7.15, the softplus activation function is monotonic,

continuous, smooth, and differentiable. Its derivative is positive, smooth,
and further differentiable.
288 Machine Learning with Python: Theory and Applications

7.8 Conditions for activation functions

We now ready to list some conditions for an activation function.

(1) It should be monotonic for uniqueness requirement (see Section 5.5.3).
(2) It should be nonlinear for linearly independence requirement (see Section
5.8.2).
(3) Analytical formula is available for gradient computation, as in all the
examples given in this chapter. Computation of gradients for ML models
will be discussed in the next chapter.
(4) It does not contribute much in gradient vanishing. Relu is an excellent
example. Its gradient stays at 1. Therefore, it does not contribute to
gradient vanishing in a deepnet architecture.

7.9 Novel activation functions

Based on the conditions given in the previous section, we can now device
some new activation functions. We present these functions below. Note that
these functions have not yet be tested in machine learning models. Interested
readers may give it a try.

7.9.1 Rational activation function

The ration function is a new type of activation function presented in
this book for the first time (to the best knowledge of the author). It is
equipped with parameter α that can be used to control the peak values of
the derivation of the function. The α may also be treated as a trainable
parameter, so that the derivative value also be trained. It has also a tunable
function Q(z). Therefore, it is a family of activation functions. The general
form is defined as follows.
Q(z)
φ(z) = ; z ∈ (−∞, ∞) (7.20)
α + |Q(z)|
where α is a constant in R, and Q(z) is an arbitrary function as long as
it is (strictly) monotonically increasing with z ∈ [0, ∞) and monotonically
increasing with z ∈ (−∞, 0). Note that φ(z) ∈ (−1, 1). Its derivative can be
easily found as
αQ (z)
φ (z) = (7.21)
(α + |Q(z)|)2
where Q (z) is the derivative of Q(z).
 Activation Functions and Universal Approximation Theory 289

Let, for example, Q(z) = z, we have the simplest rational activation

function:
z
φ(z) = (7.22)
α + |z|
Its derivative can be easily found as
α
φ (z) = (7.23)
(α + |z|)2
If we would like the rational activation function to be positive: φ(z) ∈ (0, 1),
we need simply to do a positive shifting by 1.0 and then a scaling by 0.5,
which gives,
 
z
φ(z) = 0.5 1 + (7.24)
α + |z|
In this case, its derivative becomes,
0.5α
φ (z) = (7.25)
(α + |z|)2
The Python codes for this activation function and its derivative are given
below.

# Rational activation function by GR Liu; in (-1, 1)

def raf(z, alfa):
f = [zi/(alfa-zi) if zi < 0.0 else zi/(alfa+zi) for zi in z]
return f # or return z/(alfa+ np.abs(z))
#return np.sign(z)*np.sqrt(np.abs(z))/(alfa+np.sqrt(np.abs(z)))
#return np.sign(z)*z**2/(alfa+np.abs(z**2))

def d_raf(z,alfa): # analytical derivative

return alfa/(alfa+np.abs(z))**2

alfa = 2.0
x = np.arange(-10., 10., .1) #x -> z
y = raf(x,alfa)
plt.plot(x,y)
plt.xlabel('z')
plt.ylabel('Function value')
plt.title('Rational activation function, in (-1, 1)')
plt.grid(color='r', linestyle=': ', linewidth=0.5)
plt.savefig('raf.png',dpi=500,bbox_inches='tight')
plt.show()
plt.plot(x,d_raf(x,alfa))
plt.xlabel('z')
plt.ylabel('Derivative of the function')
290 Machine Learning with Python: Theory and Applications

plt.grid(color='r', linestyle=': ', linewidth=0.5)

plt.title('Derivative of the Rational activation function')
plt.savefig('d_raf.png',dpi=500,bbox_inches='tight')
plt.show()

Figure 7.16: The rational activation function (monotonic, continuous, confined in (−1, 1),
smooth, and differentiable), and its derivative (positive, piecewise smooth, and further
differentiable).

# Rational activation function by GR Liu; in (0, 1)

def rafp(z, alfa):
#f=[.5*alfa/(alfa-zi) if zi<0. else (.5*alfa+zi)/(alfa+zi) for zi in z]
f = [zi/(alfa-zi) if zi<0. else zi/(alfa+zi) for zi in z]
return 0.5*(1.+np.array(f))

def d_rafp(z,alfa): # analytical derivative

return .5*alfa/(alfa+np.abs(z))**2

alfa = 2.0
 Activation Functions and Universal Approximation Theory 291

x = np.arange(-10., 10., .1) #x -> z

y = rafp(x,alfa)
plt.plot(x,y)
plt.xlabel('z')
plt.ylabel('Function value')
plt.title('Rational activation function, in (0, 1)')
plt.grid(color='r', linestyle=': ', linewidth=0.5)
plt.savefig('rafp.png',dpi=500,bbox_inches='tight')
plt.show()
plt.plot(x,d_rafp(x,alfa))
plt.xlabel('z')
plt.ylabel('Derivative of the function')
plt.grid(color='r', linestyle=': ', linewidth=0.5)
plt.title('Derivative of the Rational activation function')
plt.savefig('d_rafp.png',dpi=500,bbox_inches='tight')
plt.show()

Figure 7.17: The rational activation function (monotonic, continuous, confined in (0, 1),
smooth, and differentiable), and its derivative (positive, piecewise smooth, and further
differentiable).
292 Machine Learning with Python: Theory and Applications

7.9.2 Power function

The power function is another new type of activation function presented
in this book (to be best knowledge of the author) for the first time. It is
equipped with parameter α that can be used to control the degree of non-
linearity for different nets and/or for different layers in a deepnet. This is
also a family of activation functions. The general form is given as,

−(−z)α z < 0
φ(z) = (7.26)
zα z≥0

where α is a given positive constant limited by 1. Because α < 1, the negative

sign is needed when z is less than zero, or the output become a complex
number. When α = 1, it is linear. The smaller the α, the stronger the
nonlinearity. The α may also be treated as a trainable parameter, so that
the nonlinearity at different layers can also be trained.
Note that φ(z) defined in Eq.(7.26) is in (−1, 1). If we would like it to
be positive: φ(z) ∈ (0, 1), we need to do a shifting and then a scaling, as we
have done earlier for the rational activation function.
Let us also write a code to examine the behavior of the power function.

#Power function as an activation function

def powerf(z,alpha): # power function using list comprehension

f = [-(-zi)**alpha if zi < 0.0 else zi**alpha for zi in z]
return f

x0 = np.array([0.0,2.0])
x = np.arange(-2, 2, 0.11)
#alfa = np.arange(0.2, 1.2, .2)
alfa = 1.0/np.arange(1, 9, 1.0)
nx = 1
for ai in alfa:
y = powerf(x,ai)
plt.plot(x,y,label = r"$\alpha$="+"{:.2f}".format(ai))

plt.xlabel('z')
plt.ylabel("$\phi(z) $")
plt.title('Power Activation Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.savefig('paf.png',dpi=500,bbox_inches='tight')
plt.show()
 Activation Functions and Universal Approximation Theory 293

Figure 7.18: The power activation function. It is monotonic, continuous, piecewise

smooth, and differentiable.

The above curves clearly show that when α → 0, it gives similar behavior
as the tanh function. Therefore, we may choose a small α for the first hidden
layer, and then let it grow when the net approaches the last layer. At the
last layer we use α = 1, a linear function that is exactly what we want.
The derivative of the power function is as follows.

 α(−z)α−1 z < 0
φ (z) = (7.27)
α(z)α−1 z≥0
We write the following code to examine the derivative of the power
function.

def powerg(z,alpha):
a1 = alpha-1.
g = [alpha*(-zi)**a1 if zi < 0.0 else alpha*zi**a1 for zi in z]
return g

x0 = np.array([0.0,2.0])
x = np.arange(-2, 2, 0.11)
#alfa = np.arange(0.2, 1.2, .2)
alfa = 1.0/np.arange(1, 9, 1.0)

for ai in alfa:
yg = powerg(x,ai)
plt.plot(x,yg,label = r"$\alpha$="+"{:.2f}".format(ai))

plt.xlabel('z')
plt.ylabel("$\phi \ '(z) $")
plt.title('Derivative of the Power Activation Function')
294 Machine Learning with Python: Theory and Applications

plt.grid(color='r', linestyle=':', linewidth=0.5)

plt.legend()
plt.savefig('d_paf.png',dpi=500,bbox_inches='tight')
plt.show()

Figure 7.19: The derivative of the power activation function. It is positive, continuous,
piecewise smooth, and further differentiable.

At a glance, the derivative of the power function is quite similar to that of

the tanh function we have seen in the previous section, but it is not the same.
The derivative of the power function goes to infinity at z = 0! Therefore,
this can possibly lead a numerical problem.
We may now modify it, by introducing a linear piece near z = 0.

7.9.3 Power-linear function

The power-linear function is a modification to the power function, by using
a linear function in the vicinity of z = 0, so that the derivative is bounded.
It has the following form.
⎧
⎨ −(−z)α z < −zm
φ(z) = km z −zm ≤ z < zm (7.28)
⎩ α
z z ≥ zm
where zm is a positive value defining the half-width of the linear portion, and
km is the derivative of the linear portion of the function near the vicinity of
z = 0 computed using
(α−1)
km = zm (7.29)
 Activation Functions and Universal Approximation Theory 295

We examine the power-linear function, using the following code.

#Power-linear(1) function as an activation function

def powerf1(z,alpha,xm):
fm = xm**alpha
km = fm / xm # xm^(alpha-1)
f = [-(-zi)**alpha if (zi<-xm) else km*zi if (-xm<=zi<=xm)\
else zi**alpha for zi in z]
return f

x0 = np.array([0.0,2.0])
x = np.arange(-1, 1, 0.01)
#alfa = np.arange(0.2, 1.2, .2)
alfa = 1.0/np.arange(1, 9, 1.0)
xm = 0.1
for ai in alfa:
y = powerf1(x,ai,xm)
plt.plot(x,y,label = r"$\alpha$="+"{:.2f}".format(ai))

plt.xlabel('z')
plt.ylabel("$\phi(z) $")
plt.title('Power-Linear Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.savefig('plaf.png',dpi=500,bbox_inches='tight')
plt.show()

Figure 7.20: The power-linear activation function. It is monotonic, continuous, piecewise

smooth, and differentiable.
296 Machine Learning with Python: Theory and Applications

We do get a linear portion near z = 0, which connects the two pieces of

power functions one each side. Let us look at the derivative of this power-
linear function now. The formula is,
⎧
⎪
⎪ α(−z)α−1 z < −zm
⎨
φ (z) = km −zm ≤ z < zm (7.30)
⎪
⎪
⎩ α(z)α−1 z ≥ zm

#Power-linear(1) function, derivative of the function

def powerg1(z,alpha,xm):
fm = xm**alpha
km = fm / xm
a1 = alpha-1.
g = [alpha*(-zi)**a1 if zi<-xm else km if (-xm<=zi<=xm)\
else alpha*zi**a1 for zi in z]
return g

x0 = np.array([0.0,2.0])
x = np.arange(-1, 1, 0.01)
#alfa = np.arange(0.2, 1.2, .2)
alfa = 1.0/np.arange(1, 9, 1.0)

xm = 0.1
for ai in alfa:
yg = powerg1(x,ai,xm)
plt.plot(x,yg,label = r"$\alpha$="+"{:.2f}".format(ai))

plt.xlabel('z')
plt.ylabel("$\phi \' (z) $")
plt.title('Derivative of the Power-Linear Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.savefig('d_plaf.png',dpi=500,bbox_inches='tight')
plt.show()
 Activation Functions and Universal Approximation Theory 297

Figure 7.21: The derivative of the power-linear activation function. It is positive,

continuous, piecewise smooth, and differentiable.

Yes! The derivative is caped. We observe also that the smaller the α, the
large the derivative value. This may be useful in mitigating the vanishing
derivative problem in training deep nets. The vanishing derivative occurs
when back-propagation is used to update the weights and biases using a
gradient descent technique (to be discussed in great detail in Chapter 9),
the derivatives get smaller and smaller due to the multiplication of the
derivatives from the chain rule of differentiation (as shown in Chapter 8).
With the power-linear function, one may get some help in reducing the
vanishing rate, if it is placed properly in these layers.
We see also that the derivative has a plateau in the linear range. This
can be improved further by the power-quadratic function.

7.9.4 Power-quadratic function

Finally, we examine the power-quadratic function. It is further modified from

the power-linear function, by replacing the linear function portion with two
pieces of quadratic function, so that the derivative can change linearly. This
is done in the following process.
First, we compute the derivative at z = 0, which is the same as that for
the power-linear function at z = 0:

(α−1)
km = zm (7.31)
298 Machine Learning with Python: Theory and Applications

Next, we compute the derivative at z = zm , which gives,

(α−1)
kp = αzm (7.32)

We now assume that the derivative varies linearly in [−zm , zm ], which give
us
⎧
⎪ kp − km
⎪
⎨− z + km −zm ≤ z < 0
 zm
φ (z) = (7.33)
⎪
⎪ kp − km
⎩ z + km 0 ≤ z < zm
zm

We finally integrate the above function, and using the condition that it has
to be zero at z = 0, we obtain,
⎧
⎪ kp − km 2
⎪
⎨− z + km z −zm ≤ z < 0
2zm
φ(z) = (7.34)
⎪
⎪ kp − km 2
⎩ z + km z 0 ≤ z < zm
2zm

Adding up the other two pieces on both side of the equation, we now have
the full definition for the power-quadratic function:
⎧
α
⎪ −(−z)
⎪
⎪
z < −zm
⎪
⎪
⎪
⎪ kp − km 2
⎨ − 2z
⎪ z + km z −zm ≤ z < 0
m
φ(z) = (7.35)
⎪
⎪ kp − km 2
⎪
⎪ z + km z 0 ≤ z < zm
⎪
⎪ 2zm
⎪
⎪
⎩ α
z z ≥ zm

The derivative of the power-quadratic function becomes,

⎧
⎪
⎪ α(−z)α−1 z < −zm
⎪
⎪
⎪
⎪ kp − km
⎪
⎨ − zm z + km
⎪ −zm ≤ z < 0
φ (z) = (7.36)
⎪
⎪ kp − km
⎪
⎪ z + km 0 ≤ z < zm
⎪
⎪ zm
⎪
⎪
⎩
α(z)α−1 z ≥ zm
 Activation Functions and Universal Approximation Theory 299

Let us examine the power-quadratic function, using the following code.

#Power-quadratic(2) function
def apowerf2(z,alpha,xm):
a1 = alpha-1.
fm = xm**alpha
km = fm/xm
kp= alpha*xm**a1
kg = (kp-km)/xm
f = [-(-zi)**alpha if (zi<-xm) else km*zi-0.5*kg*zi**2\
if (-xm<=zi<0.) else km*zi+0.5*kg*zi**2 if (0.<=zi<=xm)\
else zi**alpha for zi in z]
return f

x0 = np.array([0.0,2.0])
x = np.arange(-2, 2, 0.02)
#alfa = np.arange(0.2, 1.2, .2)
alfa = 1.0/np.arange(1, 9, 1.0)

xm = 0.1
for ai in alfa:
y = apowerf2(x,ai,xm)
plt.plot(x,y,label = r"$\alpha$="+"{:.2f}".format(ai))

plt.xlabel('z')
plt.ylabel("$\phi(z) $")
plt.title('Power-Quadratic Activation Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.savefig('pqaf.png',dpi=500,bbox_inches='tight')
plt.show()

Figure 7.22: The power-quadratic activation function. It is monotonic, continuous,

piecewise smooth, and differentiable.
300 Machine Learning with Python: Theory and Applications

#Power-quadratic(2) function, derivative

def apowerg2(z,alpha,xm):
a1 = alpha-1.
fm = xm**alpha
km = fm/xm
kp= alpha*xm**a1
kg = (kp-km)/xm
g = [alpha*(-zi)**a1 if zi<-xm else -kg*zi+km\
if (-xm<=zi < 0.) else kg*zi+km if (0.<=zi <xm)\
else alpha*zi**a1 for zi in z]
return g

#x0 = np.array([0.0,2.0])
x = np.arange(-1, 1, 0.002)
#alfa = np.arange(0.2, 1.2, .2)
alfa = 1.0/np.arange(1, 9, 1.0)

xm = 0.1
for ai in alfa:
yg = apowerg2(x,ai,xm)
plt.plot(x,yg,label = r"$\alpha$="+"{:.2f}".format(ai))

plt.xlabel('z')
plt.ylabel("$\phi \ '(z) $")
plt.title('Derivative Power-Quadratic Activation Function')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.savefig('d_pqaf.png',dpi=500,bbox_inches='tight')
plt.show()
 Activation Functions and Universal Approximation Theory 301

Figure 7.23: The derivative of the power-quadratic activation function. It is positive,

continuous, piecewise smooth, and differentiable.

The plateau in the linear range is now removed with a nice linearly varying
derivative. We still observe that the smaller the α, the large the derivative
value. This may be useful in mitigating the vanishing derivative problem in
training deepnets. Thorough tests are required for this proposed set of power
functions, in its application to practical problems.

References

[1] G. Cybenko, Approximation by superpositions of a sigmoidal function, Mathematics

of Control, Signals and Systems, 2, 303–314, 1989.
[2] G.R. Liu, A neural element method, International Journal of Computational Methods,
17(07), 2050021, 2020.
[3] G.R. Liu, Mesh Free Methods: Moving Beyond the Finite Element Method, Taylor and
Francis Group, New York, 2010.
[4] G.R. Liu and S.S. Quek, The finite element method: A practical course, Butterworth-
Heinemann, London, 2013.
[5] V. Nair and G.E. Hinton, Rectified Linear Units Improve Restricted Boltzmann
Machines, Proceedings of the 27th International Conference on Machine Learning,
2010.
MACHINE LEARNING
WITH PYTHON
 Chapter 8

Automatic Differentiation and Autograd

8.1 General Issues on Optimization and Minimization

In machine learning, we would like the models such as neural networks to

learn by mistakes, through feeding it with teaching (or training) samples,
so that it knows where the errors are made. The measurement of error uses
the so-called loss function that evaluates, in some predefined formulas, the
difference between the predicted results (the output of current model) and
the ground truth (labels) given by training samples.
Loss function is sometimes also called cost function or objective function.
They all are essentially the same, but can have different names to make bet-
ter sense for problems of different purposes. The name of “objective function”
is more neutral and can be used for both optimization and minimization
settings. Loss or cost function implies a minimization problem, because the
purpose is clearly for minimizing the cost or loss. From now on, we do not
distinguish these names in our discussion, but use more often the name of loss
function for machine learning models. The smaller the value of the loss func-
tion, we generally say the better the model (in training accuracy). Therefore,
the training process is essentially a process of minimizing the loss function.
Optimization is equivalent to minimization in mathematics, because one
can simply flip the sign of the objective function and then the optimization
problem becomes a minimization problem or vice versa. Therefore, we do not
distinguish the optimization and minimization techniques in theory. Care is
needed when calling a routine function from an existing module to make sure
whether it is for optimization or minimization. One may need to flip the sign
of the objection function accordingly, before feeding it to the function.
The next question would be how to design optimization (or minimization)
techniques, to effectively find the maximum (or minimum). In the past,

303
304 Machine Learning with Python: Theory and Applications

quite a number of optimization techniques (https://en.wikipedia.org/wiki/

Mathematical optimization) had been developed. There are techniques that
use gradient information of the objective function, and those that do not.
In general, the techniques that used the gradient information work more
efficiently, and hence it is used whenever the gradient information can be
made available in a cost-effective manner. For machine learning models such
as NNs, gradient-based techniques have been found effective. This is because
the prediction functions (Chapter 5) and the activation function (Chapter 7)
used in NNs are continuous functions of the learning parameters, and hence
the gradient is computable. The applicability of a gradient-based minimizer
depends largely on whether we can compute the gradient of the loss function
efficiently.
In this chapter, we consider functions in general areas of study in
science and engineering, including machine learning. Therefore, without
losing generality, the function to be differentiated is denoted in general as
f (x) with x being the independent variables. Also, x is defined as a column
vector as in a more conventional way. In machine learning, however, our
function would be the loss function L(ŵ) and the independent variable would
be the learning parameters ŵ. Care is needed to avoid confusion.
There are three major types of methods to compute the gradient of a
function: analytical method, numerical method, and automatic differentia-
tion method or Autograd. We first briefly discuss the analytical method.

8.2 Analytic Differentiation

The analytical method explicitly gives the formulas for the derivatives of
functions. Table 8.1 lists some scalar functions often used in ML, f (x), with
variable x also being a scalar, together with their derivatives fdx (x)
. These
formulas can be coded directly for computing the derivatives, and symbolic
computational tools such as SymPy can be used to obtain the expression of
the derivatives. However, such formulas may not be available for all functions.
Also, not all functions can be explicitly written in such a simple closed form,

Table 8.1: Analytical formulas for derivatives of simple functions (c, n are constants).
1 ecx − e−cx
f (x) = c cxn ecx log(cx)
1 + e−cx ecx + e−cx
  cx 2 
df 1 ce−cx e − e−cx
=0 ncxn−1 cecx c 1−
dx x (ecx + e−cx )2 ecx + e−cx
 Automatic Differentiation and Autograd 305

let alone their derivatives. Most importantly, expressions of the derivatives

of many functions can be very complicated and hence difficult to implement
in a code.

8.3 Numerical Differentiation

Numerical methods compute gradient or derivative of a function using essen-

tially the finite difference method (FDM), curve fitting, and interpolation
techniques. The smoothing domain based methods used in the smoothed
finite element methods [1] and the meshfree methods [2] have been found
effective for computing the gradient of functions in lower dimensions. Numer-
ical methods usually give approximated gradient results for all differentiable
functions, as long as their numerical values are available at the nodes of a
mesh or at a set of points given in the vicinity of the location of interest.
However, it can be a challenge for high dimensional space, and also for a
chain of functions that we often encounter in machine learning. Although
the approximation in the results may not be an issue, the complexity in the
implementation could be. The numerical method is not used in this book for
creating ML models.

8.4 Automatic Differentiation

8.4.1 The concept of automatic or algorithmic differentiation

Automatic differentiation (AD) is sometimes also called algorithmic dif-

ferentiation (http://www.autodiff.org/?module=Publications&submenu=li
st%20publications&id=neidinger2010ita). It is a technique to evaluate the
value of differentiation via a computer algorithm. The essential idea is
to simultaneously compute the involved functions and their derivatives into
numerical numbers, so that the arithmetic operations among them can be
done simply in numerical numbers, avoiding such operations at expression
level. This drastically simplifies coding. The results are numerical but is exact
(in theory with machine accuracy), because all the rules of differentiation are
still followed exactly, which is fundamentally different from the numerical
differentiation mentioned earlier. It has the following features:

• It uses the chain rule of differentiation, so that the differentiation of the

final loss function becomes a product of the local derivatives (derivatives
of local function at the “node” in a computational graph).
306 Machine Learning with Python: Theory and Applications

• It constructs a process to split the complicated functions into elementary

arithmetic operations (addition, subtraction, multiplication, division,
analytic differentiation of simple functions, etc.), so that derivatives for
those functions can be computed using exact analytical formulas.
• It records the results of differentiation with respect to a variable, together
with the variable itself (as one of its attributes), for convenient recording,
automation and computation. This is because the gradient and the variable
have the same shape.

In neural network-based machine learning models, the AD is one of the key

techniques for its success. Types of algorithms based on AD are readily
available, and often known as autograd. It can be conveniently used to
compute the gradients for a number of parameter variables in automatic
fashion. Many libraries allow users to visualize a computational graph of the
net constructed. Such a graph is essentially a logical path for computing the
local derivatives in the chained nodes.

8.4.2 Differentiation of a function with respect

to a vector and matrix

The loss function for machine learning is usually in a form of scalar function.
Even if it is not, one can always make it a scalar function by, for example,
measuring its norm in some way. However, their variables can often be vectors
or matrices. Consider a scalar function f (x) ∈ R1 defined in a P -dimensional
space, the variable vector x will be in RP . The gradient of the function ∇f (x)
is a vector and will also be in RP ,
⎡ ⎤
∂f (x)
⎢ ∂x1 ⎥
⎢ ⎥
⎢ ⎥
⎢ ∂f (x) ⎥
∂f (x) ∂f (x) ⎢ ⎥
∇f (x) = = ⎡ ⎤ =⎢ ⎢
∂x2 ⎥
⎥ (8.1)
∂x x1 ⎢ .. ⎥
⎢x ⎥ ⎢ . ⎥
⎢ 2⎥ ⎢ ⎥
∂⎢ ⎥ ⎣ ∂f (x) ⎦
⎢ .. ⎥
⎣ . ⎦ ∂xm
xm

When the variables are in a matrix form, we shall have

 Automatic Differentiation and Autograd 307

⎡ ∂f (x) ∂f (x) ∂f (x)

⎤
...
⎢ ∂x11 ∂x12 ∂x1p
⎥
⎢ ∂f (x) ∂f (x) ∂f (x) ⎥
∂f (x) ∂f (x) ⎢ ∂x21 ∂x22 ... ∂x2p ⎥
⎢ ⎥
∇f (x) = = ⎡ ⎤ =⎢ . .. ⎥
∂x x11 x12 . . . x1p ⎢ . .. .. ⎥
⎢ . . . . ⎥
⎢x x22 . . . x2p ⎥ ⎣ ⎦
⎢ 21 ⎥ ∂f (x) ∂f (x) ∂f (x)
∂⎢
⎢ .. .. ..
⎥
.. ⎥ ∂xm1 ∂xm2 ... ∂xmp
⎣ . . . . ⎦
xm1 xm2 . . . xmp
(8.2)
When we seek the derivatives for a scalar function with respect to a vector
or matrix of variables, what this really means is that we want the (partial)
derivatives of the function with respect to each one of the variables in the
vector (both row or column) or matrix. Therefore, all we need to do is to
perform the differentiation of the function with respect to these variables
one by one, as we do for the scalar variable case shown in Table 8.1.
This can be easily done using the element-wise operations in Python. As
a result, we simply replace x in Table 8.1 by the vector or matrix x. In this
case, the derivative of the function shall become a vector or matrix that
contains the derivatives of the function with respect to each variable. Such a
derivative vector or matrix can have the same shape as that of the variable
vector of matrix, can be computed simultaneously, and stored together as
an additional attribute. It is essentially what is done in Autograd.
Table 8.2 gives some often used functions and their gradient. In machine
learning algorithms, we need to frequently compute the norm of vectors, and
the summation of the elements in vectors, which results in special functions:
f (x) = x2 , f (x) = sum(x). These special functions are all scalars, and
the above arguments still hold. Because their variables are a vector, the
gradient of these special functions will become a vector. For sum function
of a vector x, the gradient becomes a vector of ones. For the squared-norm
function of a vector x, the gradient becomes two times the x, which is also a
vector. In addition, when f is a constant vector c, its gradient becomes the
vector of zeros, as shown in Table 8.2.
In Table 8.2, b is a contact, c is a contact, c is a vector of constants, and
C is a matrix of constants. All these constants do not change with x. We
assume all the matrices have compatible shape (dimensions). The C-scalar
can also be a norm (C-norm, positive), if C is positive definite. The same is
true for X-scalar. The ∇f (x) has the same shape of x.
308 Machine Learning with Python: Theory and Applications

Table 8.2: Analytical formulas for the gradients of some useful scalar functions.
Function Name f (x) ∇f (x)

Constant c 0
 
Linear x c=c x c
x c+b x c+b
General exponential e ce
General logarithm log(x c + b) −c
x c+b

1 ce−x c+b
Sigmoid   
1+e−x c+b (ex c +e−x c+b )2
  2 
   
ex c+b −e−x c+b ex c+b −e−x c+b
tanh   c 1−  
ex c+b +e−x c+b ex c+b +e−x c+b

summation sum(x) 1
norm (quadratic in x) x2 = x x 2x
C-scalar (quadratic in x) x Cx 2Cx
X-scalar (linear in X) c Xc cc

Table 8.1 is obvious and writing it down needs only college calculus.
However, writing Table 8.2 requires careful derivation. The best way may be
using the tensor calculus, by first converting the matrix notation to (Ricci)
index notation, performing the differentiations, and then converting back
to matrix form. Using the tensor calculus, one can further derive formulas
for vector or even matrix functions with tensor variables. Here, we will not
discuss this in detail.

8.5 Autograd Implemented in Numpy

Autograd brings the automatic differentiation into Numpy (Oliphant, 2007)

and Scipy (Jones et al., 2001). Let us practice this. We first use the
autograd.numpy to compute the derivatives of given analytical functions.
For this, we need to install autograd (using pip install autograd or conda
install-c conda-forge autograd). We then import the autograd.numpy model
and grad() function.

#import numpy as np
import autograd.numpy as np # This is a thinly-wrapped numpy
from autograd import grad # for computing grad(f)
import matplotlib.pyplot as plt
%matplotlib inline
 Automatic Differentiation and Autograd 309

def f_sigma(z): # The sigmoid function

return 1./(1.+np.exp(-z))
def dfsa(z):
return f_sigma(z)*(1.-f_sigma(z)) # analytical derivative

df1 = grad(f_sigma) # function for the 1st derivative of f

df2 = grad(df1) # for the 2nd derivative of f
df3 = grad(df2) # ......
df4 = grad(df3)
df5 = grad(df4)
z = np.linspace(-5,5,200) # create values of arguments, z
fi = f_sigma(z) # compute the discrete values of f()
df1i = []
for i in range(len(z)): # compute df1 by passing z one-by-one
df1i.append(df1(z[i]))

df1ia = grad(f_sigma)(z[0])
df2i = list(map(df2,z)) # compute df2
plt.figure(figsize=(4.5, 3.0),dpi=100)
plt.scatter(z, df1i,c='green',label='Autograd',alpha=0.2)
plt.plot(z, dfsa(z),c='red',label='Analytical')
plt.xlabel('z')
plt.ylabel(r"$\sigma'(z)$")
plt.title('Derivative of Sigmoid: analytic vs. autograd')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.show ()

plt.figure(figsize=(4.5, 3.0),dpi=100)
plt.plot(z, fi,label='Function')
plt.plot(z, df1i,label='1st derivative')
plt.plot(z, df2i,label='2nd derivative')
plt.plot(z, list(map(df3,z)),label='3rd derivative')
plt.plot(z, list(map(df4,z)),label='4th derivative')
plt.xlabel('z')
plt.ylabel(r'$\sigma(z) $'+' and its derivatives')
plt.title('Derivatives of sigmoid function via autograd')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.show ()
310 Machine Learning with Python: Theory and Applications

Figure 8.1: Computed derivatives of functions using autograd.

As seen from Fig. 8.1, autograd is a powerful means to compute deriva-

tives of functions. It is produces exact derivatives (to machine accuracy), and
is capable of computing higher order of derivatives so long as the function
is differentiable to the order.

8.6 Autograd Implemented in the MXNet

Autograd has been implemented in many well-established modules. Next,

we use the mxnet.autograd module to demonstrate how to automatically
calculate derivatives of a loss function with respect to its variables. Once
a parameter variable is specified for gradients, and the chain of the sub-
functions for the loss function is defined, the autograd builds a graph on the
fly (https://d2l.ai/chapter preliminaries/autograd.html). It then computes
 Automatic Differentiation and Autograd 311

backwards the gradients of the loss function with respect to the variables.
The following discussion and the codes are written in reference to these at
mxnet-the-straight-dope (https://github.com/zackchase/mxnet-the-straigh
t-dope), under the Apache-2.0 License:

import numpy as np
import mxnet as mx
from mxnet import nd, autograd
mx.random.seed(1)

8.6.1 Gradients of scalar functions with simple variable

As a toy example, we take the gradient of a simple function defined as

f = 8(x2 ) for x ∈ R1 (8.3)

with respect to parameter variable x in one dimension. The gradient is just

the derivative in this one-dimensional case with only x, because there is
only one parameter variable. For this simple function, we know the gradient
(derivative) can be calculated analytically using the following equation,

df (x)
= 16x for x ∈ R1 (8.4)
dx

for any given x value in R1 .

In this demonstration, we start by assigning the values of x, at which
the derivatives of f (x) can be computed. If x is given in a vector or matrix
or in general in a tensor, the corresponding derivatives are computed and
will be in the same shape. This is quite similar to the general element-wise
operations in Python: the same differentiation is done at all the x values in
the tensor. We now define a matrix of x values as follows:

x=nd.array([[1,5,10],[20,18,28]]) # initialization of x
# in a matrix (tensor)

At all these values of x in the above matrix, we can easily compute the
gradient using analytical formula given in Eq. (8.4).

print('x=',x)
grad_at_x = 16.0*x # analytical gradient
print('Analytical grad_at_x=', grad_at_x)
312 Machine Learning with Python: Theory and Applications

x=
[[ 1. 5. 10.]
[20. 18. 28.]]
<NDArray 2x3 @cpu(0)>
Analytical grad at x=
[[ 16. 80. 160.]
[320. 288. 448.]]
<NDArray 2x3 @cpu(0)>

We note that the shape of x and grad at x is the same.

In MXNet, the gradient computation requires three steps.
First, we request a gradient for a variable, say x. We need it to be stored
in x.grad(). This is done by first invoking

x.attach_grad() # request for gradient w.r.t

# all entries in matrix x

Second, we need to define the function f , which is often given in a chain

of functions (x− > g− > f , as shown in the example below).

with autograd.record(): # record down the path

g = x * 8 # x -> g
f = g * x # g -> f

This defines a path for carrying out the differentiation and the related
evaluations. This path of differentiation is also called graph. The definition
of the chain or path builds a computation graph. The differentiation is then
done using the well-known chain rule of differentiation. MXNet builds such
a graph on the fly, using a kind of recording “device”, which record the path.
Third, with the path defined, we can finally invoke

f.backward() # compute the gradient in a back-propagation

which calls for differentiations to be carried out backwards in the defined

path, known as backprop in machine learning. To this end, the gradient
values are computed and stored at x.grad, as an attribute of object x.
We finally print out the gradient values stored at x.grad.

print('At x=',x)
print('The gradient values of f at x=', x.grad) # autograd
print('Analytical grad_at_x=', grad_at_x) # analytical
 Automatic Differentiation and Autograd 313

At x=
[[ 1. 5. 10.]
[20. 18. 28.]]
<NDArray 2x3 @cpu(0)>
The gradient values of f at x=
[[ 16. 80. 160.]
[320. 288. 448.]]
<NDArray 2x3 @cpu(0)>
Analytical grad at x=
[[ 16. 80. 160.]
[320. 288. 448.]]
<NDArray 2x3 @cpu(0)>

We note again that the shape of x and x.grad is the same. x stores all these
x values, and x.grad stores the corresponding gradients at these x values. We
can see that the gradient value obtained at x using autograd is exactly the
same as those we got using the analytic formula.

8.6.2 Gradients of scalar functions in high dimensions

Let us now consider the gradients of a scalar function with multiple variables
(multiple dimensional problems). We compute the gradients of some scalar
functions listed in Table 8.2. The first is the linear function defined in
p-dimensional space.

f (x) = x c for x ∈ Rp (8.5)

where c is vector of constants. It is also in Rp , so that f (x) become a scalar.

c = nd.array([.1, .8, .18]) # 3 constants

x = nd.array([1., 2., 3.]) # 3 independent variables
x.attach_grad() # request gradient w.r.t x

with autograd.record(): # record the path

f = x.T * c # or f = nd.dot(x, c)

f.backward() # compute gradient

print('At x=',x, ' Constants c = ',c)
print('The gradient values of f at x=', x.grad)
print('The analytical gradient values of f at x=', c)
314 Machine Learning with Python: Theory and Applications

At x=
[1. 2. 3.]
<NDArray 3 @cpu(0)> Constants c =
[0.1 0.8 0.18]
<NDArray 3 @cpu(0)>
The gradient values of f at x=
[0.1 0.8 0.18]
<NDArray 3 @cpu(0)>
The analytical gradient values of f at x=
[0.1 0.8 0.18]
<NDArray 3 @cpu(0)>

We observe the agreement of the results obtained using autograd and

the analytical method. We now examine the general exponential function in
p-dimensional space.
f (x) = exp(x c) for x ∈ Rp (8.6)
where c is vector of constants. It is also in Rp .
x.attach_grad() # Request gradient w.r.t x
# that is a vector
with autograd.record(): # Record down the path
f = nd.exp(nd.dot(x,c))

f.backward() # compute gradient

print('At x=',x, ' Constants c = ',c)
print('The gradient values of f at x=', x.grad)
print('The analytical gradient values of f at x=', \
c*nd.exp(nd.dot(x,c)))
At x=
[1. 2. 3.]
<NDArray 3 @cpu(0)> Constants c =
[0.1 0.8 0.18]
<NDArray 3 @cpu(0)>
The gradient values of f at x=
[0.93933314 7.514665 1.6907997 ]
<NDArray 3 @cpu(0)>
The analytical gradient values of f at x=
[0.93933314 7.514665 1.6907997 ]
<NDArray 3 @cpu(0)>
 Automatic Differentiation and Autograd 315

We observe again the agreement of the results obtained using autograd

and the analytical method (in machine accuracy). We now examine the
general logarithmic function in p-dimensional space.

f (x) = log(x c) for x ∈ Rp (8.7)

where c is vector of constants. It is also in Rp .

c = nd.array([.1, .8, .18]) # 3 constants
x = nd.array([1, 2, 3]) # 3 independent variables
x.attach_grad() # request gradient w.r.t. x

with autograd.record(): # record down the path

f = nd.log(nd.dot(x,c))

f.backward() # compute gradient via back-prop.

print('At x=',x, ' Constants c = ',c)
print('The gradient values of f at x=', x.grad)
print('The analytical gradient values of f at x=',\
c/(nd.dot(x,c)))
At x=
[1. 2. 3.]
<NDArray 3 @cpu(0)> Constants c =
[0.1 0.8 0.18]
<NDArray 3 @cpu(0)>
The gradient values of f at x=
[0.04464286 0.35714287 0.08035714]
<NDArray 3 @cpu(0)>
The analytical gradient values of f at x=
[0.04464286 0.35714287 0.08035715]
<NDArray 3 @cpu(0)>

Now, the general form of the sigmoid function is in Table 8.2.

x.attach_grad() # request gradient w.r.t. x

def sigma_nd(z): # Define the sigmoid function

return 1./(1.+nd.exp(-z))

def ds_nd(z): # analytical derivative

return sigma_nd(z)*(1.-sigma_nd(z))
316 Machine Learning with Python: Theory and Applications

with autograd.record(): # record down the path

f = sigma_nd(nd.dot(x,c))

f.backward() # compute gradient via back-prop.

print('At x=',x, ' Constants c = ',c)
print('The gradient values of f at x=', x.grad)
print('The analytical gradient values of f at x=',\
c*ds_nd(nd.dot(x,c)))

At x=
[1. 2. 3.]
<NDArray 3 @cpu(0)> Constants c =
[0.1 0.8 0.18]
<NDArray 3 @cpu(0)>
The gradient values of f at x=
[0.00869581 0.06956648 0.01565246]
<NDArray 3 @cpu(0)>
The analytical gradient values of f at x=
[0.00869581 0.0695665 0.01565246]
<NDArray 3 @cpu(0)>
Finally, the general form of the tanh function is in Table 8.2.

c = nd.array([.1, .8, .18]) # 3 constants

x = nd.array([1., 2., 3.]) # 3 variables
x.attach_grad() # request gradient

def tanh_nd(z): # Define the function

return (nd.exp(z)-nd.exp(-z))/(nd.exp(z)+nd.exp(-z))

def dtanh_nd(z):
return (1.-tanh_nd(z)**2) # analytical derivative

with autograd.record(): # record down the path

f = tanh_nd(nd.dot(x,c))

f.backward() # compute gradient

print('At x=',x, ' Constants c = ',c)
print('The gradient values of f at x=', x.grad)
print('The analytical gradient values of f at x=',\
c*dtanh_nd(nd.dot(x,c)))
 Automatic Differentiation and Autograd 317

At x=
[1. 2. 3.]
<NDArray 3 @cpu(0)> Constants c =
[0.1 0.8 0.18]
<NDArray 3 @cpu(0)>
The gradient values of f at x=
[0.00443234 0.03545868 0.0079782 ]
<NDArray 3 @cpu(0)>
The analytical gradient values of f at x=
[0.00443233 0.03545866 0.0079782 ]
<NDArray 3 @cpu(0)>

Next, we compute the gradient of a summation function with a vector of

variables.
from mxnet import autograd, nd
x1d = nd.arange(4) # create a 1D vector
x1d.attach_grad() # request gradients
print(x1d)

with autograd.record():
f = x1d.sum() # summation function.

f.backward() # compute gradient via back-prop.

x1d.grad
[0. 1. 2. 3.]
<NDArray 4 @cpu(0)>

[1. 1. 1. 1.]
<NDArray 4 @cpu(0)>

It gives a vector of all ones, as expected (see Table 8.2).

Dot-product (the inner product) is one of the most often used operations,
which also results in a scalar function. It is the same as the squared-
norm function in Table 8.2. We compute the squared-norm function and
its derivatives with respect to its vector variable.

from mxnet import autograd, nd

x1d = nd.arange(4) # create a 1D vector
x1d.attach_grad() # request gradients
x1d
318 Machine Learning with Python: Theory and Applications

[0. 1. 2. 3.]
<NDArray 4 @cpu(0)>

x1d.grad # current gradient value

[0. 0. 0. 0.]
<NDArray 4 @cpu(0)>
with autograd.record(): # define computation path
f = 8 * nd.dot(x1d, x1d)
#f = (x1d.norm())**2 # alternative formula

f.backward() # compute gradient

x1d.grad # query the results.
[ 0. 16. 32. 48.]
<NDArray 4 @cpu(0)>

2 * x1d # analytical formula

[0. 2. 4. 6.]
<NDArray 4 @cpu(0)>

x1d.grad == 2* x1d # correct?

[1. 0. 0. 0.]
<NDArray 4 @cpu(0)>

All correct.

8.6.3 Gradients of scalar functions with quadratic

variables in high dimensions
Consider the gradients of the scalar functions that vary quadratically with
variables in high dimensions. A typical quadratic function defined in a
p-dimensional space can have the following form:

f (x) = x Cx for x ∈ Rp (8.8)

where C is matrix of a constant with a dimension of p × p. It is clear that f (x)

is a scalar depending on x quadratically. We now use autograd to compute
the gradient of this function, and compare it with the analytical solution
given in Table 8.2.
 Automatic Differentiation and Autograd 319

c = nd.array([[1.1, .8, .18],[.8, 2.8, .18],[.18, .18, 3.8]])

# 3 by 3 matrix of constants
x = nd.array([1., 2., 3.]) # 3 independent variables
x.attach_grad() # request for gradient
with autograd.record(): # record down the path
g = nd.dot(x,c)
f = nd.dot(g,x)
f.backward() # compute gradient
print('At x=',x, ' Constants c = ',c)
print('The gradient values of f at x=', x.grad)
print('The analytical gradient values of f at x=',\
2*nd.dot(c,x))

At x=
[1. 2. 3.]
<NDArray 3 @cpu(0)> Constants c =
[[1.1 0.8 0.18]
[0.8 2.8 0.18]
[0.18 0.18 3.8 ]]
<NDArray 3x3 @cpu(0)>
The gradient values of f at x=
[ 6.48 13.88 23.88]
<NDArray 3 @cpu(0)>
The analytical gradient values of f at x=
[ 6.48 13.88 23.88]
<NDArray 3 @cpu(0)>

The autograd gradient is in agreement with the analytical gradient.

8.6.4 Gradient of scalar function with a matrix of

variables in high dimensions
Finally, we construct a scalar function with a matrix of variables. One of
such can be defined as follows:

f (X) = c Xc for X ∈ Rp×p (8.9)

where X is a matrix of p × p variables, and c is a vector of p constants that

do not change with any of the variables in X. We use autograd to compute
the gradient of this function, and compare it with the analytical solution
given in Table 8.2.
320 Machine Learning with Python: Theory and Applications

c = nd.array([1., 2., 3.]) # vector of 3 constants

c1 = nd.array([[1., 2., 3.]])
X = nd.array([[1.1, .8, .18],[.8, 2.8, .18],[.18, .18, 3.8]])
# 3 by 3 matrix of independent variables
X.attach_grad() # request for gradient
with autograd.record(): # record down the path
g = nd.dot(X,c)
f = nd.dot(g,c)
f.backward() # compute gradient
print('At x=',X, ' Constants c = ',c)
print('Gradient values of f at x=', X.grad)
print('Analytical gradient values of f at x=',\
nd.linalg.gemm2(c1.T,c1))

At x=
[[1.1 0.8 0.18]
[0.8 2.8 0.18]
[0.18 0.18 3.8 ]]
<NDArray 3x3 @cpu(0)> Constants c =
[1. 2. 3.]
<NDArray 3 @cpu(0)>
Gradient values of f at x=
[[1. 2. 3.]
[2. 4. 6.]
[3. 6. 9.]]
<NDArray 3x3 @cpu(0)>

Analytical gradient values of f at x=

[[1. 2. 3.]
[2. 4. 6.]
[3. 6. 9.]]
<NDArray 3x3 @cpu(0)>

8.6.5 Head gradient

In the above example, the chain rule can be written as

d df (g) dg(x)
f (g(x)) = (8.10)
dx dg dx
 Automatic Differentiation and Autograd 321

The function can be defined in more than one layer, which forms a chain
functions, as follows:
g =x∗2 (8.11)
f =g∗x (8.12)
h=f (8.13)

In this case, we have

d d dh(f ) df (g) dg(x)

f (g(x)) = h(f (g(x))) = (8.14)
dx dx df dg dx

The first part dhdf is called head gradient, which is often used in machine
learning. For the above simple example, we have

dh
=1 (8.15)
df

If we know the head gradient of the function, the gradient of the function
can be computed using f.backward(), by feeding the head gradient as an
input argument:

with autograd.record():
g = x * 2
f = g * x
h = f

head_gradient=nd.ones_like(g)
# default. hence is not necessary.
h.backward(head_gradient)
# this allows adding in a constant matrix to the head.
print('At x=',x, '\n The gradient values of f at x=', x.grad)

At x=
[1. 2. 3.]
<NDArray 3 @cpu(0)>
The gradient values of f at x=
[ 4. 8. 12.]
<NDArray 3 @cpu(0)>
322 Machine Learning with Python: Theory and Applications

with autograd.record():
g = 2* x ** 2
h = g
head_gradient=nd.ones_like(g)
h.backward(head_gradient)
print('At x=',x, '\n The gradient values of f at x=', x.grad)
At x=
[1. 2. 3.]
<NDArray 3 @cpu(0)>
The gradient values of f at x=
[ 4. 8. 12.]
<NDArray 3 @cpu(0)>

The above example is simple, but demonstrates a very powerful tool for
computing the exact gradients of complicated loss/objective functions one
may encounter in real-life applications to train a neural network. As long as
we can structure the loss function that is a chain of sub-functions, autograd
can be used. All one needs is to tell the code which are the parameters
(variables) that need their gradients to be “attached” to, and then “record”
the chain relationship of the sub-functions. The code will compute the
gradients, in a back-propagation process upon request, regardless of how
many layers there may be.

8.7 Gradients for Functions with Conditions

In Python, we can build complicated functions, where sub-functions are con-

trolled by conditions, allowing nonlinearity to be built into the formulations.
Autograd can also be used to compute the gradients for such functions.
Below is an example.

a = nd.random_normal(shape=6)
a
[ 0.03629 -0.49024 -0.9501 0.03751 -0.7298 -2.0401056 ]
<NDArray 6 @cpu(0)>

a.attach_grad()
with autograd.record():
b = a * 88 + 5.0
 Automatic Differentiation and Autograd 323

while (nd.norm(b) < 8888).asscalar():

b = b * 88
if (mx.nd.sum(b) > 0).asscalar():
c = b + 88
else:
c = 800 * b

head_gradient = nd.array([0.5, 1.2, .18, 0.3, 0.6, 0.9])

c.backward(head_gradient)

print(a.grad)

[3097600. 7434240.5 1115136. 1858560.1 3717120.2 5575680. ]

<NDArray 6 @cpu(0)>

Note there are operations and functions that may not be supported
by autograd, such as in-place operations, depending on the module used.
Readers are referred the online [document] (https://github.com/HIPS/
autograd/blob/master/docs/tutorial.md) and the links therein for more
detail, when encountering strange behavior.

8.8 Example: Gradients of an L2 Loss Function for

a Single Neuron

Let us consider the simplest network with only a single neuron, as shown in
Fig. 8.2.

Figure 8.2: A 1-1-1 neuron network: One input layer, one hidden layer, and one output
layer, each with a single neuron.

We first use hand calculation to get all the values at these local nodes
and stages in the feed-forward process, together with all the gradients at
the corresponding locations in the back-propagation process. The results are
shown in Fig. 8.3.
324 Machine Learning with Python: Theory and Applications

Figure 8.3: Feed-forward and Back-propagation for the simplest net with only a single
neuron. The net is split into five local nodes (red elliptic circles), as shown in the middle
of the figure. Feed-forward process: The input value of x, current values of the learning
parameters w and b, as well as the label y value are all arbitrarily set for the calculation.
All these intermediate values calculated using these values in the feed-forward process are
given above the arrows leading to these nodes. Back-propagation: All the derivatives at
the local nodes are computed backwards and the values are given in the bottom of the
figure. The final gradient of the L2 function with respect to w is calculated. The same can
be done with respect to b, and the result should be the same for this simple case.

In this case, we use the least squared error as the loss function (known
as the L2 loss function) with φ(z) being the activation function.

L = (φ(x · w + b) − y)2 (8.16)

where y is a given label at the out. In this case, we shall have

∂L
= 2(x · w + b − y)x · φ (z) (8.17)
∂w
and
∂L
= 2(x · w + b − y) · φ (z) (8.18)
∂b
Let us write a code to compute all these, using the above analytic formulae
and the autograd.

import mxnet as mx
from mxnet import nd, autograd, gluon

def logistic(z): #Define the sigmoid function

return 1. / (1. + nd.exp(-z))
 Automatic Differentiation and Autograd 325

def logisticg(z): #Analytic gradient of the sigmoid

return logistic(z)*(1-logistic(z))

def loss2(x,w,b,y): #Define L2 loss function

y_hat = logistic(nd.dot(x,w)+b)
return (y_hat-y)**2

def lossdw(x,w,b,y): #analytic derivative of L2 w.r.t w

y_hat = logistic(nd.dot(x,w)+b)
return 2*(y_hat-y)*x*logisticg(nd.dot(x,w)+b)

def lossdb(x,w,b,y): #analytic derivative of L2 w.r.t b

y_hat = logistic(nd.dot(x,w)+b)
return 2*(y_hat-y)*logisticg(nd.dot(x,w)+b)

b = nd.arange(3, 4, 5) # Set the current value for b

w1 = nd.arange(2, 3, 2) # Set the current value for w
w1.attach_grad() # request for autograd w.r.t w
b.attach_grad() # request for autograd w.r.t b
y = 2. # set the label value
x = nd.arange(1, 2, 2) # set the input value
print(' w1=',w1,'\n b=',b,'\n x=',x)

w1=
[2.]
<NDArray 1 @cpu(0)>
b=
[3.]
<NDArray 1 @cpu(0)>
x=
[1.]
<NDArray 1 @cpu(0)>

with autograd.record():
M = nd.dot(x, w1)
x2 = M + b
y_hat = logistic(x2)
s = y_hat - y
l2 = s*s
326 Machine Learning with Python: Theory and Applications

head_gradient=nd.ones_like(l2)
l2.backward(head_gradient)
# print feed-forward results of the net at all local nodes
print('x=',x[0].asscalar,'w=',w1[0].asscalar,'M=',M[0].asscalar,
'x2=',x2.asscalar, 'y_hat=',y_hat,'s=',s,'l2=',l2.asscalar)
# print the gradients from the back-propagation
print('Autograd, w1.grad',w1.grad,' shape',np.shape(w1.grad))
print('Autograd, b.grad',b.grad)
#print the analytic results.
print('Analytic loss2=',loss2(x,w1,b,y))
print('Analytic lossdw=',lossdw(x,w1,b,y))
print('Analytic lossdb=',lossdb(x,w1,b,y))

x= <bound method NDArray.asscalar of

[1.]
<NDArray 1 @cpu(0)>> w= <bound method NDArray.asscalar of
[2.]
<NDArray 1 @cpu(0)>> M= <bound method NDArray.asscalar of
[2.]
<NDArray 1 @cpu(0)>> x2= <bound method NDArray.asscalar of
[5.]
<NDArray 1 @cpu(0)>> y hat=
[0.9933072]
<NDArray 1 @cpu(0)> s=
[-1.0066929]
<NDArray 1 @cpu(0)> l2= <bound method NDArray.asscalar of
[1.0134306]
<NDArray 1 @cpu(0)>>
Autograd, w1.grad
[-0.0133851]
<NDArray 1 @cpu(0)> shape (1,)
Autograd, b.grad
[-0.0133851]
<NDArray 1 @cpu(0)>
Analytic loss2=
[1.0134306]
<NDArray 1 @cpu(0)>
Analytic lossdw=
[-0.01338505]
<NDArray 1 @cpu(0)>
 Automatic Differentiation and Autograd 327

Analytic lossdb=
[-0.01338505]
<NDArray 1 @cpu(0)>

We see again that the autograd produces the exact solutions.

8.9 Examples: Differences Between Analytical, Autograd,

and Numerical Differentiation

We now check out the differences between analytical, autograd, and numer-
ical differentiation. We use the sigmoid function to do so, because it is one
of the most often used functions in machine learning.

import numpy as np

def logistic(z): # Define the sigmoid function

return 1. / (1. + nd.exp(-z))

x1 = nd.arange(-10, 10, 1) # Define x1

x1.attach_grad() # Request for gradient
with autograd.record(): # Define the path
g = logistic(x1)
h = g
head_gradient=nd.ones_like(g)
h.backward(head_gradient)
print('The gradient values of f at x=', x1.grad)

The gradient values of f at x=

[4.53958128e-05 1.23379359e-04 3.35237652e-04 9.10221133e-04
2.46650912e-03 6.64805667e-03 1.76627077e-02 4.51766588e-02
1.04993582e-01 1.96611941e-01 2.50000000e-01 1.96611971e-01
1.04993574e-01 4.51766588e-02 1.76627059e-02 6.64805621e-03
2.46650935e-03 9.10221308e-04 3.35237681e-04 1.23379359e-04]
<NDArray 20 @cpu(0)>

import mxnet as mx
from mxnet import nd, autograd, gluon
import matplotlib.pyplot as plt
from matplotlib import rc
%matplotlib inline
328 Machine Learning with Python: Theory and Applications

yg = logistic(x1)*(1.-logistic(x1)) # Analytic derivatives

nyg = nd.array(np.gradient(logistic(x1).asnumpy(),x1.asnumpy()))
# numerical gradient using np.gradient()
dgag=yg-x1.grad # difference between the analytical and autograd
dgng=yg-nyg # difference between the analytical and numerical
m = int(len(dgag)/10)
print('dgag=',dgag[0:len(dgag):m])
print('dgng=',dgng[0:len(dgng):m])
plt.plot(x1.asnumpy(),yg.asnumpy(),dashes=[30, 5, 5, 5],color='red')
plt.plot(x1.asnumpy(),x1.grad.asnumpy(),alpha=0.5,linewidth=7.0)
plt.plot(x1.asnumpy(),nyg.asnumpy(),alpha=0.9,linewidth=1.0,color='black')
plt.plot(x1.asnumpy(),dgag.asnumpy(), color='green')
plt.plot(x1.asnumpy(),dgng.asnumpy(), color='orange')
plt.xlabel('z')
plt.ylabel(r"$\sigma \ '(z) $")
plt.title('Gradient of the sigmoid Function')
plt.show()
l0 = int(len (x1)/2) #take look at its value at the center
print('x=',x1[l0], 'sigmoid(0)=', logistic(x1[l0]))

dgag=
[-3.6379788e-12 2.9103830e-11 0.0000000e+00 -1.8626451e-09
0.0000000e+00 0.0000000e+00 5.2154064e-08 2.7939677e-08
-4.3539330e-08 -1.4406396e-08]
<NDArray 10 @cpu(0)>
dgng=
[-3.2600899e-05 -5.8590609e-05 -4.2439066e-04 -2.7038064e-03
-5.7642013e-03 1.8941417e-02 -5.7641417e-03 -2.7037859e-03
-4.2444887e-04 -5.8584614e-05]
<NDArray 10 @cpu(0)>

Figure 8.4: Derivative of the sigmoid function computed using different methods, and the
error in the results.
 Automatic Differentiation and Autograd 329

x=
[0.]
<NDArray 1 @cpu(0)> sigmoid(0)=
[0.5]
<NDArray 1 @cpu(0)>

The results plotted in Fig. 8.4 show clearly that the autograd gives the
exact solution (the green line above is zero). The numerical gradient gives
only an approximated solution (the orange line is not zero).

8.10 Discussion

The numerical autograd is a powerful tool for computing the gradient of a

function and the result is exact (to machine accuracy). It can be used not
only in machine learning but also for many other computations in engineering
and sciences.
Our discussion focused on scalar functions with variables that can be a
scalar, or a vector. When the variable is a scalar, we have a derivative, and
when the variable is a vector, we have a gradient vector at a local node. In
an MLP layer, the function itself at the layer can be a vector. In these cases,
the gradient of the vector function can be formulated as a matrix, known
as the Jacobian matrix. The formulation, procedure, and the chain rule still
holds, but gradient vectors become Jacobian matrices.

References
[1] G.R. Liu and T.T. Nguyen, Smoothed Finite Element Methods. Taylor and Francis
Group, New York, 2010.
[2] G.R. Liu, Mesh Free Methods: Moving Beyond the Finite Element Method. Taylor and
Francis Group, New York, 2010.
MACHINE LEARNING
WITH PYTHON
 Chapter 9

Solution Existence Theory

and Optimization Techniques

9.1 Introduction

Developing and training neural network models is to find a set of parameters

(weights and biases) that minimize the loss function (or cost function or
objective function). The first question that needs to be answered is on the
existence of such a set of parameters for the model to be created given a
dataset. Our theoretical study in Chapter 5 has concluded that such a set of
parameters do exist, if affine transformation is used and the hypothesis space
is sufficiently large in relation to the dataset. The uniqueness of the solution
of the set of parameters needs further examination. This requires a good
understanding on the hypothesis used in the model as well as the dataset.
This chapter will first examine the solution existence and uniqueness using
an analytical manageable ML problem, known as linear regression problem.
Our analysis reveals important solution existence theory, predictability
of the solution, and effects of parallel data-points in a dataset, which
are applicable to general complex machine learning models.
Next, a proper tool known as optimization technique is required to find
these parameters for general ML problems that do not have a closed form.
Optimization techniques are not unique to machining learning algorithms
and are widely used in many design and analysis problems in engineering
and science. Developing more powerful new or using the existing techniques
properly to their fullest potential is believed to be one of the most challenging
works for a deep learning project. Techniques for optimization are essential to
machine learning for a number of reasons. For one, the optimization efficiency
largely controls the efficiency of training. Second, convergence of training is
often rooted in the convergence property of the optimization techniques used.

331
332 Machine Learning with Python: Theory and Applications

In addition, optimization methods often with some tunable parameters and

proper tuning of the optimization parameters can also be critical in training
a model.
Our discussion will focus on gradient-based techniques for finding an
optimal solution. For this part, we referenced excellent materials available
at https://github.com/zackchase/mxnet-the-straight-dope/tree/master/
chapter06 optimization (under the Apache License 2).

9.2 Analytic Optimization Methods: Ideal Cases

For optimization problems with very simple objective functions, one may
be able to find analytical solutions. For example, for an objective function
f (W ) = (W − a)2 with a given fixed real number a, one can immediately find
the global and unique minimum at W ∗ = a. This function is positive and
quadratic in W , and hence has one and only one solution in R1 . This simple
problem implies that finding the minimum is straightforward, if the function
is quadratic.

9.2.1 Least square formulation

However, very few machine learning problems have such a simple objective
function. There are problems that can be formulated with the least square
type of objective functions, such as the linear regression problems (https://
en.wikipedia.org/wiki/Linear regression). This type of problem is quite
common in sciences and engineering, yet simple enough to be formulated
analytically to obtain kind of close-form solution. It is an excellent type
of problem for us to reveal important theoretical properties on solution
existence and predictability. These properties can be generalized for complex
machine learning models that are too complex to analyze.
Given a dataset with m (measurement) data pair in the form of

[yi , xi1 , xi2 , . . . , xip ], i = 1, 2, . . . , m (9.1)

where yi ∈ Y1 is the label obtained at the ith label (observed or measured),

when p variables (features) are set at xi1 , xi2 , . . . , xip . To predict y based on
the given linearly independent variables x1 , x2 , . . . , xp ∈ Xp , we can construct
a simple linear regression model in p-dimensional space, by assuming that
the predictor ŷi is the affine transformation as discussed in Chapter 5:

ŷi ≡ zi = W 0 + W 2 xi1 + W 2 xi2 , . . . , + W p xip , i = 1, 2, . . . , m (9.2)

 Solution Existence Theory and Optimization Techniques 333

where W i , i = 0, 1, 2, . . . , p are the learning parameters in Wp+1 .

Equation (9.2) can be formulated in the following matrix form:
ŷ = Xŵ (9.3)

where ŷ is a vector of predictions:

ŷ = {ŷ1 , ŷ2 , . . . , ŷm } (9.4)

and X is an m × (p + 1) matrix, and its ith row is given as

p
[1, xi1 , xi2 , . . . , xip ] ∈ X (9.5)

and
ŵ = [W 0 , W 1 , W 2 , . . . , W p ] ∈ Wp+1 (9.6)

Note that Eq. (9.3) requires ŵ being a column vector. This allows us to
obtain a conventional form of the normal equation later. Our task is to find
a vector ŵ that gives a best ŷ closest to the labels in the given m data
pairs, y:

[y1 , y2 , . . . , ym ] (9.7)

For existence of a unique solution (see later), we require

m ≥ (p + 1) (9.8)

Meaning that the number of rows in X is at least no less than its

number of columns. In other words, we require more (usually a lot more)
observation/measurement data, and hence it results in a “tall” matrix X.
Equation (9.8) also gives the basic relationship between the number of
samples m and the number of training parameters (p + 1). More detailed
discussion will be given in Section 13.2 for deepnets.

9.2.2 L2 loss function

We define now a loss function using the L2 norm that is the sum of the
squared differences between the predictor ŷ and the label data y for all the
observations. It is often called L2 loss function.

L(y, ŷ(ŵ)) = ||y − Xŵ||22 = (y − Xŵ) (y − Xŵ) (9.9)

Notice Eq. (9.9) mimics the idea function discussed at the beginning of this
section, and is essentially quadratic. Clearly, the loss function L is zero, if the
prediction produces the dataset exactly. Otherwise it is a positive number.
334 Machine Learning with Python: Theory and Applications

The regression problem is casted as the following minimization problem using

the least-square formulation:

min L(y, ŷ(ŵ)) = min (y − Xŵ)T (y − Xŵ) (9.10)

ŵ∈Wp+1 ŵ∈Wp+1

Because L is a function of ŵ, and is also quadratic in ŵ. It is thus hopeful

that we can expect a unique set of optimal ŵ∗ ∈ Wp+1 that minimize L,
which can be found analytically by evaluating its gradient (see Chapter 8),
and then setting to 0:

∂L(y, ŷ(ŵ)) T
∇ŵ L(y, ŷ(ŵ)) = = −2X (y − Xŵ∗ ) = 0 (9.11)
∂ ŵ

9.2.3 Normal equation

Equation (9.11) gives
 
X Xŵ∗ = X y (9.12)

This equation is known as Normal Equation. Assuming matrix X X is
invertible (X has full rank, as discussed in detail in Chapter 3), the solution
can be found as
 
ŵ∗ = (X X)−1 X y (9.13)

which is a kind of close-form solution.

Note that for large systems, we do not usually form the normal matrix
and solve the normal equation. Instead, the QR transformation type of
algorithms are used to compute numerically the least-square solution. This is
because of the numerical instability reasons mentioned in Section 3.4.2. For
our theoretical analysis here, this approach works well, because no numerical
operations involved. It can reveal important inside on solution existence.

9.2.4 Solution existence analysis


In obtaining the above solution ŵ∗ , X needs to have full rank so that X X
is invertible. Let us examine what this really means, by considering a simple
case of 2D feature space: p = 2, m = p + 1. If X has full rank, the normal
equation (9.12) can be reduced to

Xŵ∗ = y (9.14)
 Solution Existence Theory and Optimization Techniques 335

Figure 9.1: Vectors of data-points in the feature space X2 (blue) and in the augmented
2
feature space X (red). Vectors x1 , x2 , and x3 in space X2 are linearly dependent, but x1 ,
2
x2 , and x3 on space X in R3 are not, because of the elevation.

It can be written explicitly as

⎡ ⎤⎡ ⎡ ⎤⎤
1 x11 x12 b∗ y1
⎢ ⎥ ⎢ ∗⎥ ⎢ ⎥
⎣1 x21 x22 ⎦ ⎣w1 ⎦ = ⎣y2 ⎦ (9.15)
1 x31 x32 w2∗ y3

It is clear now that X is exactly the same as the “moment matrix” in the for-
mulation of the triangular elements given in Eq. (7.10) in FEM [1]. The con-
dition for X having full rank is that the area formed by these 3 data-points
in the feature space X2 (2-dimensional) is nonzero. Such an area is half of the
determinate of the moment matrix. Figure 9.1 plots 3 data-points x1 , x2 , and
x3 in space X2 , the area for which the triangle is clearly nonzero. If we replace
x3 by x3 , x1 , x2 , and x3 become in line, the triangle becomes degenerated,
the area becomes zero, X will be singular, and its rank reduces to 2.
Notice that while x1 , x2 , and x3 are linearly dependent in space X2 ,
vector x3 can be presented by a linear combination of x1 and x2 , because
these 3 vectors (in blue) are in plane. However, when these 3 data-points are
2
projected onto the affine plane X (see Section 1.4.2), these 3 vectors (in red)
in R3 are no longer in plane and become linearly independent. Therefore,
alternatively the condition for X having a full rank is that the vectors of
2
these 3 data-points on plane of x0 = 1 on the space X in R3 are linearly
independent. If we replace x3 by x3 , x1 , x2 , and x3 are now in a plane, and
they become linearly dependent. In this case, X will become singular and
336 Machine Learning with Python: Theory and Applications

its rank will become 2. Therefore, the rank of X depends on the maximum
2
number of the linearly independent vectors of the data-points in the X
space.
Consider now cases of m > p + 1 with p = 2. This means that we still
have 2 features, but more than 3 data-points. In these cases, the rank of X
depends on the maximum number of the linearly independent vectors in R3
2
of all the data-points in the X space. Hence, the chance should be higher.
For general high-dimensional feature spaces Xp , the same argument shall
hold for X being full rank. If we examine the data-points in Xp space, we need
the “volume” of the largest polyhedron with any set of p + 1 data-points as
p
its vertexes to be non-zero. If we examine the data-points in X space, we
need the minimum number of the linearly independent vectors in Rp+1 of all
p
the data-points in the X to be p + 1. Once X has full rank, we shall have
a least-square solution given by Eq. (9.13). A general analysis for arbitrary
dimension can be found in [2].
We now summarize our discussion in the following solution existence
theory.

9.2.5 Solution existence theory

Solution Existence Theory: Consider a dataset with m ≥ p + 1 data-
points. If the maximum number of the linearly independent vectors among
p
all the data-points in the dataset in the affine space X is no-less than p + 1,
then matrix X will have full rank, and there exists a unique least-square
solution given in by Eq. (9.13).
Alternatively, the condition may also be set for non-degeneration of the
largest polyhedron formed by any set of p + 1 data-points in the dataset as
its vertexes in the feature pace Xp . In high dimensions, the geometry of such
a polyhedron becomes very complicated.
In practice, when we have many data-points, X is usually a very tall

matrix, and hence matrix X X is often invertible and SPD.
Remark: The above analysis reveals also an important fact that the existence
of unique solution depends solely on the number of data-points and the
distribution of the data-points in Xp .
When X has full rank, we can obtain the exact optimal value ŵ∗ via
single-step computation using Eq. (9.13), with one matrix multiplication,
one matrix inversion, and two matrix-vector products.
If the dataset contains higher-order latent features, one can add
in higher-order bases, such as x1 , x2 , x21 , . . .. The solution existence theory
 Solution Existence Theory and Optimization Techniques 337

still holds, and the solution existence depends on whether the resulting
moment matrix X has full rank. Alternatively, one can use a more powerful
hypothesis, such as MLP a large number of learning parameters discussed in
Section 13.2. Precise analysis of solution existence for MLP becomes more
complicated, but we can assert that it depends on whether or not the number
and the distribution of the data-points support the MLP hypothesis with
the learning parameters. This assertion is important in ML model creation,
because if one fails to train a model, checking and enriching the dataset, or
reduce the complexity of MLP could be effective solutions.

9.2.6 Effects of parallel data-points

Let us dig in further into another important issue: the effects of duplicated
or parallel data-points in datasets, which may bother many in building
machine learning models. Assume in Eq. (9.15), that any two data-points
are identical or parallel. The moment matrix will be singular, based on the
Solution Existence Theory. If there are many data-points in the dataset, in
which there still be at least three data-points linearly independent in its
2
corresponding affine space X , we shall still have a least-square solution. In
this case, the duplication (or parallelism) of the data-points has no effect
on the solution existence. However, it may affect the value of the solution,
because the same data-points are used twice in L2 loss function. If this is not
the intention of the analyst, clearing up the dataset, using PCA for example,
is needed. This finding applies to higher dimensions and to other complicated
models.
In terms of obtaining a solution in a single step (without iteration) in
the form of Eq. (9.13), this least square linear regression probably is as far
as we can go. At most, we may add in some constraints for this type of
optimization problem leading to quadratic programming problems, as we
have done in studying the SVM.

9.2.7 Predictability of the solution against the label

Note that yi ∈ Y does not affect the solution existence. It affects the
values of the solution ŵ∗ . In constructing the loss function in Eq. (9.9),
we compare ŷi with yi . This implies that the prediction ŷi must be in the
label space Y. This relates to our discussion in Chapter 5 on predictability
of the hypothesis using affine transformation. We have proved there that
the prediction function given in Eq. (9.2) can predict any arbitrary linear
function of feature variables x that may be hidden in the dataset. In this
338 Machine Learning with Python: Theory and Applications

particular case, such a linear function gives the least error defined in Eq. (9.9)
in fitting the dataset.
To predict higher-order and nonlinear latent features relating to the labels
in a dataset, one has to either add in higher-order bases, or other types of
nonlinear bases, or more powerful nonlinear models like MLP, as mentioned
earlier. The predictions at the output layer of neurons will still be, in Yk .
This is because of the continuous coverage over the affine spaces by each
neuron in the hidden layers ensured by both affine transformations wrapped
with activation functions. The prediction functions at the output layer are
continuous and differentiable with respect to the learning parameters, as
discussed in detail in Chapter 5 for affine transformations and in Chapter 7
for activation functions. In addition, the gradients can be found via autograd
as demonstrated in Chapter 8. Thus, a smooth and continuous coverage
of the prediction function can be realized by the learning parameters in
hypothesis space WP . As results, the predictability of the solution against
the label is ensured in theory. Regardless of the actual values of the labels
in the label space, the prediction shall be able to match those values to
the desired accuracy, by tuning the learning parameters in the hypothesis
space. The Universal Prediction Theory (see, Chapter 5) has ensured the
capability of an MLP, as long as proper configuration of the MLP is used
with a sufficiently large number of learning parameters. Effective numerical
techniques are, however, needed to find the optimal learning parameters that
minimize the loss function.
For most of the machine training models, the loss function is complicated.
It is often not possible to convert the minimization problem similar to
Eq. (9.10) to a single normal equation system. Therefore, we have to find
a way to solve the minimization problems directly by numerical means.
Iterative methods are often used, and finding a set of the minimizers requires
many steps.

9.3 Considerations in Optimization for Complex Problems

Careful considerations are needed to solve the minimization problems

numerically. For example, the loss functions may not be convex globally,
the optimum point is not unique, multiple local optima, and presence of
saddle points.
In the rest of the discussion in this chapter, we consider general opti-
mization techniques that can be used in many areas of study in science and
 Solution Existence Theory and Optimization Techniques 339

engineering, including machine learning. Therefore, without losing generality,

the function to be optimized is denoted in general as f (x) with x being the
independent variables. If there is only one variable, x simply becomes x. In
machine learning, however, our function would be the loss function L(ŵ)
and the independent variables would be the learning parameters ŵ. Care is
needed to avoid confusion.

9.3.1 Local minima

For most real-life machine learning, the objective function f (x) with x being
the variable may have a number of local minimum values, at each of which
f (x) is smaller than at its nearby points. Only the x∗ , at which f (x∗ ) has the
smallest value over the entire domain, is regarded as the global minimizer,
and is usually the one we would like to find.
Consider the following function:

x
f (x) = · cos(πx), −1.0 ≤ x ≤ 4.0. (9.16)
2

Let us plot the curve of this function for an explicit view.

%config InlineBackend.figure_format ='retina' # high quality plot

%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt

def f(x):
return 0.5*x * np.cos(np.pi * x)

x = np.arange(-1.0, 4.0, 0.1)

fig = plt.figure()
subplt = fig.add_subplot(111)
subplt.annotate('local minimum',xy=(-.3,-.1),xytext=(-.8,-.5),
arrowprops=dict(facecolor='black', shrink=0.01))
subplt.annotate('local minimum',xy=(1.1,-.5),xytext=(.5, .05),
arrowprops=dict(facecolor='black', shrink=0.05))
subplt.annotate('global minimum',xy=(3.,-1.5),xytext=(3.,.05),
arrowprops=dict(facecolor='red', shrink=0.05))
plt.title('A cost function with multiple minima')
plt.plot(x, f(x))
plt.show()
340 Machine Learning with Python: Theory and Applications

Figure 9.2: An example of cost functions with multiple minimums.

It is clear Fig. 9.2 that this function has two local minimums, and one
global minimum in the domain we plotted. Existence of local minima presents
a challenge, and unfortunately we do not have effective means to ensure the
finding of the global minimum for general ML models. In practice, what we
can do is to perform the optimization multiple times, each starting from a
different initial point, and then take the minimum of these multiple minima.

9.3.2 Saddle points

Saddle points present another challenge for optimization techniques. The
gradient of the loss function at the saddle points is zero, but the loss function
itself may be minimum there along some directions and maximum along
other directions. For high-dimensional problems (like machine learning),
saddle points can be in numbers. The problem with saddle points can be
evidenced using the following example.
Consider an objective function in one dimension defined as
1 3
f (x) = x + 2, −2.0 ≤ x ≤ 2.0 (9.17)
2
The derivative of this function is
3
f  (x) = x2 , −2.0 ≤ x ≤ 2.0 (9.18)
2
which is zero at x = 0, but all positive elsewhere in −2.0 ≤ x ≤ 2.0. The
2nd derivative of this function is
f  (x) = 3x − 2.0 ≤ x ≤ 2.0 (9.19)
which is zero at x = 0. However, it is negative in [ − 2.0, 0.0), and positive
in (0.02.0]. This means that x = 0 seems to be a maximizer when viewed
 Solution Existence Theory and Optimization Techniques 341

from left towards 0 and seems to be a minimizer when viewed away from 0
to its right, as shown in Fig. 9.3. The gradients at both opposite sides of the
saddle point are the same.

x = np.arange(-2.0, 2.0, 0.1)

fig = plt.figure()
subplt = fig.add_subplot(111)
subplt.annotate('saddle point',xy=(0,1.8),xytext=(-.4,-1.),
arrowprops=dict(facecolor='red', shrink=0.001))
plt.plot(x, 0.5 * x**3 + 2.0)
plt.show()
plt.annotate('saddle point', xy=(0., 0.1), xytext=(0., 2.),
arrowprops=dict(facecolor='red', shrink=0.001))
plt.plot(x, 1.5 * x**2 )
plt.show()

Figure 9.3: An example of cost functions with saddle points.

342 Machine Learning with Python: Theory and Applications

Figure 9.3 shows an example of cost functions with saddle points. We

have seen saddle functions in 2D space earlier when studying the SVM. It was
handled nicely via analytical means, which leads to a quadratic programming
problem. In machine learning, many complicated functions can have multiple
similar but more complicated saddle points. It is not likely we can find an
analytical means for a set of minimizers. Let us write a code to generate
such a function to get a feel for it.

import matplotlib as mpl

#mpl.rcParams['figure.dpi']= 80
import matplotlib.pyplot as plt
from mpl_toolkits import mplot3d
from mpl_toolkits.mplot3d import axes3d
from IPython.core.interactiveshell import InteractiveShell
#InteractiveShell.ast_node_interactivity = "all"
#%matplotlib notebook #%matplotlib tk
x1 = np.linspace(.0 ,3.5,100) # np.arange(0.5, 5.0, 5.0)
x2 = np.linspace(-1.,4.0,100) # np.arange(2.0, 6.0, 5.0)
lx1, lx2 = len(x1), len(x2)
print('length of x1=',lx1,'length of x2=',lx2)
X1, X2 = np.meshgrid(x1, x2)
a, b = 1.0, 0.5
Z = a*X1 * np.cos(np.pi * X1) + b*X2 * np.cos(np.pi * X2)
fig_s = plt.figure(figsize=(8,5))
ax = fig_s.add_subplot(1,1,1,projection='3d')
ax.set_xlabel('$x_1$', fontsize=9, rotation=80)
ax.set_ylabel('$x_2$', fontsize=9)
ax.set_zlabel('Function Value', fontsize=9, rotation=0)
ax.yaxis._axinfo['label']['space_factor'] = 2.0
#ax.plot_wireframe(X1,X2,Z,color='b',rstride=5,cstride=5)
ax.plot_surface(X1,X2,Z,color='b',rstride=1,cstride=1,
shade=False,cmap="jet", linewidth=1)
plt.title('Parabolic Cylinder function of $x_1$ and $x_2$'\
' with, a='+str(a)+', b=' +str(b))
#plt.legend()
fig_s.tight_layout() # otherwise right y-label is clipped
plt.show()

length of x1= 100 length of x2= 100

 Solution Existence Theory and Optimization Techniques 343

Figure 9.4: An example of cost functions in 2D space with multiple saddle points.

For this 2D case plotted in Fig. 9.4, there are multiple saddle points, at
which the gradient of the function is zero, but the values of the function are
maxima along the x1 direction and minima along the x2 direction.

9.3.3 Convex functions

Elliptic paraboloidal functions are easy to deal with, because such a function
is convex and has only one minimum. Let us write a code and plot one.
a, b = 1.0, 2.0 # minor & major radii for an ellipse
x1 = np.linspace(-2.0,2.0,100)
x2 = np.linspace(-4.,4.0,100)
X1, X2 = np.meshgrid(x1, x2) # Grides for 3D plot
Z = X1*X1/a**2 + X2*X2/b**2 # an ellipse
#%matplotlib notebook
%matplotlib inline
fig_s = plt.figure(figsize=(8,6))
ax = fig_s.add_subplot(1,1,1,projection='3d')
ax.set_xlabel('$x_1$', fontsize=12, rotation=80)
ax.set_ylabel('$x_2$', fontsize=12)
ax.set_zlabel('Function Value',fontsize=12,rotation=0)
ax.yaxis._axinfo['label']['space_factor'] = 3.0
ax.plot_wireframe(X1,X2,Z,color='b',rstride=5,cstride=5)
344 Machine Learning with Python: Theory and Applications

plt.title('Elliptic Paraboloidal function of $x_1$ and '\

'$x_2$ with, a='+str(a)+', b=' +str(b))
#plt.legend()
fig_s.tight_layout() # otherwise y-label is clipped
plt.show();

Figure 9.5: An ideal example of quadratic cost functions in 2D space with only one
minimum.

As shown in Fig. 9.5, the ellipse function behaves very well and has only
one minimum. Understanding this type of property is useful in dealing with
complicated loss functions because near a local minimum, many complicated
functions are locally elliptic, and hence can be taken advantage of. The only
difficulty that may arise is when the ratio of the minor and major radii is
too large (or too small). We shall discuss this again later.

9.4 Gradient Descent (GD) Method for Optimization

The gradient descent method is the basic technique for minimization of a

function when its gradient is computable. It is widely used in science and
engineering, including machine learning. It is in general an iterative method,
meaning the solution (approximate) is obtained through a number of guided
 Solution Existence Theory and Optimization Techniques 345

iterations. This type of iterative method is most practical especially for

training machine learning models, because the parameter space is usually
of extremely high dimension.
Consider a 2D space, where a cost function surface is defined. The surface
of the cost function can be viewed as a terrain surface with many valleys and
hills. Imagine that one is currently standing at a point on a hill and would like
to get to the lowest point in the valley. The fastest path is going downhill in
the direction where the slope is steepest. This is essentially the gradient
descent method (sometimes called the steepest descent method). The
gradient is just a formal and more rigorous term in mathematics for the slope.
A typical cost function defined in machine learning problems can be in
much higher dimensions, possibly in the order of as much as millions. It is
a space far beyond human imagination. However, the surface can still be
viewed as a hyper-dimensional surface, and the direction should still be the
steepest downhill direction (or negative gradient direction). To develop such
a computer algorithm, we shall

• Compute the gradient that gives the learning direction.

• Determine the step to take in that direction, which gives the learning rate.

Let us examine how this would work for the simplest one-dimensional
problem.

9.4.1 Gradient descent in one dimension

Assume a cost function f (x) defined in a one-dimensional (1D) space with
variable x. Assume also that it has at least the first derivative (that is the
gradient in the 1D case) at any x in the domain of interest. We can then use
the Taylor series expansion in the vicinity of x, which gives

f (x + δx) ≈ f (x) + f  (x)δx (9.20)

where (δx) is a small real number. Because the function has the first
derivative, it must be a finite number (otherwise, it is not differentiable).
Therefore, we can choose a very small positive real number −η as a scalar,
so that −ηf  (x) is as small as δx. This implies that η should carry a unit of
[(δx)f −1 (x)]. Equation (9.20) can now be rewritten as

f (x − ηf  (x)) ≈ f (x) − η(f  (x))2 . (9.21)

Because η is chosen positive, we have

f (x − ηf  (x)) ≤ f (x). (9.22)

346 Machine Learning with Python: Theory and Applications

Figure 9.6: Gradient descent in a one-dimensional case.

This means that an update to x with

x := x − ηf  (x) (9.23)

would lead to a reduction of the value of f (x). Note the negative sign, which
implies that updating in x is in the negative (or downhill) direction of the
gradient.
Figure 9.6 shows that regardless of which side the x is located on, the
gradient descent takes a step closer to the minimum of the function that is
convex.

9.4.2 Remarks
Based on the above analysis, we note the following important points:
1. Updating in x is in the negative direction of the gradient, and hence the
name of gradient descent.
2. The amount of reduction is proportional to the gradient of the function.
The larger the gradient, the larger the reduction, provided that ηf  (x) is
sufficiently small so that the Taylor expansion (and our analysis) can still
hold without overshooting.
3. Scalar η is called the learning rate or learning step size in machine
learning, and choices of it can be tricky. It often requires some trial and
error.
 Solution Existence Theory and Optimization Techniques 347

4. On the one hand, we must take the direction at which the derivative is the
largest. On the other hand, we must not be too greedy to take too big an η
that could undermine the basis of our analysis. A larger learning rate may
lead to overshooting and hence oscillating behavior in the convergence
process, which is often observed in practicing machine learning.
5. When the derivative value f  (x) at x is zero, we make no progress. This
means also that we are already at the minimum (including the saddle
point discussed in the previous section).
6. If the derivative value f  (x) at x is small, we make very (one more order
smaller) little progress (note the (f  (x))2 term). This implies that to speed
up the process, one would need to access the higher-order derivatives
of the cost function (such as the higher-order methods like Newton’s
method).

9.4.3 Gradient descent in hyper-dimensions

Consider a cost function f (x), where argument x = [x1 , x2 , . . . , xp ] has p

variables (or parameters in ML models). This means that f (x), a scalar, is
defined in a space with a dimension of p.
The gradient of f (x) is thus with respect to all these variables/parameters
x, which is a vector (see Chapter 8) written in the form of
 
∂f (x) ∂f (x) ∂f (x)
∇x f (x) = , ,..., . (9.24)
∂x1 ∂x2 ∂xp

where ∂f (x)/∂xi is the partial derivative of the function with respect to the
ith parameter. It is the change in f at x with respect only to a small change
in xi .
To measure the change in f at x with respect to a small of change in
length (δl) in an arbitrary direction defined by a unit arbitrary vector u, we
may write
f (x + (δl)u) − f (x)
Du f (x) = lim (9.25)
(δl)→0 (δl)

which is in fact the derivative of f (x) in an arbitrary direction u. Notice

that Du f (x) is a scalar (for a scalar f (x)).
On the other hand, using the Taylor series expansion in the vicinity of x
gives
f (x + (δl)u) ≈ f (x) + ∇x f (x) · (δl)u (9.26)
348 Machine Learning with Python: Theory and Applications

which can be rewritten as

f (x + (δl)u) − f (x)
∇x f (x) · u = lim (9.27)
(δl)→0 (δl)

Using Eqs. (9.25) and (9.27), we obtain

Du f (x) = ∇x f (x) · u. (9.28)

This is expected, because the derivative of the function (scalar) in an

arbitrary direction u is the projection (inner product) of the gradient of
the function (vector) on the direction of u.
For our minimization purposes, we want the change in cost function
Du f (x) as negative as possible. We thus minimize Du f (x) with respect to
all possible directions of the unit vector u. Equation (9.28) is then written as

Du f (x) = ∇x f (x) · u · cos(θ) = ∇x f (x) · cos(θ) (9.29)

where θ is the angle between ∇x f (x) and u, with 0 ≤ θ ≤ 2π. The minimum
value of cos(θ) is −1 at θ = π. Therefore, Du f (x) is minimized when u is at
the opposite direction of the gradient ∇x f (x).
We can now reduce the value of f by updating x in the following manner:

xi := xi−1 − η∇x f (x) (9.30)

where η is a positive scalar that shall carry a unit of [(δx)∇x f (xi )−1 ], the
learning rate.
The analysis for hyper-dimensional problems is similar to what we have
done for the one-dimensional problems, but needs an additional analysis on
the direction. We also note that the gradient of a scalar function with respect
to a vector variable is a vector of the same shape of the variable vector. In
other words, taking the gradient to a scalar function is essentially the same
as taking the derivative. All one needs to do is to change the scalar variable
to a vector variable. By the same argument, all the remarks made in the
previous subsection also hold for hyper-dimensional problems.

9.4.4 Property of a convex function

Convexity is an important feature of a function in terms of optimization, and
many optimization algorithms work well for convex functions. If a function
f (x) is said to be convex, it means that (a) there is an x∗ at which f (x∗ ) is
a minimum and (b) for any xa and xb in the domain where f (x) is defined,
we always have
 Solution Existence Theory and Optimization Techniques 349

Figure 9.7: Pictorial demonstration of the properties and important inequalities of a

convex function in one dimension.


xa + xb f (xa ) + f (xb )
f ≤ (9.31)
2 2
This condition can be given in a more general statement as follows:

f (xb ) − f (xa ) ≤ ∇xb f (xb ) · (xb − xa ) (9.32)

These properties are useful to prove convergence theory for gradient descent
methods. They are shown pictorially in Fig. 9.7.

9.4.5 The convergence theorem for the Gradient

Decent algorithm
Theorem: Consider a Gradient Descent (GD) algorithm for minimizing
function f (x). At time t + 1, it updates x using the following GD algorithm:

xt+1 = xt − η [∇xt f (xt )] = xt − η∇t (9.33)

  
∇t

where we denote for short

∇t = ∇xt f (xt ) (9.34)

If (1) function f (x) is convex;

(2) the norm of the gradient has an upper bound for all t:

∇t  = ∇xt f (xt ) ≤ G (9.35)

where G is a constant independent of t;

350 Machine Learning with Python: Theory and Applications

(3) there is no overshooting in the updating process, meaning that the

(Euclidean) distance between any point xa in the domain and the minimum
point x∗ is bounded; We shall have

xa − x∗  ≤ R (9.36)

where R is a constant independent of t.

After T iterations, we obtain
 T

1 RG
f xt ≤ f (x∗ ) + √ (9.37)
T t T
R
by setting η = √
G T
that has a unit of [f −1 R].

Proof. We first examine the length (amplitude) of vector xt+1 − x∗ 2 .

xt+1 − x∗ 2 = (x∗ − xt+1 ) · (x∗ − xt+1 )

= (x∗ − xt + η∇t ) · (x∗ − xt + η∇t )
= x∗ − xt 2 − 2η∇t · (xt − x∗ ) + η 2 ∇t 2
≤ x∗ − xt 2 − 2η∇t · (xt − x∗ ) + η 2 G2 (9.38)

The above equation gives

1 η
∇t · (xt − x∗ ) ≤ (x∗ − xt 2 − xt+1 − x∗ 2 ) + G2 (9.39)
2η 2
We now examine
 T
 T
1 ∗ 1
f xt − f (x ) ≤ (f (xt ) − f (x∗ ))
T t T t
T
1
≤ (∇t · (xt − x∗ ))
T t
T
1  ∗ η
≤ [x − xt 2 − xt+1 − x∗ 2 ] + G2
2ηT t 2
1 η
= [x∗ − x0 2 − x∗ − xT 2 ] + G2
2ηT 2
R2 η RG
≤ + G2 = √ (9.40)
2ηT 2 T
R
In the last step, we have set η = √
G T
. This completes the proof. 
 Solution Existence Theory and Optimization Techniques 351

9.4.6 Setting or the learning rates

In the proof given above, we simply assumed that the gradient is bounded
by a constant G < ∞. The setting of η implies that the learning rate shall
depend on step t. More detailed analysis can be performed by linking G
with the so-called Lipschitz continuity constant for convex functions. Such an
analysis would lead to the following conditions for general convex functions.
∞
 ∞

ηt2 < ∞ and ηt → ∞ (9.41)
n n

The first equation ensures a convergence, meaning that f (xt ) will eventually
approaches to f (x∗ ), and the 2nd equation ensures the convergence can be
achieved in finite steps. Note that in practice, the tuning rate can often be a
challenging task. Equation (9.41) a theoretical guideline to design a schedule
for learning rates. The schedule give below satisfies both foregoing equations.
η0
ηt = β (9.42)
t
in which η0 is a pre-specified initial learning rate, and 1/2 < β < 1, in order
to satisfy conditions in Eq. (9.41). The following is a simple code for this
(inverse) power scheduling.

%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np

def lr_schedulerP(epoch, lr0, beta):

'''epoch: epoch number in the learning process
lr0: starting linearing rate
beta: (inverse) power '''
return lr0/(epoch+1)**beta

ep_list = np.array(range(2000))
lr0, beta, lr_list= 0.01, 0.51, []
for epoch in ep_list:
lr_list.append(lr_schedulerP(epoch, lr0, beta))

fig = plt.figure()
ax = plt.axes()
ax.plot(ep_list, lr_list);
352 Machine Learning with Python: Theory and Applications

Figure 9.8: An exponentially decaying learning rate.

Figure 9.8 plots an example of an exponentially decaying learning rate.

The author often found a simple linear schedule can work well for many
problems. Readers may try the following codes with modification for his/her
problems.

def lr_scheduler1(epoch, lr, lr1, lrT, ep_linear):

'''Linear schedule of learning rates
epoch: epoch number in the learning process
lr: current learning rate
lr1: starting point of the linear schedule
lrT: learning rate at the final learning stage
ep_linear: number of epochs for the linear schedule\
portion
lr1_T: learning rages in the linear portion from lr1\
to lrT'''

lr1_T = (lrT-lr1)/ep_linear * epoch + lr1

lr = [lrT if epoch>ep_linear else lr if lr1_T>lr\
else lr1_T][0]
return lr

ep_list = np.array(range(1500))
lr, lr_list= 0.01, []
for epoch in ep_list:
lr_list.append(lr_scheduler1(epoch,lr,0.02,1e-5,1000))
 Solution Existence Theory and Optimization Techniques 353

fig = plt.figure()
ax = plt.axes()
ax.plot(ep_list, lr_list)

Figure 9.9: An example of piecewise linearly decaying learning rate.

This schedule shown in Fig. 9.9 has constant portion at the beginning,
which can be useful to have a fast convergence at the initial stage. The
learning rate used at the final stage is set at a minimum rate that is
acceptable to the analyst. The learning rate in between varies linearly.

9.5 Stochastic Gradient Descent

Computation of the gradient in a gradient descent algorithm is the major

consumption of the computer power. Especially for hyper-dimensional
problems with extremely large number of variables with a huge dataset, the
gradient descent methods can be too costly. The stochastic gradient descent
(SGD) method offers an effective way to reduce computations. It makes use
of the statistics property when the training datasets are sufficiently large.
The key idea in SGD is as follows.
When training a model, we use a dataset with many data-points. The
cost function is in fact a sum of a number of cost functions for each data-
point. The gradient of the whole cost function is then estimated using the
gradient of the cost function for the ith data-point, with i being randomly
sampled at uniform from the dataset. Such a sampling gives, of course, less
accurate gradient, but it is an unbiased estimation of the true gradient of the
354 Machine Learning with Python: Theory and Applications

whole cost function. For a dataset with one million data-points, for example,
the computation of the gradient at each iteration is reduced by one million
times! Thus, this can be very cost-effective.
When the SGD is used in this kind of extreme manner, the convergence
of the cost function may become less stable and behaves in an oscillatory
manner. This is when the mini-batch concept comes in handy; instead of
randomly sampling only one data-point, we can sample randomly at uniform
a small batch of data-points at each iteration (epoch). This can significantly
improve the convergence behavior with a little increase of cost, and hence
the mini-batch technique can be very effective. It is one of the most efficient
algorithms in machine learning when dealing with a big dataset.
The SGD is even more appealing when the training data-points have
high redundancy. The stochastically sampled gradients can be a very good
estimation of the true gradient of the whole cost function using a small batch.
When the training data-points are large enough with high redundancy, a
smaller batch size can be used, and the saving in cost can be very significant.
In addition, using the stochastically sampled gradients can be considered as
a type of regularization, providing some effect in alleviating overfitting.

9.5.1 Numerical experiment

We now demonstrate gradient-based algorithms, by creating a simple neural

network model for linear regression, using the code made available at mxnet-
the-straight-dope (https://github.com/zackchase/mxnet-the-straight-dope/
blob/master/chapter01 crashcourse/autograd.ipynb), under the Apache-2.0
License.
We first define a function for minimization based on the SGD method
with a mini-batch algorithm. The key steps in a basic mini-batch algorithm
are as follows:

1. Randomly shuffle all the data-points or samples in the entire dataset.

2. Select batches each with a small number of samples (known as mini-batch)
from the shuffled dataset.
3. Loop over the mini-batches, and compute the gradient of the loss function
with respect to all its training parameters and perform minimization
(update the training parameters) until the process is roughly converged.
This completes one training epoch.
4. Select another set of mini-batches randomly from the entire training
dataset, and repeat the above for another epoch of training. This is done
 Solution Existence Theory and Optimization Techniques 355

for multiple epochs, until all the samples in the entire dataset are more
or less exhausted and the loss function is converged to an acceptable low
level.

Because of the random sample selection for the mini-batches, one can
hope that any mini-batch is statistically a good representation of the entire
dataset. The following function is defined for the mini-batch SGD:

def sgd(params, lr, batch_size): # Mini-batch SGD.

for param in params: # lr: learning rate
param[:] = param-lr*param.grad/batch_size

import mxnet as mx
import numpy as np
from mxnet import autograd,gluon
from mxnet import ndarray as nd
import random
%matplotlib inline
import matplotlib.pyplot as plt

mx.random.seed(1) # seed value for random number generation

random.seed(1)
# Generate data for training and testing.
n_v = 1 # n_v = 2 # number of variables/features
# or dimension of the space for regression.
n_s = 1000 # number of samples (data pints in the dataset)
true_w = [0.8] # one may try this: true_w = [2, -3.4]
true_b = 4.2
start_w,start_b = 1.0,3.5 # initialize w and b
X = nd.random_normal(scale=1, shape=(n_s, n_v))
#generate coordinates in variable space
y = true_w[0] * X[:, 0] + true_b
# generate the function values
# using the coordinates
#y = true_w[0] * X[:,0]+true_w[1]*X[: 1]+true_b
y += .2 * nd.random_normal(scale=1, shape=y.shape)# add noise
dataset = gluon.data.ArrayDataset(X, y) #generate the dataset
356 Machine Learning with Python: Theory and Applications

plt.scatter(X[:, 0].asnumpy(),y.asnumpy(),color='r')
plt.xlabel('$x$')
plt.ylabel("$\hat{y}$")
plt.title('data-points for Linear Regression')
plt.show()

Figure 9.10: Computer-generated data-points with noises for regression study.

# Data iterator for randomly generating mini-batches of data-points

def data_iter(batch_size):
idx = list(range(n_s)) # n_s: number of samples
random.shuffle(idx)
for batch_i,i in enumerate(range(0,n_s,batch_size)):
j = nd.array(idx[i: min(i + batch_size, n_s)])
yield batch_i, X.take(j), y.take(j) # take a mini-batch

# Initialize model training parameters.

def init_params():
w = nd.random_normal(scale=1, shape=(n_v, 1))
b = nd.zeros(shape=(1,))
w[0,0] = start_w # for examination, fix starting point.
b[0] = start_b # print('in init;',w[0,0],b[0])
params = [w, b]
for param in params:
param.attach_grad()
return params

# Linear regression net: an affine mapping.

def net(X, w, b):
return nd.dot(X, w) + b
 Solution Existence Theory and Optimization Techniques 357

# Least squared loss function.

def square_loss(yhat, y):
return (yhat - y.reshape(yhat.shape)) ** 2 / 2

def cost_function(X,W11,B1,y): # for plotting cost function

W12 = np.expand_dims(W11, axis=2)
# W11:(100,100)->(100,100,1) for X:(1000,1)
print('shape of W12:',W12.shape,'shape of X:',X.shape)
yh = np.inner(X.asnumpy(), W12)+ B1
yy = np.expand_dims(y.asnumpy(), axis=-1)
yz = np.expand_dims(yy, axis=-1)
print('shape of yh:',yh.shape,'shape of yz is:',yz.shape)
yhy = np.mean((yh - yz) ** 2/2,axis=0)
return yhy

With all the necessary functions defined for use, we now write the
algorithm to train the model.

%matplotlib notebook
import matplotlib as mpl
mpl.rcParams['figure.dpi']= 120
import matplotlib.pyplot as plt
from mpl_toolkits import mplot3d
from mpl_toolkits.mplot3d import axes3d
import numpy as np

def train(batch_size, lr, epochs, period):

assert period >= batch_size and period % batch_size == 0
w, b = init_params()
total_loss=[np.mean(square_loss(net(X, w, b), y).asnumpy())]
wi, bi = [], []
wi.append(w.asnumpy()[0,0])
bi.append(b.asnumpy()[0])

# Epoch starts from 1.

for epoch in range(1, epochs + 1):
if epoch > 2:
lr *= 0.9 # Decay learning rate.
for batch_i, data, label in data_iter(batch_size):
#for samples in a mini-batch
with autograd.record(): # define computational graph.
output = net(data, w, b)
loss = square_loss(output, label)
358 Machine Learning with Python: Theory and Applications

loss.backward() # compute the gradient

sgd([w, b], lr, batch_size) # update the parameters

if batch_i * batch_size % period == 0:

total_loss.append(np.mean(square_loss
(net(X,w,b),y).asnumpy()))
bi.append(b.asnumpy()[0])
wi.append(w.asnumpy()[0,0])

print("Batch size %d,Learning rate %f,Epoch %d,loss %.4e"

%(batch_size,lr,epoch, total_loss[-1]))

return wi,bi,total_loss,epochs

Wi,Bi,T_cost,Epochs=train(batch_size=10,lr=.2,epochs=5,
period=10)

print(Wi[-1],Bi[-1]) #the final converged weight and bias

Batch size 10,Learning rate 0.200000,Epoch 1,loss 2.0546e-02

Batch size 10,Learning rate 0.200000,Epoch 2,loss 2.0513e-02
Batch size 10,Learning rate 0.180000,Epoch 3,loss 2.0435e-02
Batch size 10,Learning rate 0.162000,Epoch 4,loss 2.1135e-02
Batch size 10,Learning rate 0.145800,Epoch 5,loss 2.1392e-02
0.78678846 4.246581

Finally, let us examine the training process.

nonlic =['linear','sigmoid','tanh','relu'] #nonlinearities

nsf = ['(a) ','(b) ','(c) ', '(d) ']

w11 = np.linspace(-2.0,4.0,100) # np.arange(0.5, 5.0, 5.0)

b1 = np.linspace(0.,8.0,100) # np.arange(2.0, 6.0, 5.0)
lw, lb = len(w11), len(b1)
print('length of w11=',lw,'length of b1=',lb)
W11, B1 = np.meshgrid(w11, b1)
Z = cost_function(X,W11,B1,y)
 Solution Existence Theory and Optimization Techniques 359

tw,tb = [true_w],[true_b]
true_z = cost_function(X,tw,tb,y)

print('length of Wi = ',len(Wi))
#%matplotlib notebook
%matplotlib inline

fig_s = plt.figure(figsize=(8,6)) # Convergence 3D plot

ax = fig_s.add_subplot(1,1,1,projection='3d')
ax.set_xlabel('$w11$', fontsize=12, rotation=80)
ax.set_ylabel('$b1$', fontsize=12)
ax.set_zlabel(r'$error$', fontsize=12, rotation=0)
ax.yaxis._axinfo['label']['space_factor'] = 3.0
ax.plot_wireframe(W11,B1,Z,color='b', rstride=5, cstride=5)
plt.title('Error function of $w11$ and
$b1$ (' + nonlic[0] +')')
marker_style = dict(markersize=20)
ax.scatter3D(Wi,Bi,T_cost, color='red',marker='.',
cmap='hsv',s=20)
ax.scatter3D(Wi[-1],Bi[-1],T_cost[-1],color='red',
marker='X', s=80)
ax.scatter3D(Wi[0], Bi[0],T_cost[0],color='green',
marker='o', s=50)
ax.scatter3D(true_w[0],true_b,true_z,color='green',
marker='o', s=50)
#plt.legend()
fig_s.tight_layout() # otherwise y-label is clipped
plt.show()

length of w11= 100 length of b1= 100

shape of W12: (100, 100, 1) shape of X: (1000, 1)
shape of yh is: (1000, 100, 100) shape of yz is: (1000, 1, 1)
shape of W12: (1, 1, 1) shape of X: (1000, 1)
shape of yh is: (1000, 1, 1) shape of yz is: (1000, 1, 1)
length of Wi = 501
360 Machine Learning with Python: Theory and Applications

Figure 9.11: The loss function in 2D space varying quadratically with weight and bias
for the linear regression problem.

fig, ax = plt.subplots(figsize=(8,6))
plt.rcParams["font.size"] = "15"
CS = ax.contour(W11, B1, Z,20)
ax.clabel(CS, inline=1, fontsize=9)
ax.set_title('A cost function of least square error')
plt.scatter(Wi, Bi, s=20, c='r',marker='^', alpha=0.5)
plt.scatter(Wi[-1], Bi[-1], s=80, c='r',marker='X',alpha=0.9)
plt.scatter(Wi[0], Bi[0], s=80, c='r', marker='o', alpha=0.9)
plt.scatter(true_w[0],true_b,s=80,c='g',marker='o',alpha=0.5)
ax.set_xlabel('$w11$', fontsize=12, rotation=0)
ax.set_ylabel('$b1$', fontsize=12)
fig.tight_layout()
plt.show()
 Solution Existence Theory and Optimization Techniques 361

Figure 9.12: Contours of the loss function in 2D space varying quadratically with weight
and bias for the linear regression problem. A minimization algorithm (SGD) finds the
minimum (green dot) starting from the initial values of weight and bias (red dot).

#line plots: https://jakevdp.github.io/PythonDataScience

#Handbook/04.06-customizing-legends.html

fig, ax1 = plt.subplots(figsize=(8,6))

plt.rcParams["font.size"] = "15"
x_axis = np.linspace(0, Epochs, len(T_cost), endpoint=True)
plt.semilogy(x_axis, T_cost,color='black',linestyle='solid',\
label='Loss function')
plt.xlabel('Epoch')
plt.ylabel('Loss')
leg = ax1.legend(loc='center', frameon=False)
ax2 = ax1.twinx() # second axes that shares the same x-axis
color = 'tab:blue'
ax2.set_ylabel('w11 and b1 ', color=color)
ax2.plot(x_axis,Wi,color='red',linestyle='dotted',
label='w11')
ax2.plot(x_axis,Bi,color='blue',linestyle='dashed',
label='b1')
ax2.tick_params(axis='y', labelcolor=color)
362 Machine Learning with Python: Theory and Applications

ax1.grid() # linestyle='-', linewidth='0.5', color='red')

fig.tight_layout() # otherwise y-label is slightly clipped
leg = ax2.legend()
plt.show()

Figure 9.13: Convergence of the loss function, the weight, and bias for the linear regression
problem, during the SGD searching for the minimizer.

%matplotlib inline
import matplotlib.pyplot as plt
print('true_w:',true_w,' true_b:',true_b)
x = np.arange(-4, 4, .1)
yh = Wi[-1] * x + Bi[-1]
print('Predicted_w:', Wi[-1],'Predicted_b:',Bi[-1])
fig = plt.figure(figsize=(8,6))
plt.rcParams["font.size"] = "12"
plt.scatter(X[:, 0].asnumpy(),y.asnumpy(),color='r')
plt.plot(x,yh,label = "$\hat{y} = x w + b$")
plt.xlabel('$x$')
plt.ylabel("$\hat{y}$")
plt.title('Linear Regression by Minimizing the Distances')
plt.legend()
plt.show()

true w: [0.8] true b: 4.2

Predicted w: 0.78678846 Predicted b: 4.246581
 Solution Existence Theory and Optimization Techniques 363

Figure 9.14: The results of the linear regression. It is found that a linear function fits
best to the data-points in the sense of the least squares of errors.

9.6 Gradient Descent with Momentum

The gradient descent (GD) and its variation of the stochastic gradient
descent (SGD) move in each iteration based solely on the current status of
the gradient information of the cost function. It does not use any information
from the past search history. Therefore, techniques that making use of the
history may perform better. The method of gradient descent momentum
(GDM) is one of these techniques.

9.6.1 The most critical problem with GD methods

Let us first examine the most critical problem with gradient descent type
of methods, including the GD and SGD. We use a simple example in two-
dimensional parameter space for easy intuitive analysis and understanding.
Consider a cost function f (x) with two parameters x = [x1 , x2 ]. In Fig. 9.15,
we plotted the contours of the functions value in 2D space. On each contour
line, the function has the same value. The green dot near the center is the
true minimum point of the function. Let the GD algorithm start to search
for the green dot from the initial point marked with a red star. Based on
the concept of steepest descent, the negative gradient direction should be
perpendicular to the contour line at the current point, because that direction
is the steepest. Because of the highly elliptical shape of these contours, a GD
step could overshoot and arrive at a contour line at the next lower level, and
get to the other side of the hill. At this new position, GD makes the next
364 Machine Learning with Python: Theory and Applications

Figure 9.15: Pictorial description of the zigzag behavior of the gradient descent algorithm
for searching the minimizer of an elliptic loss function in an extreme case. The long axis
is in the horizontal direction, and the short axis is in the vertical direction. “Productive”
gradient components are in the horizontal direction, and they are all aligned in the positive
direction towards the minimizer, but they are small, and hence progressing slowly. The
“unproductive” gradient components are in the vertical direction, because they reverse
directions each iteration resulting in a waste. The search traces on the right-lower part
(starting from the red dot) are the actual case that we have seen in the previous example.
We managed to reach the minimizer with rather a smoother path (at least at the earlier
iterations). This was at a cost using a very small learning rate.

marching step. This time, the negative gradient point is yet again on the
opposite side. This cycle continues, forming a zigzag path moving slowly
toward the minimum point. It is clear that the GD method wasted quite a
lot of effort in the process.
The root of the problem is the drastic change in direction of the
consecutive (negative) gradient vectors, from one step to the next. For this
type of cost functions with heavy elliptical contours, the two components
of the gradient vector are very much different (it corresponds to the bad
conditioning of Hessian matrix that contains the 2nd derivatives of the cost
function). All these marching arrows have two components pointing to the x1
(along the long axis of the ellipse) and x2 (along the short axis of the ellipse)
directions. They are basically two components of the (negative) gradient
vector. It is easy to see that the x1 components for all the marching arrows
are largely aligned and pointing more or less to the target. However, the x2
component of a marching arrow is largely opposite that of the next arrow.
With this insightful observation, we can now design an algorithm that
uses some kind of weighted average of the gradient vectors in consecutive
steps. If this can be done, all the x1 components are still encouraged to point
to the target, and all the x2 components will cancel each other, leading to a
significant reduction of the zigzag. This is essentially the idea of the method
of gradient descent with momentum.
 Solution Existence Theory and Optimization Techniques 365

9.6.2 Formulation
The gradient descent with momentum can be found in Ref. [3]. It is given
as follows:
vi := γvi−1 + η∇f (xi ) (9.43)
xi := xi−1 − vi (9.44)

where v is the velocity at the current step, and γ is the momentum parameter
that is between 0 and 1, and is usually set to 0.9. The learning rate η and
the gradient ∇f (x) are defined in the previous section.
To analyze the working behavior of the GDM, let us consider two
simplified cases. The first case assumes that all the ∇f (x) at consecutive
steps are pointing to the same direction and have the same values as a
constant vector g. This situation is quite similar to the component of the
gradient in the x1 -direction for the case shown in Fig. 9.15. We now start
the iteration as follows:
v0 := 0, (9.45)
v1 := ηg, (9.46)
v2 := γv1 + ηg = ηg(γ + 1), (9.47)
v3 := γv2 + ηg = ηg(γ 2 + γ + 1), (9.48)
... (9.49)
vt := γvt−1 + ηg = ηg(γ t−1 + γ t−2 + · · · + γ 2 + γ + 1), (9.50)
... (9.51)
ηg
vinf := . (9.52)
1−γ

Let us now examine the following cases:

1. For a large γ that close to 1, such as γ = 0.9, the final velocity would
be 10 times faster than the original gradient descent using the constant
g. Therefore, iterations with momentum can speed up the advancement
in the direction at which all the gradients are aligning on the same
direction, such as the component of the gradient in the x1 -direction shown
in Fig. 9.15.
2. For a small γ that close to 0, such as γ = 0.1, the effect on the acceleration
of the iteration is very small, and hence the GDM will not have much
difference from the original GD algorithm.
366 Machine Learning with Python: Theory and Applications

3. For a middle ranged γ, such as γ = 0.5, the speed up is about 50% in the
alignment direction.
This simple example shows that a momentum is built up to speed up the
marching faster and faster, if the gradients are all well aligned and a larger
γ is used.
In the second example, let us assume that all the ∇f (x) at consecutive
steps are pointing to the opposite direction, but still have the same values as a
constant vector g. This is similar to the x2 component under this assumption;
the standard GD method will not make any advancement in that direction.
Let us examine what would happen when the GDM is used.
We now start the iteration as follows:

v0 := 0, (9.53)
v1 := ηg, (9.54)
v2 := γv1 − ηg = −ηg(1 − γ), (9.55)
v3 := γv2 + ηg = ηg(1 − γ + γ 2 ), (9.56)
v4 := γv3 − ηg = −ηg(1 − γ + γ 2 − γ 3 ), (9.57)
... (9.58)
v2i−1 := ηg(1 − γ + γ 2 , . . . , −γ 2i−3 + γ 2i−2 ), (9.59)
v2i := −ηg(1 − γ + γ 2 , . . . , +γ 2i−2 − γ 2i−1 ), (9.60)

For any γ < 1, we note the following:

1. The learning is always reduced, as long as γ < 1.
2. The series behind ηg in Eqs. (9.59) and (9.60) will all bounded by 1, and
hence we shall have

v2i−1 ≤ ηg (9.61)
v2i ≥ −ηg (9.62)

Let us now examine the following cases:

1. For a large γ close to 1, such as γ = 0.9, the value of the (negative)

velocity at the 2nd iteration v2 is reduced by 90%, because of the
cancellation. Although v3 at the 3rd become larger again, it will always be
smaller than the original velocity. The same trend is kept in the following
iterations. Therefore, iterations with momentum will reduce the speed in
the direction at which all the gradients are opposite to each other, such
 Solution Existence Theory and Optimization Techniques 367

as the component of the gradient in the x2 -direction as shown Fig. 9.15.

Note that because the velocity in the x1 -direction is significantly speeding
up, the zigzag behavior will be significantly reduced.
2. For a small γ that close to 0, such as γ = 0.1, the cancellation effect is
very small, and hence the GDM will not have much difference from the
original GD algorithm.
3. For a middle ranged γ, such as γ = 0.5, the cancellation at early iterations
is about 50%. The cancellation effect is reduced in the later iterations,
but will always be smaller than the original velocity. Because the velocity
in the x1 -direction is speeding up about 50%, the zigzag behavior will be
significantly reduced. The convergence is expected to be more than 50%
faster.

Based on the above analysis, we can conclude that a large γ that close to 1,
say γ = 0.9, is preferred. In summary, the gradient descent with momentum
converges much faster, compared to the original GD, because updates in
the long axis are much faster, and in the short axis there are some levels of
cancellation. The zigzag behavior can be significantly suppressed.
From the algorithm, it is seen that the gradients obtained in all the pre-
vious iterations are involved in the current iteration. The next question one
may ask is what the extent of such involvements. Let us analyze this further.
We start the iteration again as follows considering the gradient is changing
in each iteration:
v0 := 0,
(9.63)
v1 := η∇f (x1 ), (9.64)
v2 := γv1 + η∇f (x2 ) = γη∇f (x1 ) + η∇f (x2 ), (9.65)
v3 := γv2 + η∇f (x3 ) = γ 2 η∇f (x1 ) + γη∇f (x2 ) + η∇f (x3 ), (9.66)
... (9.67)

vt := γvt−1 + η∇f (xt ) = γ t−1 η∇f (x1 ) + γ t−1 η∇f (x2 )

+ · · · + γη∇f (xt−1 ) + η∇f (xt ) (9.68)

Because γ < 1, it is clear that the influence of the gradient of the earlier
iterations slowly reduced at the rate of γ.
Finally, one may be inspired from this analysis that a properly varying
γ could be even better! We also note that to solve this kind zigzag behavior
entirely, the conjugate gradient method should be used, but it requires the
valuation of the 2nd gradient (the Hessian matrix), which can be a lot more
expensive and the scalability can also be a problem.
368 Machine Learning with Python: Theory and Applications

9.6.3 Numerical experiment

We now demonstrate an algorithm of GDM, by creating a linear regression

model.

def sgd_momentum(params, vs, lr, mom, batch_size):

for param, v in zip(params, vs):
v[:] = mom * v + lr * param.grad / batch_size
param[:] = param - v

import mxnet as mx
from mxnet import autograd, gluon
from mxnet import ndarray as nd
import random
import numpy as np

mx.random.seed(1); random.seed(1)
# Generate data.
n_v = 2 # number of variables (features)
n_s = 1000 # number of samples

true_w, true_b = [2, -3.4], 4.2

X=nd.random_normal(scale=1,shape=(n_s,n_v))
y=true_w[0] * X[:, 0] + true_w[1] * X[:, 1] + true_b
y+=.01 * nd.random_normal(scale=1, shape=y.shape)
dataset = gluon.data.ArrayDataset(X, y)
# Construct data iterator.
def data_iter(batch_size):
idx = list(range(n_s))
random.shuffle(idx)
for batch_i,i in enumerate(range(0,n_s,batch_size)):
j = nd.array(idx[i: min(i + batch_size, n_s)])
yield batch_i, X.take(j), y.take(j)

# Initialize model parameters.

def init_params():
w = nd.random_normal(scale=1, shape=(n_v, 1))
b = nd.zeros(shape=(1,))
 Solution Existence Theory and Optimization Techniques 369

params = [w, b]
vs = []
for param in params:
param.attach_grad()
#
vs.append(param.zeros_like())
return params, vs

# Linear regression.
def net(X, w, b):
return nd.dot(X, w) + b

# Loss function.
def square_loss(yhat, y):
return (yhat - y.reshape(yhat.shape)) ** 2 / 2

%matplotlib inline
import matplotlib as mpl
mpl.rcParams['figure.dpi']= 120
import matplotlib.pyplot as plt
import numpy as np

def train(batch_size, lr, mom, epochs, period):

assert period >= batch_size and period % batch_size == 0
[w, b], vs = init_params()
total_loss=[np.mean(square_loss(net(X,w,b),y).asnumpy())]
# Epoch starts from 1.
for epoch in range(1, epochs + 1):
# Decay learning rate.
if epoch > 2:
lr *= 0.1
for batch_i, data, label in data_iter(batch_size):
with autograd.record():
output = net(data, w, b)
loss = square_loss(output, label)
loss.backward()
sgd_momentum([w, b], vs, lr, mom, batch_size)
if batch_i * batch_size % period == 0:
total_loss.append(np.mean(square_loss\
(net(X,w,b),y).asnumpy()))
370 Machine Learning with Python: Theory and Applications

print("Batch size %d, Learning rate %f, Epoch %d,loss\

%.4e" % (batch_size, lr, epoch, total_loss[-1]))
print('true_w:',true_w,' true_b:',true_b)
print('Predicted_w:', np.reshape(w.asnumpy(), (1, -1)),
'Predicted_b:', b.asnumpy()[0], '\n')
x_axis=np.linspace(0,epochs,len(total_loss),endpoint=True)
plt.semilogy(x_axis, total_loss)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.show()

train(batch_size=10, lr=.2, mom=.9, epochs=3, period=10)

Batch size 10, Learning rate 0.20, Epoch 1,loss 1.4316e-02
Batch size 10, Learning rate 0.20, Epoch 2,loss 6.2213e-05
Batch size 10, Learning rate 0.02, Epoch 3,loss 5.2174e-05
true w: [2, -3.4] true b: 4.2
Predicted w: [[ 1.9999688 -3.4014332]] Predicted b: 4.1995444

Figure 9.16: Convergence of the loss function for the same linear regression problem,
during the SGD with momentum searching for the minimizer.

9.7 Nesterov Accelerated Gradient

9.7.1 Formulation

The Nesterov accelerated gradient (NAG) [4] method may be viewed as a

modified GDM. In computing the gradient in Eq. (9.43), instead of using xi ,
 Solution Existence Theory and Optimization Techniques 371

it uses (xi − γvi−1 ). The equations become

vi := γvi−1 + η∇f (xi − γvi−1 ) (9.69)

xi := xi−1 − vi (9.70)

It is clear that it looks ahead in computing the gradient at a near future

instead of at the current, aiming for faster convergence. We will observe its
behavior in the case study section at the end of this chapter.

9.8 AdaGrad Gradient Algorithm

Adaptive gradient algorithm (AdaGrad) is another modified stochastic

gradient descent (SGD) algorithm. It was first published in 2011 [5], allowing
the use of per-iteration and even per-parameter learning rates. It is designed
to increase the learning speed for sparser parameters and decrease it for dense
ones. It is claimed that this strategy often improves convergence performance
over standard SGD, and has been successfully applied to both convex and
non-convex problems. More detailed discussions may be found at the wike-
pedia page (https://en.wikipedia.org/wiki/Stochastic gradient descent).

9.8.1 Formulation
The Adagrad scales the base learning rate η and updates the parameters
using
η
xi := xi−1 − √ ∇f (xi ) (9.71)
Gi +
where is a small positive just for preventing a zero denominator during
the computation. Note that the above operations for these vectors are all
element-wise including the square-root operation, and Gi is also a vector
given by
i

Gi = (∇f (xq ))2 (9.72)
q=1

which is the sum of the element-wise product of the gradients of all the
previous iterations (which are accumulated during the iteration).
Note that this is done for each parameter. For parameters that received
excessive updates in the previous iterations, the current update is suppressed,
while parameters that get small updates receive larger learning rates.
372 Machine Learning with Python: Theory and Applications

For the strong elliptical objective functions discussed earlier, the Adagrad
tends promote advancement along the long axis (x1 -direction) and discour-
ages moves along the short axis (x2 -direction). This practically suppresses
the zigzag behavior shown in Fig. 9.15.
Let us examine the Adagrad algorithm using the same linear regression
problem.

9.8.2 Numerical experiment

# Adagrad.
def adagrad(params, sqrs, lr, batch_size):
eps_stable = 1e-7
for param, sqr in zip(params, sqrs):
g = param.grad / batch_size
sqr[:]+=nd.square(g) # Accumulate squared gradient
div=lr*g/nd.sqrt(sqr+eps_stable) #element-wise
param[:] -= div

import mxnet as mx
from mxnet import autograd, gluon
from mxnet import ndarray as nd
import random

mx.random.seed(1); random.seed(1)
# Generate data.
n_v = 2 # number of variables (features)
n_s = 1000 # number of samples
true_w, true_b = [2, -3.4], 4.2
X=nd.random_normal(scale=1,shape=(n_s,n_v))
y=true_w[0] * X[:, 0] + true_w[1] * X[:, 1] + true_b
y+=.01 * nd.random_normal(scale=1, shape=y.shape)
dataset = gluon.data.ArrayDataset(X, y)

# Construct data iterator.

def data_iter(batch_size):
idx = list(range(n_s))
random.shuffle(idx)
for batch_i,i in enumerate(range(0,n_s,batch_size)):
j = nd.array(idx[i: min(i + batch_size, n_s)])
yield batch_i, X.take(j), y.take(j)
 Solution Existence Theory and Optimization Techniques 373

# Initialize model parameters.

def init_params():
w = nd.random_normal(scale=1, shape=(n_v, 1))
b = nd.zeros(shape=(1,))
params = [w, b]
sqrs = []
for param in params:
param.attach_grad()
#
sqrs.append(param.zeros_like())
return params, sqrs

# Linear regression.
def net(X, w, b):
return nd.dot(X, w) + b

# Loss function.
def square_loss(yhat, y):
return (yhat - y.reshape(yhat.shape)) ** 2 / 2

%matplotlib inline
import matplotlib as mpl
mpl.rcParams['figure.dpi']= 120
import matplotlib.pyplot as plt
import numpy as np

def train(batch_size, lr, epochs, period):

assert period >= batch_size and period % batch_size == 0
[w, b], sqrs = init_params()
total_loss=[np.mean(square_loss(net(X, w, b),y).asnumpy())]

# Epoch starts from 1.

for epoch in range(1, epochs + 1):
for batch_i, data, label in data_iter(batch_size):
with autograd.record():
output = net(data, w, b)
loss = square_loss(output, label)
loss.backward()
adagrad([w, b], sqrs, lr, batch_size)
if batch_i * batch_size % period == 0:
total_loss.append(np.mean(square_loss\
(net(X,w,b),y).asnumpy()))
374 Machine Learning with Python: Theory and Applications

print("Batch size %d, Learning rate %f, Epoch %d",

"loss %.4e" %(batch_size,lr,epoch,total_loss[-1]))
print('true_w:',true_w,' true_b:',true_b)
print('Predicted_w:', np.reshape(w.asnumpy(), (1, -1)),
'Predicted_b:', b.asnumpy()[0], '\n')
x_axis=np.linspace(0,epochs,len(total_loss),endpoint=True)
plt.semilogy(x_axis, total_loss)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.show()

train(batch_size=10, lr=0.9, epochs=3, period=10)

Batch size 10, Learning rate 0.90, Epoch 1, loss 5.1843e-05

Batch size 10, Learning rate 0.90, Epoch 2, loss 5.2248e-05
Batch size 10, Learning rate 0.90, Epoch 3, loss 5.1507e-05
true w: [2, -3.4] true b: 4.2
Predicted w: [[ 2.0000997 -3.4004908]] Predicted b: 4.199482

Figure 9.17: Convergence of the loss function for the same linear regression problem,
during the AdaGrad searching for the minimizer.

9.9 RMSProp Gradient Algorithm

RMSProp is short for Root Mean Square Propagation [6]. It scales the learn-
ing rate for each of the parameters. It has been shown that RMSProp has
good adaptation of learning rate for different applications. More details can
 Solution Existence Theory and Optimization Techniques 375

be found at the Wikipedia page (https://en.wikipedia.org/wiki/Stochastic

gradient descent).

9.9.1 Formulation
RMSProp scales the base learning rate η and updates the parameters using
η
xi := xi−1 − √ ∇f (xi ) (9.73)
vi +
where is a small positive number for preventing a zero denominator during
the computation. Notice again that the operations are element-wise. And,
vi at the ith iteration is given as follows:

vi := γvi−1 + (1 − γ)(∇f (xi ))2 (9.74)

where γ is the tunable positive number called forgetting factor.

Let us observe the behavior.

9.9.2 Numerical experiment

# RMSProp.
def rmsprop(params, sqrs, lr, gamma, batch_size):
eps_stable = 1e-8
for param, sqr in zip(params, sqrs):
g = param.grad / batch_size
sqr[:]=gamma*sqr+(1.-gamma)*nd.square(g)
#note the in-place computation here
div = lr * g / nd.sqrt(sqr + eps_stable)
param[:] -= div

import mxnet as mx
from mxnet import autograd, gluon
from mxnet import ndarray as nd
import random

mx.random.seed(1); random.seed(1)
# Generate data.
n_v = 2 # number of variables (features)
n_s = 1000 # number of samples
376 Machine Learning with Python: Theory and Applications

true_w, true_b = [2, -3.4], 4.2

X = nd.random_normal(scale=1, shape=(n_s, n_v))
y = true_w[0] * X[:, 0] + true_w[1] * X[:, 1] + true_b
y += .01 * nd.random_normal(scale=1, shape=y.shape)
dataset = gluon.data.ArrayDataset(X, y)
# Construct data iterator.
def data_iter(batch_size):
idx = list(range(n_s))
random.shuffle(idx)
for batch_i,i in enumerate(range(0,n_s,batch_size)):
j = nd.array(idx[i: min(i + batch_size, n_s)])
yield batch_i, X.take(j), y.take(j)

# Initialize model parameters.

def init_params():
w = nd.random_normal(scale=1, shape=(n_v, 1))
b = nd.zeros(shape=(1,))
params = [w, b]
sqrs = []
for param in params:
param.attach_grad()
sqrs.append(param.zeros_like())
return params, sqrs

# Linear regression.
def net(X, w, b):
return nd.dot(X, w) + b

# Loss function.
def square_loss(yhat, y):
return (yhat - y.reshape(yhat.shape)) ** 2 / 2

%matplotlib inline
import matplotlib as mpl
mpl.rcParams['figure.dpi']= 120
import matplotlib.pyplot as plt
import numpy as np
 Solution Existence Theory and Optimization Techniques 377

def train(batch_size, lr, gamma, epochs, period):

assert period >= batch_size and period % batch_size == 0
[w, b], sqrs = init_params()
total_loss=[np.mean(square_loss(net(X,w,b),y).asnumpy())]

# Epoch starts from 1.

for epoch in range(1, epochs + 1):
for batch_i, data, label in data_iter(batch_size):
with autograd.record():
output = net(data, w, b)
loss = square_loss(output, label)
loss.backward()
rmsprop([w, b], sqrs, lr, gamma, batch_size)
if batch_i * batch_size % period == 0:
total_loss.append(np.mean(square_loss\
(net(X, w, b), y).asnumpy()))
print("Batch size %d,Learning rate %f, Epoch %d,loss\
%.4e" %(batch_size, lr, epoch, total_loss[-1]))

print('true_w:',true_w,' true_b:',true_b)
print('Predicted_w:', np.reshape(w.asnumpy(), (1, -1)),
'Predicted_b:', b.asnumpy()[0], '\n')
x_axis=np.linspace(0,epochs,len(total_loss),
endpoint=True)
plt.semilogy(x_axis, total_loss)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.show()

train(batch_size=10, lr=.02, gamma=.9, epochs=3, period=10)

Batch size 10,Learning rate 0.020, Epoch 1,loss 6.2608e+00

Batch size 10,Learning rate 0.020, Epoch 2,loss 5.9297e-01
Batch size 10,Learning rate 0.020, Epoch 3,loss 5.6944e-05
true w: [2, -3.4] true b: 4.2
Predicted w: [[ 2.0020406 -3.397886 ]] Predicted b: 4.2007647
378 Machine Learning with Python: Theory and Applications

Figure 9.18: Convergence of the loss function for the same linear regression problem,
during the RMSProp searching for the minimizer.

9.10 AdaDelta Gradient Algorithm

9.10.1 The idea

Adadelta gradient algorithm is a more robust extension of Adagrad [7]. It

adapts learning rates using a moving window of gradients, instead of accu-
mulating all past gradients as in the Adagrad. In using AdaDelta, one does
not have to set an initial learning rate. A short and clearly written note on
AdaDelta can also be found at http://akyrillidis.github.io/notes/AdaDelta.

9.10.2 Numerical experiment

# Adadelta.
def adadelta(params, sqrs, deltas, rho, batch_size):
eps = 1e-5 # epsilon for stabilization
for param, sqr, delta in zip(params, sqrs, deltas):
g = param.grad / batch_size
sqr[:] = rho * sqr + (1. - rho) * nd.square(g)
cur_delta=nd.sqrt(delta+eps)/nd.sqrt(sqr+eps)*g
delta[:]=rho*delta+(1.-rho)*cur_delta*cur_delta
# update weight
param[:] -= cur_delta

import mxnet as mx
from mxnet import ndarray as nd
from mxnet import autograd, gluon
import random
 Solution Existence Theory and Optimization Techniques 379

mx.random.seed(1)
# Generate data.
n_v = 2 # number of variables (features)
n_s = 1000 # number of samples
true_w, true_b = [2, -3.4], 4.2
X = nd.random_normal(scale=1,shape=(n_s,n_v))
y = true_w[0] * X[:, 0] + true_w[1] * X[:, 1] + true_b
y += .01 * nd.random_normal(scale=1, shape=y.shape)
dataset = gluon.data.ArrayDataset(X, y)
# Construct data iterator.
def data_iter(batch_size):
idx = list(range(n_s))
random.shuffle(idx)
for batch_i,i in enumerate(range(0,n_s,batch_size)):
j = nd.array(idx[i: min(i + batch_size, n_s)])
yield batch_i, X.take(j), y.take(j)

# Initialize model parameters.

def init_params():
w = nd.random_normal(scale=1, shape=(n_v, 1))
b = nd.zeros(shape=(1,))
params = [w, b]
sqrs = []
deltas = []
for param in params:
param.attach_grad()
#
sqrs.append(param.zeros_like())
deltas.append(param.zeros_like())
return params, sqrs, deltas

# Linear regression.
def net(X, w, b):
return nd.dot(X, w) + b

# Loss function.
def square_loss(yhat, y):
return (yhat - y.reshape(yhat.shape)) ** 2 / 2
380 Machine Learning with Python: Theory and Applications

%matplotlib inline
import matplotlib as mpl
mpl.rcParams['figure.dpi']= 120
import matplotlib.pyplot as plt
import numpy as np

def train(batch_size, rho, epochs, period):

assert period >= batch_size and period % batch_size == 0
[w, b], sqrs, deltas = init_params()
total_loss=[np.mean(square_loss(net(X,w,b),y).asnumpy())]

# Epoch starts from 1.

for epoch in range(1, epochs + 1):
for batch_i, data, label in data_iter(batch_size):
with autograd.record():
output = net(data, w, b)
loss = square_loss(output, label)
loss.backward()
adadelta([w, b], sqrs, deltas, rho, batch_size)
if batch_i * batch_size % period == 0:
total_loss.append(np.mean(square_loss\
(net(X, w, b), y).asnumpy()))
print("Batch size %d, Epoch %d, loss %.4e" %
(batch_size, epoch, total_loss[-1]))

print('true_w:',true_w,' true_b:',true_b)
print('Predicted_w:', np.reshape(w.asnumpy(), (1, -1)),
'Predicted_b:', b.asnumpy()[0], '\n')
x_axis=np.linspace(0,epochs,len(total_loss), endpoint=True)
plt.semilogy(x_axis, total_loss)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.show()

train(batch_size=10, rho=.9999, epochs=3, period=10)

Batch size 10, Epoch 1, loss 5.3295e-05

Batch size 10, Epoch 2, loss 5.2909e-05
Batch size 10, Epoch 3, loss 5.2828e-05
true w: [2, -3.4] true b: 4.2
Predicted w: [[ 1.999908 -3.4017406]] Predicted b: 4.19924
 Solution Existence Theory and Optimization Techniques 381

Figure 9.19: Convergence of the loss function for the same linear regression problem,
during the AdaDelta searching for the minimizer.

9.11 Adam Gradient Algorithm

Adaptive Moment Estimation (Adam) gradient algorithm [8] is a modified

version of the RMSProp algorithm. It uses running averages of both the
RMSProp gradients and the second moments of the gradients.

9.11.1 Formulation
At the ith training iteration, for given parameters xi and a loss function
f (xi ), Adam algorithm scales the base learning rate η and updates the
parameters using
ηmi
xi := xi−1 − √ (9.75)
v̂i +
where is a small positive number for preventing a zero denominator during
the computation, and mi and v̂i are given as follows:
mi
mi := (9.76)
1 − β1
vi
v̂i := (9.77)
1 − β2
in which mi and vi , are, respectively, updated using

mi := β1 mi−1 + (1 − β1 )(∇f (xi )) (9.78)

vi := β2 vi−1 + (1 − β2 )[∇f (xi )]2 (9.79)

In Eqs. (9.78) and (9.79), β1 and β2 are all positive constants of forgetting
factors, respectively, for the gradients and the second moments of the
gradients.
382 Machine Learning with Python: Theory and Applications

The Adam algorithm is essentially a combination of the GDM and the

RMSProp. It has features from these two. It is carefully formulated with
statistical bias corrected by dividing (1 − β) for both the first and second
momentums. To see this, let us examine the case of the first momentum, as
an example.
t = 0 : m0 = 0 (9.80)
t = 1 : m1 = (1 − β1 )∇f (x1 ) (9.81)
t = 2 : m2 = β1 (1 − β1 )∇f (x1 ) + (1 − β1 )∇f (x2 ) (9.82)
t = 3 : m3 = β12 (1 − β1 )∇f (x1 ) + β1 (1 − β1 )∇f (x2 )
+(1 − β1 )∇f (x3 ) (9.83)
... (9.84)

t
t = t : mt = (1 − β1 ) β1t−i ∇f (xi ) (9.85)
i=1

The expectation for Equation (9.85) becomes

 t

 1 − β1t
t−i
E[mt ] = E (1 − β1 ) β1 ∇f (xi ) = (1 − β1 ) E[∇f (xt )]
1 − β1
i+1

= (1 − β1t )E[∇f (xt )] (9.86)

To ensure the statistical equivalence in expectation between mt and ∇f (xt ),
mt needs to be divided by (1 − β1t ).
Let us observe the behavior.

9.11.2 Numerical experiment

# Adam.
def adam(params, vs, sqrs, lr, batch_size, t):
beta1, beta2 = 0.9, 0.999
eps = 1e-8 # epsilon for stabilization
for param, v, sqr in zip(params, vs, sqrs):
g = param.grad / batch_size
v[:] = beta1 * v + (1.-beta1)*g # momentum
sqr[:] = beta2 * sqr + (1.-beta2)* nd.square(g)
v_bias_corr = v / (1. - beta1 ** t)
sqr_bias_corr = sqr / (1. - beta2 ** t)
 Solution Existence Theory and Optimization Techniques 383

div=lr*v_bias_corr/(nd.sqrt(sqr_bias_corr)+eps)
param[:] = param - div #RMSprop

import mxnet as mx
from mxnet import autograd, gluon
from mxnet import ndarray as nd
import random

mx.random.seed(1); random.seed(1)
# Generate data.
n_v = 2 # number of variables (features)
n_s = 1000 # number of samples
true_w, true_b = [2, -3.4], 4.2
X = nd.random_normal(scale=1, shape=(n_s, n_v))
y = true_w[0] * X[:, 0] + true_w[1] * X[:, 1] + true_b
y += .01 * nd.random_normal(scale=1, shape=y.shape)
dataset = gluon.data.ArrayDataset(X, y)

# Construct data iterator.

def data_iter(batch_size):
idx = list(range(n_s))
random.shuffle(idx)
for batch_i, i in enumerate(range(0, n_s, batch_size)):
j = nd.array(idx[i: min(i + batch_size, n_s)])
yield batch_i, X.take(j), y.take(j)

# Initialize model parameters.

def init_params():
w = nd.random_normal(scale=1, shape=(n_v, 1))
b = nd.zeros(shape=(1,))
params = [w, b]
vs = []
sqrs = []
for param in params:
param.attach_grad()
vs.append(param.zeros_like())
sqrs.append(param.zeros_like())
return params, vs, sqrs
384 Machine Learning with Python: Theory and Applications

# Linear regression.
def net(X, w, b):
return nd.dot(X, w) + b

# Loss function.
def square_loss(yhat, y):
return (yhat - y.reshape(yhat.shape)) ** 2 / 2

%matplotlib inline
import matplotlib as mpl
mpl.rcParams['figure.dpi']= 120
import matplotlib.pyplot as plt
import numpy as np

def train(batch_size, lr, epochs, period):

assert period >= batch_size and period % batch_size == 0
[w, b], vs, sqrs = init_params()
total_loss=[np.mean(square_loss(net(X,w,b),y).asnumpy())]

t = 0
# Epoch starts from 1.
for epoch in range(1, epochs + 1):
for batch_i, data, label in data_iter(batch_size):
with autograd.record():
output = net(data, w, b)
loss = square_loss(output, label)
loss.backward()
# Increment t before invoking adam.
t += 1
adam([w, b], vs, sqrs, lr, batch_size, t)
if batch_i * batch_size % period == 0:
total_loss.append(np.mean(square_loss\
(net(X, w, b), y).asnumpy()))
print("Batch size %d, Learning rate %f, Epoch %d",\
"loss %.4e"%(batch_size,lr,epoch,total_loss[-1]))

print('true_w:',true_w,' true_b:',true_b)
print('Predicted_w:', np.reshape(w.asnumpy(), (1, -1)),
'Predicted_b:', b.asnumpy()[0], '\n')
 Solution Existence Theory and Optimization Techniques 385

x_axis = np.linspace(0,epochs,len(total_loss),
endpoint=True)
plt.semilogy(x_axis, total_loss)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.show()

train(batch_size=10, lr=0.1, epochs=3, period=10)

Batch size 10, Learning rate 0.100, Epoch 1, loss 5.7667e-04

Batch size 10, Learning rate 0.100, Epoch 2, loss 5.2377e-05
Batch size 10, Learning rate 0.100, Epoch 3, loss 5.1530e-05
true w: [2, -3.4] true b: 4.2
Predicted w: [[ 2.0000498 -3.4008262]] Predicted b: 4.1996875

Figure 9.20: Convergence of the loss function for the same linear regression problem,
during the Adam searching for the minimizer.

9.12 A Case Study: Compare Minimization Techniques

Used in MLPClassifier

We now present a case study done by the scikit-learn team [9]. This study
examines the performance and plots the loss function evolution, using differ-
ent stochastic minimization techniques, including SGD and Adam. The code
is available at Sklearn example site (https://scikit-learn.org/stable/auto
examples/neural networks/plot mlp training curves.html#). Figure 9.21 is
the result obtained using the code.
It is observed that SDM (momentum), Nesterov’s SDM, and Adam are
among the best performers, and Adam seems to be the overall best, based
386 Machine Learning with Python: Theory and Applications

Figure 9.21: Results of a comparison study on minimization techniques based on gradient

descent algorithms using Sklearn MLPClassifiers.

on this study on four datasets (iris, digits, circles, and moons). Note that
there may be some bias to some techniques, and the outcome of the tests
can change for different datasets. This study shall still give a good general
indication on these different techniques. Note also that those results can be
highly dependent on some parameters. A more detailed examination of this
study can be found at the Sklearn example site.

9.13 Other Algorithms

• There are many other versions of gradient descent algorithms available in

the open literature.
• Another direction of development is on higher-order methods and algo-
rithms, including Newton Methods, Quasi-Newton methods, and Conju-
gate Gradient methods.
• These methods often use line search algorithms to ensure a reasonable
quality solution at each iteration so that the matching can be significantly
faster and also provide necessary conditions for convergence.
 Solution Existence Theory and Optimization Techniques 387

• In general, high-order methods can drastically reduce the number of

iteration (by as much as hundreds of times).
• In each iteration, however, high-order methods take a lot more time for
line search and new directions.
• In general, high-order methods converge very fast near the minimum,
because of the improved local convexity there and also the quadratic
convergence property.
• The robustness (sensitive to line search and noise data) and efficiency for
higher-order methods for complex and large machine learning problems
may need further improvements.
• It may be a good idea to use a first-order robust algorithm to drive the
solution to near a minimum, and use a higher-order algorithm to finish
it up.

References

[1] G.R. Liu and S.S. Quek, The Finite Element Method: A Practical Course, Butterworth-
Heinemann, London, 2013.
[2] G.R. Liu, Solution existence theory for artificial neural networks, International Journal
of Computational Methods, 19(8), in-printing, 2022.
[3] D.E. Rumelhart, G.E. Hinton, and R.J. Williams, Learning representations by
back-propagating errors, Nature, 323(6088), 533–536, 1986. http://www.nature.com/
articles/323533a0.
[4] Y. Nesterov, A method for unconstrained convex minimization problem with the rate
of convergence o(1/kˆ2), Doklady ANSSSR (translated as Soviet.Math.Docl.), 269,
543–547, 1983.
[5] J.C. Duchi, H. Elad and S. Yoram, Adaptive subgradient methods for online learning
and stochastic optimization, J. Mach. Learn. Res., 12, 2121–2159, 2011. http://dblp.
uni-trier.de/db/journals/jmlr/jmlr12.html#DuchiHS11.
[6] T. Tieleman and G. Hinton, Lecture 6.5-rmsprop: Divide the gradient by a running
average of its recent magnitude, COURSERA, Neural Networks for Machine Learning,
4(1), 26–31, 2012.
[7] M. Zeiler, ADADELTA: An adaptive learning rate method, arXiv, 1212.5701, 2012.
[8] D.P. Kingma and J. Ba, Adam: A method for stochastic optimization, 2014. http://
arxiv.org/abs/1412.6980.
[9] P. Fabian, V. Gae, G. Alexandre et al., Scikit-learn: Machine Learning in Python,
Journal of Machine Learning Research, 12(85), 2825–2830, 2011. http://jmlr.org/
papers/v12/pedregosa11a.html.
MACHINE LEARNING
WITH PYTHON
 Chapter 10

Loss Functions for Regression

Regression is one of the most widely used tools in analyses of engineering

systems. It is also an important technique in the construction of numerical
models for both physics-law-based methods (such as the point interpolation
methods PIM [1], and the moving least square or MLS used in the meshfree
methods [2]) and machine learning models. In a machine learning setting, it
takes inputs of a dataset and produces a prediction of a real number, once
a regression model is established/trained.
In Chapter 9, we discussed a linear regression problem that has an
analytic close-form solution using an L2 loss function, revealing conditions
for solution existence. This chapter introduces first various types of loss
functions for regression problems. The common feature of all loss functions
is that they all produce a positive real number for given learning parameters
and a dataset. Our discussion will focus on their differences. Python codes
are presented for solutions to the normal equation derived in Chapter 9.
Linear regressions are performed on both linear and nonlinear datasets, using
corresponding order of bases as features.
A simple neural network will then be built with one layer of input
neurons and one neuron in the output layer. It is trained to perform a
task of linear regression, using stochastic gradient descent (SGD) as min-
imizer to numerically obtain the solution. Detailed procedures will be pre-
sented using the materials and code from mxnet-the-straight-dope (https://
github.com/zackchase/mxnet-the-straight-dope/tree/master/chapter06
optimization) under the Apache License 2. Codes developed by the scikit-
learn team have also been used [3].

389
390 Machine Learning with Python: Theory and Applications

10.1 Formulations for Linear Regression

10.1.1 Mathematical model

The mathematical model for regression is to predict a real valued y for a
given dataset with p linearly independent features defined in p-dimensional
spaces Xp . An arbitrary point in the feature space can be expressed as

x = [x1 , x2 , . . . , xp ] (10.1)

In engineering design problems, these features may be the design parameters

of an engineering structure, for example.
In linear regression, we assume that the prediction function ŷ is the
affine transformation z(w, b; x):
p

ŷ ≡ z(w, b; x) = W 1 x1 + · · · + W p xp + b = Wi xi + b = x · w + b (10.2)
i=1

where w = [W 1 , W 2 , . . . , W p ] ∈ Wp is called weights and b ∈ W1 is called

bias. The learning (or training) parameters are in Wp+1 .

10.1.2 Neural network configuration

If we build a neural network to perform linear regressions, the typical
architecture of this type of nets can be shown as follows.
As shown in Fig. 10.1, the loss function L(ŷ, y) needs to be minimized for
a given set of data-points, so as to force the prediction as close as possible to

Figure 10.1: A simple p → 1 neural network with one input layer of multiple neurons
each takes one feature input, and one output layer of a single neuron that produces a
prediction ŷ, using the xw+b formulation.
 Loss Functions for Regression 391

the label y. This is done through training of the network, using a dataset as
inputs together with their corresponding label value of y. The data structure
can be exactly the same as those discussed in Chapter 5, by simply replacing
z by ŷ. This replacement is usually done for the last layer of output neurons
where a comparison of the prediction is performed against the label y.

10.1.3 The xw formulation

By the xw formulation, Eq. (10.2) can be written as

p
 p

ŷ = Wi xi + b = Wi xi = xŵ (10.3)
i=1 i=0

in which x0 = 1, and W 0 = b. It is more obvious to state that the information

taken by a neuron is simply a weighted aggregation for the inputs. In this
case, the neural network has the following architecture is shown in Fig. 10.2.

Figure 10.2: A simple p → 1 neural network with one input layer of multiple neurons and
one output layer of a single neuron, using the xw formulation.

10.2 Loss Functions for Linear Regression

An ideal loss function should be largely quadratic in learning parameters

for a given type of dataset and hypothesis of the ML model used, so that a
minimization technique can work effectively to find the minimizers. The L2
loss function is such a function for linear regression problems.
392 Machine Learning with Python: Theory and Applications

10.2.1 Mean squared error loss or L2 loss function

Loss (cost, objective) function is needed to measure how far the model
predictions are from the true label data. It is used by the optimizer to
tune the training parameters. Here, we introduce a widely used loss function
in engineering and sciences, known as mean-squared-error (MSE) or L2
loss function, which is used in Chapter 9. Assume we have m labels or
measured/observed actual data. The loss function can be defined as follows:
m m
1  1  1  1
L(ŷ, y) = (ŷi − yi )2 = (ri )2 = (r r) = r22 (10.4)
2m    2m 2m 2m
i=1 ri i=1

where i is the dummy index for the ith training sample and m is the total
number of samples in the training dataset. ri is the ith entry in residual
vector r and is the residual caused by the ith prediction against the label.
Equation (10.4) gives a loss functional that maps m sets of data to a positive
real number. The MSE loss function is called L2 loss function because it uses
the L2 norm, as seen in Eq. (10.4).
The MSE or L2 loss function has the following properties:

1. It is quadratic in predictions, and hence will be quadratic in weights and

biases when affine transformation is used in linear regression. We can
have a close-form solution for the minimizers, as shown in the previous
chapter.
2. It is smooth and convex with respect to the residual and hence convenient
to use with optimization algorithms that use gradients.
3. The L2 loss is (Euclidean) distance based because it depends only on the
residual, and hence it is also called residual loss. When the residual is
zero, it is zero.
4. In addition, it is translation invariant, because the L2 loss is distance
based. This means that

L(ŷ + c, y + c) = L(ŷ, y) (10.5)

where c is an arbitrary constant vector. Therefore, it may not be used for

measures using relative error, such as ŷiy−y
i
i
.
5. Because of the square operation, it overestimates the error when the
residual is larger than 1, and hence it is sensitive to the outliers (few
data samples that stand out of the reset). It underestimates the error
when the residual is less than 1. When the error level is low, the loss
function becomes insensitive to the change of the residual and in turn to
 Loss Functions for Regression 393

Figure 10.3: Schematic drawing for linear regression problems using loss function
measuring the average squared distances from the data-points to the linear prediction
function.

the training parameters. Therefore, one can expect a fast convergence at

the beginning and slow convergence when coming near the minimum.

For our linear regression problems, the loss function measures the average
squared distances from the data-points to the straight line of prediction, as
shown in Fig. 10.3 for one-dimensional problems, where we have only one
design parameter or feature, one weight, and one bias.
The error is linear in ŷ, which is in turn linear in weights and biases.
Therefore, the squared error loss function is quadratic in weights and biases.
The minimization would lead to a set of linear system equations, and we
can have a close-form solution for the minimizers, as shown in Chapter 9. In
this chapter, we compute the solution numerically and demonstrate how a
neural network can be trained to do this task.

10.2.2 Absolute error loss or L1 loss function

Another loss function is the mean absolute error (MAE) loss function or L1
loss for short. It also measures the distance between the model predictions
and the true label data. Assume we have m labels or measured/observed
actual data. The L1 loss function is given by
m m
1  1  1
L(ŷ, y) = (| ŷi − yi |) = |ri | = r1 (10.6)
m    m m
i=1 ri i=1
394 Machine Learning with Python: Theory and Applications

The MAE or L1 loss function is also called Laplace loss. It has the
following properties:
1. It is still convex with respect to the residual by definition or by the fact
that all norms are convex.
2. The L1 loss is, however, not differentiable at the sample point where
the residual is zero. Special techniques (sub-gradient) are needed to use
gradient-based optimization algorithms.
3. It is distance based depending only on the residual. When the residual is
zero, it is zero. It is also translation invariant (distance-based).
4. Because L1 measures the absolute residuals, it is linearly related to
the residuals. It is thus more robust to outliers compared to its L2
counterpart.

10.2.3 Huber loss function

Huber loss is a combination of the L1 and L2 losses. It is less sensitive to
outliers in the dataset and behaves more like the L1 in this regard. It is made
differentiable at 0 by letting it be quadratic when the residual is small. The
transition from linear to quadratic depends on a hyperparameter δ that can
be tuned. Huber loss approaches the L1 loss when δ is near zero, and L2 loss
when δ become a very large number. The formula is as follows:
⎧ 1
⎪
⎨ ŷ − y22 when ŷ − y1 ≤ δ;
2m
L(ŷ, y, δ) = (10.7)
⎪
⎩ 1 δŷ − y − 1 δ2 Otherwise.
1 2m
m
The Huber loss has the following properties:
1. It is convex, and is quadratic when the residual is near zero.
2. It is differentiable everywhere.
3. It has a tunable hyperparameter.
4. The error measure is linear in residuals when it is large. It is thus less
sensitive to outliers compared to the L2 counterpart.

10.2.4 Log-cosh loss function

Log-cosh loss is a logarithm of the hyperbolic cosine of the residual. The
formula is as follows:
m m
1  1 
L(ŷ, y) = log(cosh(ŷi − yi )) = log(cosh(ri )) (10.8)
m    m
i=1 ri i=1
 Loss Functions for Regression 395

2
1. It is approximately equal to r2 for a small residual, and hence it behaves
more like the L2 loss near zero.
2. It approaches abs(r)−log(2) when the residual is large, and hence is more
like the L1 loss.
3. It is convex, smooth, and differentiable everywhere.
4. It is less sensitive to outliers compared to the L2 loss.

Let us now write a Python code to examine these loss functions in more
detail in numbers and in graphs, using only one sample. We first check all
of them in numbers.

10.2.5 Comparison between these loss functions

import numpy as np
huber=lambda r,dlt: np.where(np.abs(r)<dlt,0.5*((r)**2),\
dlt*np.abs(r)-0.5*(dlt**2))
for r in np.linspace(0.1,1.0,10): #range(0.1,1, 11):
print(f'r={r:.3f} |r|= {r:.3f} r^2= {r*r:.3f}\
Huber={huber(r,1.0):.3f} logcosh={np.log(np.cosh(r)):.3f}')
r=0.100 |r|= 0.100 r^2= 0.010 Huber=0.005 logcosh=0.005
r=0.200 |r|= 0.200 r^2= 0.040 Huber=0.020 logcosh=0.020
r=0.300 |r|= 0.300 r^2= 0.090 Huber=0.045 logcosh=0.044
r=0.400 |r|= 0.400 r^2= 0.160 Huber=0.080 logcosh=0.078
r=0.500 |r|= 0.500 r^2= 0.250 Huber=0.125 logcosh=0.120
r=0.600 |r|= 0.600 r^2= 0.360 Huber=0.180 logcosh=0.170
r=0.700 |r|= 0.700 r^2= 0.490 Huber=0.245 logcosh=0.227
r=0.800 |r|= 0.800 r^2= 0.640 Huber=0.320 logcosh=0.291
r=0.900 |r|= 0.900 r^2= 0.810 Huber=0.405 logcosh=0.360
r=1.000 |r|= 1.000 r^2= 1.000 Huber=0.500 logcosh=0.434

We notice that the L2, Huber, and log-cosh losses are very insensitive to
small residuals.

for r in np.linspace(10,100,10): #range(0.1,1, 11):

print(f'r={r:.1f} |r|= {r:.1f} r^2= {r*r:.1f}\
Huber={huber(r,1.0):.1f} logcosh={np.log(np.cosh(r)):.1f}')
r=10.0 |r|= 10.0 r^2= 100.0 Huber=9.5 logcosh=9.3
r=20.0 |r|= 20.0 r^2= 400.0 Huber=19.5 logcosh=19.3
r=30.0 |r|= 30.0 r^2= 900.0 Huber=29.5 logcosh=29.3
396 Machine Learning with Python: Theory and Applications

r=40.0 |r|= 40.0 r^2= 1600.0 Huber=39.5 logcosh=39.3

r=50.0 |r|= 50.0 r^2= 2500.0 Huber=49.5 logcosh=49.3
r=60.0 |r|= 60.0 r^2= 3600.0 Huber=59.5 logcosh=59.3
r=70.0 |r|= 70.0 r^2= 4900.0 Huber=69.5 logcosh=69.3
r=80.0 |r|= 80.0 r^2= 6400.0 Huber=79.5 logcosh=79.3
r=90.0 |r|= 90.0 r^2= 8100.0 Huber=89.5 logcosh=89.3
r=100.0 |r|= 100.0 r^2= 10000.0 Huber=99.5 logcosh=99.3

We notice this time that the L2 is very sensitive to large residuals.

10.2.6 Python codes for these loss functions

Next, we take a look at the graphs for all these loss functions by running
our codes.
# The mean squared error loss (function)
def mse(ylabel, yhat):
r = ylabel-yhat
return 0.5*r**2

# The mean absolute error loss

def mae(ylabel, yhat):
r = ylabel-yhat
return np.abs(r)

# The huber loss

def huber(ylabel, yhat, dlt=1.0):
r = ylabel-yhat
abr = np.abs(r)
loss=np.where(abr<dlt,0.5*((r)**2), dlt*abr-0.5*(dlt**2))
return loss

# The log-cosh loss

def logcosh(ylabel, yhat):
r = ylabel-yhat
return np.log(np.cosh(r))

%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
from matplotlib import rc
 Loss Functions for Regression 397

plt.figure(figsize=(4.5, 2.8),dpi=100)

xleft,xright,xdelta =-1.5,1.6, 0.1

rs = np.arange(xleft,xright,xdelta) # Residuals
lxd = len(rs)
y = np.arange(xleft,xright,xdelta) # y_label

y_hat = y + rs # simulated prediction

plt.plot(rs,mse(y,y_hat),label = "MSE loss")

plt.plot(rs,mae(y,y_hat),label = "MAE loss")
dlt = 1.0
plt.plot(rs,huber(y,y_hat,dlt),label="Huber loss, $\delta=$"\
+str(f"{dlt:.1f}"))
plt.plot(rs,logcosh(y,y_hat),label = "logcosh loss")

plt.xlabel('Residual')
plt.ylabel('Loss (Residual)')
plt.title('Comparison of loss functions')
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.show()

Figure 10.4. Four typical loss functions based on residuals.

The key features mentioned earlier are evident in Fig. 10.4. Readers
may play with the code, make changes to the hyperparameter, and have a
different view of these loss functions. For more detailed discussions, readers
are referred to the article entitled 5 Regression Loss Functions All Machine
Learners Should Know (heartbeat.fritz.ai/5-regression-loss-functions-all-
machine-learners-should-know-4fb140e9d4b0.
398 Machine Learning with Python: Theory and Applications

Figure 10.5: Regression result using linear polynomial and noisy data-points.

10.3 Python Codes for Regression

Let us now write a code to perform linear regressions, based on the

formulations given above. We first perform linear regression with a linear
polynomial function using the following code by setting reg type =1. At
the same time, we can examine how sensitive the L2 loss is to the outlier
data-points. The results are plotted in Fig. 10.5.
It is clearly shown that these two outliers have pushed the regression line
significantly downward. Figure 10.5 is generated using the following code.
Readers may play with the code using different feature functions by slightly
modifying the code.

import numpy as np
%matplotlib inline
import matplotlib.pyplot as plt

def datagen1(x,noise=0.3):
# Generate data with features of 1, x (linear)
lx = len(x)

One = np.ones(lx)
np.random.seed(8)
X = np.stack((One, x), axis=-1) # for linear polynomial
y = x + np.random.rand(lx)*noise # y vector,
return X,y
 Loss Functions for Regression 399

def datagen2(x,noise=0.3): # data with features of 1, x, x^2

lx = len(x)
One = np.ones(lx)
np.random.seed(8)
X = np.stack((One, x,x2), axis=-1) # 2nd order polynomial
y = x**2.+ x + np.random.rand(lx)*noise # form y vector
return X, y

def datagen3(x,noise=0.3): # data with features 1,x,sin(x)

lx = len(x)
One = np.ones(lx)
np.random.seed(8)
X = np.stack((One, x,np.sin(x)), axis=-1)
y = 1.+ 0.1*x + np.sin(x) + np.random.rand(lx)*noise
return X, y

xleft,xright,xdelta = 0.0, np.pi, 0.1

x = np.arange(xleft,xright,xdelta)
lx = len(x)
print('Length of x=',lx, x)
noise = 0.3
reg_type = 3 # 1:linear, 2:2nd order polynomial, 3:sin(x)

if reg_type ==1:
X, y = datagen1(x,noise) # Data, linear polynomial
elif reg_type ==2:
X, y = datagen2(x,noise) # Data, 2nd order polynomial
else:
X, y = datagen3(x,noise=0.1) # Data with sin(x)

yout = y.copy()
yout[int(lx/2)],yout[int(lx-2)] = 1.0, 0.8 # 2 outliers
print('y=',y,y.shape)

XTy = np.dot(X.T,y) #compute X.T * y

XTyout = np.dot(X.T,yout)
print('XTy=',XTy,XTy.shape)

XTX = np.matmul(X.T,X) #Compute X.T * X

print('XTX=',XTX,XTX.shape)

def y_hat(X,cr): # function for the predicted regression

# line for given solution cr*
return np.dot(X,cr) # X = [1 x1] ; cr* = [b* w*]^T
400 Machine Learning with Python: Theory and Applications

plt.figure(figsize=(4.5, 2.9),dpi=100)
c = np.linalg.solve(XTX,XTy)
cout = np.linalg.solve(XTX,XTyout)
print(' c*=',c,c.shape)
plt.plot(x,y_hat(X,c),label = "$\hat{y}$ without outliers")
plt.plot(x,y_hat(X,cout), c='r',alpha=0.5,
label = "$\hat{y}$ with outliers")
# compute the predicted lines and plot out.

plt.scatter(x, y, s=15., c='b', alpha=0.8)

plt.scatter(x, yout, s=45., c='r', alpha=0.2)
plt.xlabel('x')
plt.ylabel('$\hat{y}$')
plt.title('Linear Regression')
plt.grid(color='r', linestyle=':', linewidth=0.3)
plt.legend()
plt.show()

Length of x= 32 [0. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.
1.1 1.2 1.3 1.4 1.5 1.6 1.7
1.8 1.9 2. 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3. 3.1]
y= [1.08734294 1.20668748 1.30558878 1.37860578 1.45269118
1.53056542
1.66768936 1.75445282 1.84962356 1.92116609 1.99700663
2.05554596
2.12812864 2.16479564 2.18741794 2.19010416 2.18848111
2.25905033
2.18722504 2.15818019 2.11587827 2.17149642 2.04128198
2.00791829
1.92255747 1.87094914 1.81486486 1.78699438 1.64953431
1.62772427
1.44399321 1.38674914] (32,)
XTy= [58.51029079 93.63832045 39.77958804] (3,)
XTX= [[ 32. 49.6 19.9954796 ]
[ 49.6 104.16 31.42797769]
[ 19.9954796 31.42797769 15.70789511]] (3, 3)
c*= [1.06733004 0.09226871 0.98918366] (3,)
 Loss Functions for Regression 401

Figure 10.6: Linear regression of data-points with a quadratic polynomial.

Figure 10.7: Regression result using 2nd-order polynomial and noisy data-points.

10.3.1 Linear regression using high-order polynomial and

other feature functions

We now use the code given above to do a regression using 2nd-order

polynomials, simply by adding a quadratic term x2 as the 2nd feature (basis)
in the above code. In the code, we just set reg type = 2. The results are
plotted in Fig. 10.7.
Figure 10.8 shows regression results using sine feature (basis) function.
402 Machine Learning with Python: Theory and Applications

Figure 10.8: Regression result using sine feature function and noisy data-points.

It is convenient to use the classes built in Scikit-Learn to perform the

regression (curve fitting).

from sklearn.preprocessing import PolynomialFeatures

from sklearn.pipeline import make_pipeline
from sklearn.linear_model import LinearRegression
%matplotlib inline
import matplotlib.pyplot as plt
import seaborn as sns; sns.set()
import numpy as np

# Create a 6th-degree polynomial model

poly7_model = make_pipeline(PolynomialFeatures(6),
LinearRegression())

rng = np.random.RandomState(1)
x = 10 * rng.rand(50)
y = np.sin(x) + 0.05 * rng.randn(50)
xfit = np.linspace(0, 10, 1000)
poly7_model.fit(x[:, np.newaxis], y)
yfit = poly7_model.predict(xfit[:, np.newaxis])
plt.figure(figsize=(3.5, 3.0),dpi=100)
plt.scatter(x, y, s=15., c='b', alpha=0.8)
plt.plot(xfit, yfit,c='b',label = "Fitted with 6th order\
polynomial", alpha=0.9)
plt.plot(xfit, np.sin(xfit),c='r',label = "Original true sine\
function,"alpha=0.9)
plt.grid(color='r', linestyle=':', linewidth=0.3)
 Loss Functions for Regression 403

plt.legend(bbox_to_anchor=(2.02, 0.5),loc='center right')

plt.show;

Figure 10.9: Linear regression of data-points with a higher-order polynomial using

Sklearn.

It is seen from Fig. 10.9 that using a 6th-order polynomial basis functions,
we can fit a set of nonlinear data created using a sine function and 5% white
noise. Note that such a high-order polynomial fitting can be very dangerous
when it is used for making predictions outside of the dataset. We can print
out the features of a 6th-order polynomial (variable powered with all the
orders):

x1 = np.arange(2,4,4) # one-dimensional, 6 data-points

poly7_data = PolynomialFeatures(7, include_bias=False)
poly7_data.fit_transform(x1[:, None])
array([[ 2., 4., 8., 16., 32., 64., 128.]])

We can see that it grows very fast with the increase of the order of the
polynomial. Let us print out the prediction a little out of the fitted domain.

xfit = np.linspace(0, 11, 1000)

yfit = poly7_model.predict(xfit[:, np.newaxis])
plt.figure(figsize=(4.5, 2.9),dpi=100)
plt.scatter(x, y, s=15., c='b', alpha=0.8)
plt.plot(xfit, yfit,c='b',label = "Fitted with 6th order\
polynomial")
plt.plot(xfit, np.sin(xfit),c='r',label = "Original true\
sine function")
plt.legend();
404 Machine Learning with Python: Theory and Applications

Figure 10.10: Linear regression of data-points with a higher-order polynomial. Note the
danger of extrapolation.

It is clear from Fig. 10.10 that high-order polynomial fitting should be

used with care especially when extrapolation is used.

10.3.2 Linear regression using Gaussian basis functions

We will now use the Scikit-Learn and the code for linear regression linear
regression (https://jakevdp.github.io/PythonDataScienceHandbook/05.06-
linear-regression.html), using Gaussian basis functions as the feature
function.

from sklearn.base import BaseEstimator, TransformerMixin

class GaussianFeatures(BaseEstimator, TransformerMixin):
"""Uniformly spaced Gaussian features for\
one-dimensional input"""

def __init__(self, N, width_factor=2.0):

self.N = N
self.width_factor = width_factor

@staticmethod
def _gauss_basis(x, y, width, axis=None):
arg = (x - y) / width
return np.exp(-0.5 * np.sum(arg ** 2, axis))

def fit(self, X, y=None):

# create N centers spread along the data range
self.centers_ = np.linspace(X.min(), X.max(), self.N)
self.width_=self.width_factor*(self.centers_[1]-self.
centers_[0])
 Loss Functions for Regression 405

return self
def transform(self, X):
return self._gauss_basis(X[:, :, np.newaxis], self.
centers_,self.width_, axis=1)
xfit = np.linspace(0, 10, 1000)
gauss_model = make_pipeline(GaussianFeatures(20),
LinearRegression())
gauss_model.fit(x[:, np.newaxis], y)
yfit = gauss_model.predict(xfit[:, np.newaxis])
plt.figure(figsize=(4.5, 2.9),dpi=100)
plt.scatter(x, y, s=15., c='b', alpha=0.8)
plt.plot(xfit,yfit,c='b',label="Fitted with Gaussian\
basis")
plt.plot(xfit,np.sin(xfit),c='r',label="Original true\
sine function")
plt.legend(bbox_to_anchor=(1.72, 0.5),loc='center right');
#plt.xlim(0, 10)

Figure 10.11: Linear regression of data-points with RBFs.

This is indeed an excellent fit, as shown in Fig. 10.11. Let us print out
the prediction a little out of the fitted domain.

xfit = np.linspace(0, 11, 1000)

yfit = gauss_model.predict(xfit[:, np.newaxis])
plt.figure(figsize=(4.5, 2.9),dpi=100)
plt.scatter(x, y, s=15., c='b', alpha=0.8)
plt.plot(xfit,yfit,c='b',label="Fitted with Gaussian basis")
plt.plot(xfit,np.sin(xfit),c='r',label="Original true sine\
function")
plt.legend()
plt.ylim(-1.5,3)

(-1.5, 3.0)
406 Machine Learning with Python: Theory and Applications

Figure 10.12: Linear regression of data-points with RBFs. Note the danger of
extrapolation.

We observe a similar phenomenon when the Gaussian basis functions

are used. Care is needed for curve fitting using nonlinear feature functions,
especially when extrapolation is used.

10.4 Neural Network Model for Linear Regressions

with Big Datasets

10.4.1 Setting up neural network models

A neural network model can be built to form a systematic procedure to
effectively minimize the loss function that can have a number of forms as
discussed earlier. The loss function is a function of the weights and biases of
the neurons, which are the learning (or training) parameters. The training
is to find the minimizers, that is, a set of the weights and biases at which
the loss function is a minimum. Because our loss functions are differentiable
(or sub-differentiable), we can use a gradient-based minimization technique
to achieve the goal.
We now use the code provided by mxnet-the-straight-dope (https://
github.com/zackchase/mxnet-the-straight-dope/tree/master/chapter02
supervised-learning), under the Apache-2.0 License, to complete the process
of training a linear regression model using the neural network. We shall do
the following:

• Generate a set of synthetic data and the corresponding labels, similar as

what we did in Section 9.5.1.
• The dataset shall be generated using nd.random normal with mean close
to 0 and variance close to 1. Therefore, the covariance of the dataset will
be close to diagonal, and the corresponding features in the dataset will
thus be almost independent.
 Loss Functions for Regression 407

• The labels are generated using the “true” feature and with artificially
added noise, synthesizing the measurement error. The noise generation
uses a random Gaussian distribution with zero mean and a small
percentage (say, 10%; readers may change it) of variance.
We first import necessary libraries for this task.
# The following codes are modified from these at https://
# github.com/zackchase/ mxnet-the-straight-dope/tree/master/
# chapter02_supervised- learning
# Under Apache-2.0 License.

from __future__ import print_function

import mxnet as mx
from mxnet import nd, autograd, gluon
import random
import numpy as np
mx.random.seed(1) # fix seed for repeatable random numbers

# Specify GPU or CPU computation

data_ctx = mx.cpu()
model_ctx = mx.cpu()
num_inputs = 2 # 2D space with two features: x1 and x2
num_outputs = 1 # one prediction y
num_samples = 1000 # Sample number, use 10 for demonstration,
# but 100 or more for actual running.
def real_fn(X): # linear func.,
# w1,w2,b1: 3 learning parameters
return 2*X[:,0]-3.4*X[:,1]+4.2 #2*x1-3.4*x2+4.2

X = nd.random_normal(shape=(num_samples, num_inputs), ctx=data_ctx)

# randomly generate samples for x1 and x2
noise = .1 * nd.random_normal(shape=(num_samples,), ctx=data_ctx)
# 10% noise, zero mean, 1 variance
y = real_fn(X) + noise # Generate labels with noise.

To relate the data with the equations presented earlier, we have

p=num inputs, m=num samples. Matrix X has a dimension of num samples
x num inputs, and y is a one-dimensional array with num samples entries.
As a good practice, we shall always try to understand the dataset as much
as possible. First, to check the data structure, use the following:

print('y=',y[0:10])
print(y.shape)
print(X[:5])
print(X.shape)
408 Machine Learning with Python: Theory and Applications

y=
[ 6.0136046 2.2423 9.592255 3.569859 2.2450843
4.9734535 5.285636 6.4735103 10.942213 8.004087 ]
<NDArray 10 @cpu(0)>
(1000,)

[[ 0.03629482 -0.49024424]
[-0.9501793 0.03751943]
[-0.7298465 -2.0401056 ]
[ 1.4821309 1.040828 ]
[-0.45256865 0.3116043 ]]
<NDArray 5x2 @cpu(0)>
(1000, 2)

Next, let us compute the covariance matrix of X just to see the level of
independence of these two variables.

print('Covariance Matrix of X:\n', np.cov(X.asnumpy().T))

Covariance Matrix of X:
[[ 1.05065641 -0.01452139]
[-0.01452139 1.08663854]]

These two variables are quite independent, which is good news, and a
quality model can likely be built. Further, let us check how far the prediction
is from the label for the 1st sample:

print('y_hat=',2 * X[0, 0] - 3.4 * X[0, 1] + 4.2,

'Label y_1=', y[0])

y hat=
[5.9394197]
<NDArray 1 @cpu(0)> Label y 1=
[6.0136046]
<NDArray 1 @cpu(0)>

These values are quite close. We can hope for quite a fast training.
Finally, let us use matplotlib to plot all the data-points for easy viewing.
In order to use matplotlib, we need to convert data arrays in NDArray to
NumPy arrays using the .asnumpy() function.
 Loss Functions for Regression 409

%matplotlib inline
import matplotlib.pyplot as plt
plt.scatter(X[:, 0].asnumpy(),y.asnumpy(),color='r')
plt.scatter(X[:, 1].asnumpy(),y.asnumpy(),color='b')
plt.show()

Figure 10.13: Computer-generated data-points for linear regression study using NNs.

These data-points for two different features are quite distinct. Our model
shall be able to learn these features.

10.4.2 Create data iterators

For fast computation, the data are packaged in ways for easy operation in
bulk. In MXNet, data iterators provide utilities for fetching and manip-
ulating data. In our case, we load training data pair X and y into an
ArrayDataset, by calling gluon.data.ArrayDataset(X, y). X can be a
multi-dimensional input array (say, of images) and y can be just a one-
dimensional array. They shall have equal lengths along the first axis, i.e.,
len(X) == len(y).
Given an ArrayDataset, we create a DataLoader which grabs random
batches of data from the ArrayDataset. We need decide the batch size for
group training samples, which specify how many examples we want to use at
a time (epoch). We shall specify whether or not to shuffle over the dataset
between the iterations.
batch_size = 4
train_data = gluon.data.DataLoader(gluon.data.
ArrayDataset(X, y), batch_size=batch_size, shuffle=True)
410 Machine Learning with Python: Theory and Applications

We have now train data initialized using DataLoader. It becomes an

iterable object, and can be easily fetched in batches, as if it were a Python
list. To unpack the data, we simply use the following loop to iterate:

for i, (data, label) in enumerate(train_data):

# Get samples after shuffling the data pool.
print(data, label)
break

[[-0.7022906 0.6857335 ]
[-0.02577175 0.43850085]
[-0.38564655 0.6393674 ]
[-0.6317411 -1.3817437 ]]
<NDArray 4x2 @cpu(0)>
[0.22829296 2.823702 1.2765092 7.6593714 ]
<NDArray 4 @cpu(0)>

When we run the same code again, we shall get a different set of 4 samples,
because we have set shuffle=True.

for i, (data, label) in enumerate(train_data):

print(data, label)
break

[[-2.206704 2.3855615 ]
[ 0.7445208 -0.2434989 ]
[-1.0826081 -1.027927 ]
[ 0.09882921 -0.86426044]]
<NDArray 4x2 @cpu(0)>
[-8.413907 6.1686425 5.5676003 7.5225515]
<NDArray 4 @cpu(0)>

Let us see how many batches we have, and then ensure that makes sense.

for i, (data, label) in enumerate(train_data):

pass
print(i+1)

250
 Loss Functions for Regression 411

10.4.3 Training parameters

First, initialize the training parameters for both the weights and bias.
Usually, this is done by assigning some random numbers.

w = nd.random_normal(shape=(num_inputs, num_outputs),
ctx=model_ctx)
b = nd.random_normal(shape=num_outputs, ctx=model_ctx)
params = [w, b]
print(params)
print('w=',w, '\n b=',b)

[
[[-2.3237102]
[-1.109485 ]]
<NDArray 2x1 @cpu(0)>,
[-0.48563406]
<NDArray 1 @cpu(0)>]
w=
[[-2.3237102]
[-1.109485 ]]
<NDArray 2x1 @cpu(0)>
b=
[-0.48563406]
<NDArray 1 @cpu(0)>

Allocate some memory for the gradient of loss functions with respect to
all learning parameters.

for param in params:

param.attach_grad()
print(param.grad)

[0.]
<NDArray 1 @cpu(0)>

It is seen that only the location is assigned.

412 Machine Learning with Python: Theory and Applications

10.4.4 Define the neural network

Following the formulas given in the formulation section, the net can be
defined simply as follows:

def net(X):
return mx.nd.dot(X, w) + b

10.4.5 Define the loss function

We use the L2 loss function that is the squared distance between the
prediction and the label value.

def square_loss(yhat, y):

return nd.mean((yhat - y) ** 2)

10.4.6 Use of optimizer

Let us use the stochastic gradient descent (SGD) to train the net. This
is done in mini batches. At each iteration, a batch of training samples is
randomly drawn from the dataset. We shall then compute the gradient of
the loss function with respect to all the training parameters: the weights
and biases. The gradient provides the direction for updating the training
parameters, and a learning rate lr, usually a small number, will be specified
to determine how fast the parameter should be updated at each step along
that direction.

def SGD(params, lr): # lr: the learning rate determine

# how fast the parameters are updated
for param in params:
param[:] = param - lr * param.grad

10.4.7 Execute the training

The training is done in stages. Each stage is called an epoch, that is, the
number of drawings of the data batches over the entire dataset. Then, for
each pass, we iterate through the train data, grabbing batches of examples
and their corresponding labels.
 Loss Functions for Regression 413

epochs = 10
learning_rate = .01
num_batches = num_samples/batch_size

for e in range(epochs):
cumulative_loss = 0
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(model_ctx)
label = label.as_in_context(model_ctx).reshape((-1, 1))
with autograd.record(): # record the path for gradient
output = net(data)
loss = square_loss(output, label)
loss.backward() # compute gradient via backward prop.
SGD(params, learning_rate) # perform optimization
cumulative_loss += loss.asscalar() # update the loss.
print(cumulative_loss / num_batches)

4.694164189037867
0.010493556420027745
0.010353110232390463
0.01032697731954977
0.010341777712805197
0.010346036787843332
0.010346328813116997
0.010331496173050254
0.01031649975432083
0.010330414514523

10.4.8 Examining training progress

# Re-initialize parameters because they were already

# trained earlier
w[:] = nd.random_normal(shape=(num_inputs, num_outputs),
ctx=model_ctx)
b[:] = nd.random_normal(shape=num_outputs, ctx=model_ctx)
# Function to plot the losses over epoch
def plot(losses, X, sample_size=100):
xs = list(range(len(losses)))
f, (fg1, fg2) = plt.subplots(1, 2)
fg1.set_title('Loss during training')
fg1.plot(xs, losses, '-r')
fg2.set_title('Estimated vs real function')
414 Machine Learning with Python: Theory and Applications

fg2.plot(X[:sample_size, 1].asnumpy(),
net(X[:sample_size,:]).asnumpy(),'or', label='Estimated')
fg2.plot(X[:sample_size, 1].asnumpy(),
real_fn(X[:sample_size,:]).asnumpy(),'*g', label='Real')
fg2.legend()
plt.show()

learning_rate = .01
losses = []
plot(losses, X)

for e in range(epochs):
cumulative_loss = 0
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(model_ctx)
label=label.as_in_context(model_ctx).reshape((-1, 1))
with autograd.record():
output = net(data)
loss = square_loss(output, label)
loss.backward()
SGD(params, learning_rate)
cumulative_loss += loss.asscalar()

print("Epoch %s, batch %s. Mean loss: %s" % (e, i,

cumulative_loss/num_batches))
losses.append(cumulative_loss/num_batches)

plot(losses, X)

Figure 10.14: Results from the NN before training.

 Loss Functions for Regression 415

Epoch 0, batch 249. Mean loss: 4.427614943831228

Epoch 1, batch 249. Mean loss: 0.010421961726155131
Epoch 2, batch 249. Mean loss: 0.010315687943017111
Epoch 3, batch 249. Mean loss: 0.01035601595134358
Epoch 4, batch 249. Mean loss: 0.01037310311215697
Epoch 5, batch 249. Mean loss: 0.010365772993071004
Epoch 6, batch 249. Mean loss: 0.01032915258151479
Epoch 7, batch 249. Mean loss: 0.010371889414265752
Epoch 8, batch 249. Mean loss: 0.010340319244773128
Epoch 9, batch 249. Mean loss: 0.010389118599705397

Figure 10.15: Results from the NN after training.

10.5 Neural Network Model for Nonlinear Regression

We now present a case study done by the scikit-learn team [3]. This study
uses a multi-layer perceptron network (MLPRegressor class) to perform non-
linear regression using the Boston housing price dataset, together with the
decision tree regression [4, 5]. A code is provided to perform the regressions
and then plot the partial dependence curves. The plot partial dependence
function returns a PartialDependenceDisplay class object that contains
the necessary attributes, which can be used for plotting without recalculating
the partial dependence.
416 Machine Learning with Python: Theory and Applications

%matplotlib inline
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.datasets import load_boston
from sklearn.neural_network import MLPRegressor
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import make_pipeline
from sklearn.tree import DecisionTreeRegressor
from sklearn.inspection import plot_partial_dependence

10.5.1 Train models on the Boston housing price dataset

We train a decision tree (will not be discussed in this book) and a multi-layer
perceptron on the Boston housing price dataset. One may use help to figure
out these classes build in Sklearn. For example, use “help(MLPRegressor)”
to find out the setting of the neural network, when using the MLPRegressor
class.

boston = load_boston()
X = pd.DataFrame(boston.data, columns=boston.feature_names)
y = boston.target
tree = DecisionTreeRegressor()
mlp = make_pipeline(StandardScaler(),
MLPRegressor(hidden_layer_sizes=(100, 100), tol=1e-2,
max_iter=500, random_state=0)) # 2 hidden layers, Relu.
tree.fit(X, y)
mlp.fit(X, y); # you may remove # to view outputs.

10.5.2 Plotting partial dependence for two features

We plot partial dependence curves for two features “LSTAT” and “RM” for
the decision tree, using plot partial dependence(). The spacing of the grids
of these two plots is defined by ax.

fig, ax = plt.subplots(figsize=(12, 6))

ax.set_title("Decision Tree")
tree_disp=plot_partial_dependence(tree,X,["LSTAT", "RM"],ax=ax)
 Loss Functions for Regression 417

Figure 10.16: Results using the decision tree in Sklearn for the Boston housing price
dataset.

Next, the partial dependence curves are plotted for the MLP.

fig, ax = plt.subplots(figsize=(12, 6))

ax.set_title("Multi-layer Perceptron")
mlp_disp = plot_partial_dependence(mlp, X, ["LSTAT", "RM"], ax=ax,
line_kw={"c": "red"})

Figure 10.17: Results using the MLP in Sklearn for the Boston housing price dataset.
418 Machine Learning with Python: Theory and Applications

10.5.3 Plot curves on top of each other

Finally, we plot these partial dependence curves on top of each other using
the following code:
# Sets this image as the thumbnail for sphinx gallery
# sphinx_gallery_thumbnail_number = 4
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(10, 6))

tree_disp.plot(ax=[ax1,ax2],line_kw={"label":"Decision Tree"})
mlp_disp.plot(ax=[ax1, ax2], line_kw={"label":"Multi-layer
Perceptron","c": "red"})
ax1.legend()
ax2.legend();

Figure 10.18: Comparison of results obtained using the decision tree and MLP in Sklearn
for the Boston housing price dataset.

10.6 On Nonlinear Regressions

Note that the above-discussed regression models are in fact linear models,
even if we use nonlinear feature functions, such as higher-order polynomial,
many other special feature functions, and kernels in SVMs. The term feature
function used in regression is essentially the same as the basis function in
FEM. It is just a different name.
We may also note here that when a regression model uses a nonlinear
feature function, it is sometimes called nonlinear regression. This is different
 Loss Functions for Regression 419

from the FEM types of models, where an FEM model is still called linear,
even if nonlinear basis functions are used. When we say nonlinear FEM
model, it is truly nonlinear because the system matrix will depend on the
unknown field variables, and hence a solution to a nonlinear FEM model is
usually iterative.
In fact, one should not be confused that the regression model itself is a
linear one, even if a nonlinear feature function is used, because the system
matrix X X does not depend on the regression parameter [ŵ], and [ŵ] can
be obtained by a single step with a linear equation solver. The prediction
is assumed to be a linear combination of the features and the regression
parameters, shown in the formulation above and will also be seen explicitly
in the code to be written later. This seems just to be a minor terminology
problem, but it may have conceptual significance. Moreover, there are also
true nonlinear regression models, in which the regression parameters and
some (or all) features cannot be written in linear combination. For such true
nonlinear regression models, the solution should be obtained via iterative or
successive fitting processes.

10.7 Conclusion

We end this chapter with the following remarks:

1. Regression is a process of deriving a function of features from a dataset.
The form of the function is set once the model is defined. The model
parameters (e.g., weights and biases for MLP) are determined when the
model is trained. These parameters now record the features of the dataset.
2. Linear regression models can use nonlinear feature functions to better fit
complicated datasets. The use of nonlinear feature functions is similar to
the use of nonlinear kernels in the SVM.
3. Regression can be performed using various machine learning models.
4. A properly built and trained neural network is capable of producing
prediction functions from big datasets containing complicated label-
feature relations. It is capable of extracting features that are both linear
and nonlinear from datasets.

References

[1] G.R. Liu and G.Y. Zhang, Smoothed Point Interpolation Methods: G Space Theory
and Weakened Weak Forms, World Scientific, Singapore, 2013.
420 Machine Learning with Python: Theory and Applications

[2] G.R. Liu, Mesh Free Methods: Moving Beyond the Finite Element Method, Taylor
and Francis Group, New York, 2010.
[3] P. Fabian, V. Gae, G. Alexandre et al., Scikit-learn: Machine Learning in Python,
Journal of Machine Learning Research, 12(85), 2825–2830, 2011. http://jmlr.org/
papers/v12/pedregosa11a.html.
[4] X. Wu, R.J. Kumar Vipin et al., Top 10 algorithms in data mining, Knowledge and
Information Systems, 14, 1–37, 2007.
[5] W. Pushkar and G.R. Liu, Real-time prediction of projectile penetration to laminates
by training machine learning models with finite element solver as the trainer, Defence
Technology, 17(1), 147–160, 2021. https://www.sciencedirect.com/science/article/
pii/S2214914720303275.
 Chapter 11

Loss Functions and Models for Classification

In many practical applications, we often only need to make a “yes” or “no”

type of decision. Probability prediction is often the best choice for making
such a decision. For example, what is the likelihood that a machinery may
fail in operation in the coming month? What is the probability that a virus
will become a pandemic in a region in the coming season? This type of
problem requires an assessment on probability of the occurrence of an event,
and is a typical classification problem. The simplest classification problem
is the binary classification: there are only two classes, “yes” or “no”, for
decision making. This chapter shall focus on this type of problem, focusing
on necessary methods, formulation, and techniques that can be used for best
prediction for binary labels, using neural networks. The reference materials
for this chapter include the following:

• Lecture series by David S. Rosenberg, Office of the CTO at Bloomberg.

• Hastie, Tibshirani, and Friedman’s book: The Elements of Statistical
Learning.
• James, Witten, Hastie, and Tibshirani’s book: An Introduction to Statis-
tical Learning.
• Shalev-Shwartz and Ben-David’s book, Understanding Machine Learning:
From Theory to Algorithms.
• The mxnet-the-straight-dope (https://github.com/zackchase/mxnet-the-
straight-dope) (under the Apache License 2).

11.1 Prediction Functions

To formulate proper loss functions, we must define a prediction function for

the last layer of a neural network model. Thus, our discussion starts from

421
422 Machine Learning with Python: Theory and Applications

formulation of different types of prediction functions that may be used for

binary classification. The prediction function is also called score function and
the value of the prediction function is called score achieved by the model.

11.1.1 Linear function

The first and simplest prediction function for a given input x is the affine
transformation function discussed in Chapter 5:
p

z= Wi xi + b = x · w + b = x w + b (11.1)
i=1

Or by the xw formulation
p

z= Wi xi = xŵ (11.2)
i=0

in which x0 = 1, W 0 = b, and ŵ includes all the training parameters (both

the weights and bias). Output values z are in general real numbers.
The linear prediction function uses the z value as it is at the output (last)
layer of the neural network (linear):

ŷ = z (11.3)

The values of the linear prediction function are in Y1 , and are real
numbers in (−∞, ∞).

11.1.2 Logistic prediction function

A two-classification problem has two labels 1 or 0 for a given input xi :
positive class labeled with yi = 1 and negative class labeled with yi = 0.
Our problem can be formulated to predict values of probability alike for an
event against these labels. Therefore, it makes good sense to use the following
logistic (or sigmoid) activation function σ that maps an arbitrary real output
z to be within (0, 1):
1
ŷ = σ(z) = (11.4)
1 + e−z
It squashes any argument value z in (−∞, ∞) to be in (0, 1). Using the affine
transformation our prediction formula becomes
1 1
ŷ = = (11.5)
1+ e−(xw+b) 1 + e−(xŵ)
 Loss Functions and Models for Classification 423

Because the value of a logistic prediction function falls in the region of (0, 1),
it is now comparable with the given labels.

11.1.3 The tanh prediction function

The positive and negative classes in a two-classification problem may also be
labeled, respectively, with 1 and −1. This type of problem can be re-labeled
to 1 and 0. Alternatively, the hyperbolic tangent (tanh) prediction function
becomes handy. It is defined as follows:
ez − e−z
ŷ = tanh(z) = (11.6)
ez + e−z
where z is given in Eq. (11.1) or (11.2). It is clear that for any given argument
of a real number z in (−∞, ∞), the tanh prediction function returns a real
number within (−1, 1).

11.2 Loss Functions for Classification Problems

Loss function is a measure of the correctness of the prediction against the true
label that is discrete for binary classification problems. Constructing a loss
function requires careful considerations. We first define some terminology.

11.2.1 The margin concept

The margin m(x) is also called margin function. It is often used to define loss
functions for binary classification problems. Let ŷ be the prediction function.
There are two major cases of setting for binary classification problems.
Case I: the label is given as y ∈ {−1, +1}. In this case, the margin m(x)
is defined as
⎧
⎨−ŷ, when y = −1
mg (x) = y ŷ(x) = (11.7)
⎩ŷ, when y = 1

The above definition tells us the following:

• Assuming the label y and a prediction ŷ have the same sign, the prediction
is correct, and the margin mg is positive.
• Assuming the label y and a prediction ŷ have a different sign, the
prediction is incorrect, and the margin mg is negative.
• Therefore, margin mg is a measure of the level of the correctness of a
prediction. We thus want to maximize the margin mg (x), to achieve correct
prediction with a largest possible margin.
424 Machine Learning with Python: Theory and Applications

Case II: the label is given as y ∈ {0, 1}. In this case, the margin m(x) is
defined as ⎧
⎨(1 − ŷ), when y = 0
mg (x) = y ŷ(x) + (1 − y)(1 − ŷ(x)) = (11.8)
⎩ŷ, when y = 1

This definition ensures the following:

• When the label y = 0, ŷ should be close to 0 for the prediction to be
correct, and the margin mg can be positive and maximized for the best
possible prediction.
• When the label y = 1, ŷ should be close to 1 for the prediction to be
correct, and the margin mg can be positive and maximized for the best
possible prediction.
• In either situation, the margin mg becomes a measure of how correct the
prediction is. Maximizing margin mg (x) achieves correct prediction with
the largest possible margin.
Once the margin is properly defined as above according to these
two different label cases, the definition on the loss functions for binary
classification problems can be unified for these two cases. All one needs is to
have a maximized margin by minimizing the loss function. These types of
loss functions are called margin-based loss, in contrast to the distance-based
loss discussed in Chapter 10.
We have seen a similar concept in Chapter 6 when discussing SVM, where
we call the margin the width of the decision street. We went through a
procedure to maximize it, using a nice prime-dual (Lagrangian multipliers)
formulation leading to a standard quadratic programming problem, which
was solved using routinely available packages.

11.2.2 0–1 loss

For a given margin function, the 0–1 loss function can be defined as follows:

L(mg ) = 1(mg ≤ 0) (11.9)

where 1() is the indicator function. Figure 11.1 shows schematically the 0–1
loss.
The 0–1 loss function has the following properties:
1. This function has 1 as its value when the margin is smaller or equal to zero.
It is zero otherwise, where it gives the correct prediction. Minimization
of the loss is to maximize the margin.
 Loss Functions and Models for Classification 425

Figure 11.1: 0–1 loss function for classification.

2. It is clearly discontinuous and not differentiable. Therefore, it is difficult

to be used with a gradient-based minimization algorithm.
3. It is also not convex, and hence there is no guarantee that a minimum
can be found, even if a gradient-based minimization algorithm is workable
for it.

11.2.3 Hinge loss

For a given margin function, the hinge loss function has the following form:

L(mg ) = max(1 − mg , 0) (11.10)

Figure 11.2 shows schematically the hinge loss together with the 0–1 loss.

Figure 11.2: Hinge loss and 0–1 loss functions for classification.
426 Machine Learning with Python: Theory and Applications

The hinge loss function has the following properties:

1. When the margin is smaller then 1, it is a linear function with a slope

of −1 and passing through [0, 1]. When the margin is larger than 1, it is
zero. We see again that minimizing the loss maximizes the margin.
2. It is convex, but not strictly and has multiple minima.
3. It is an upper bound of the 0–1 loss.
4. Not differentiable at mg = 1.
5. It has a margin error when 0 < mg < 1: it has a loss even if it gives a
correct prediction there.

The hinge loss can be used to train an SVM model using a gradient-based
minimization algorithm.

11.2.4 Logistic loss

For a given margin function, the logistic loss function has the following form:

L(mg ) = log(1 + e−mg ) (11.11)

Figure 11.3 shows schematically the logistic loss together with the hinge loss
and the 0–1 loss.
The logistic loss function has the following properties:

1. It is smooth, differentiable, and convex.

2. It never goes to zero (minimizer always wants more margin).
3. When the margin is negatively small, its slope is close to the hinge loss.

Figure 11.3: Logistic, Hinge, and 0–1 loss functions for classification.
 Loss Functions and Models for Classification 427

Figure 11.4: Exponential, Logistic, Hinge, and 0–1 loss functions for classification.

11.2.5 Exponential loss

For a given margin function, the exponential loss function has the following
form:
L(mg ) = exp(−mg ) (11.12)

Figure 11.4 shows schematically the exponential loss together with the hinge
loss and the 0–1 loss.
The exponential loss function has the following properties:

1. It is smooth, differentiable, and convex.

2. It never goes to zero (minimizer always wants more margin).
3. It is an upper bound of the 0–1 loss.
4. It penalizes the negative margins very heavily.

11.2.6 Square loss

Square loss function uses the L2 loss and it can be also written as a function
of margin:

L(ŷ, y) = (ŷ − y)2 = (1 − ŷy)2 = (1 − mg )2 = L(mg ) (11.13)

This is because y 2 = 1 for the label y ∈ {−1, +1}.

The properties of the square loss function have been discussed in detail
in Chapter 10:

1. It is smooth, differentiable, convex, and mg = 1 is the minimizer of the

loss function.
428 Machine Learning with Python: Theory and Applications

Figure 11.5: A neural network for binary classification.

2. It is sensitive to outliers.
3. It may have higher complexity and hence need for samples compared to
the hinge and logistic lasses [1].

Figure 11.5 plots all these four together for examination.

We write a code to compute these loss functions and plot figures in the
above.

# Loss functions for classification

# the 0-1 loss
def zerone(mg):
return np.where(mg <= 0 , 1.0,0.)

# the hinge loss

def hinge(mg):
return np.where(mg <= 1. , 1.-mg,0.) #np.max(1-mg,0)

# the logloss
def logloss(mg):
return np.log(1.+np.exp(-mg))

# the exploss, exponential loss

def exploss(mg):
return np.exp(-mg)
 Loss Functions and Models for Classification 429

# the squared error loss

def squared(mg):
return (1-mg)**2

%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
from matplotlib import rc

xleft,xright,xdelta =-3.0,3.01, .01

mg = np.arange(xleft,xright,xdelta) # Residuals
lxd = len(mg)
#print('Length of mg=',lxd, mg)

plt.figure(figsize=(4.5, 2.8),dpi=100)
plt.plot(mg,zerone(mg),label = "0-1 loss") #+str(f"{L:.2f}"))
plt.plot(mg,hinge(mg),label = "Hinge loss")
plt.plot(mg,logloss(mg),label = "Logistic loss")
plt.plot(mg,exploss(mg),label = "exponential loss")
plt.plot(mg,squared(mg),label = "Squared loss")
dlt = 1.0
plt.xlabel('Margin, $m_g=y\hat{y}$')
plt.ylabel('Loss ($m_g$)')
plt.title('Loss functions for classification')
plt.grid(color='r', linestyle=':', linewidth=0.2)
plt.legend()
plt.ylim(-0.2, 5)
plt.show()

The code above produces the previous figure.

11.2.7 Binary cross-entropy loss

Consider a given dataset of m samples, xs , (s = 1, 2, . . . , m). We now want to

evaluate the overall probability of labels being true, given these m samples,
using a neural network with training parameter ŵ. This requires computing
the following probability:

Pŵ ((y1 , . . . , ym )|x1 , . . . , xm ) (11.14)

430 Machine Learning with Python: Theory and Applications

If each sample is independent of the rest, and each label is only for its own
corresponding sample, these m events are independent. Thus, the overall
probability is the product of all the probabilities for each sample. Equation
(11.14) becomes

Pŵ (y1 |x1 )Pŵ (y2 |x2 ) . . . Pŵ (ym |xm ) (11.15)

This is a product of multiple probabilities, which is not very convenient to

use directly as a loss function. In addition, the probabilities are all smaller
or equal to 1.0, whose products can be a very small number, if m is a very
large number. Taking a log for the product of the probabilities gives

log Pŵ (y1 |x1 ) + · · · + log Pŵ (ym |xm ) (11.16)

This is preferred, and we shall use it instead. Because log function is a

monotonically increasing function with an increasing positive argument, it
does not change the trend of the original function, and hence will not alter the
locations of the extremes of the original function that fed to the log function.
This can be observed using the following codes that produces Fig. 11.6:

x = np.arange(0.001, 10, .1)

y = np.log(x)
plt.plot(x,y)
plt.xlabel('x')
plt.ylabel('$log(x) $')
plt.title('Log Function: Monotonic!')
plt.show()
print('log(0.5)=',np.log(0.5),'log(1.0)=',np.log(1.0))

Figure 11.6: The curve of a logarithmic function. It varies monotonically.

 Loss Functions and Models for Classification 431

log(0.5)= -0.6931471805599453 log(1.0)= 0.0

Now, our goal is to find the prediction yˆi as close as possible to the labels
yi , but needs to be in terms of probability. Also, we already know that yˆi
will be a value between 0 and 1, when using the sigmoid logistic function to
squash any input given. Moreover, because a label yi is given as either 1 or 0
for a binary classification problem, the prediction ŷi shall have the following
correspondence to label yi :

• When yi belongs to the positive class (i.e., yi = 1), the probability for
yi being observed for the given ith sample should be set as yˆi . This is
because when yˆi is maximized, over the training parameters ŵ, we obtain
the correct correspondence.
• On the other hand, when yi belongs to the negative class (i.e., yi = 0), the
probability for yi being observed for the given ith sample should be set as
1 − yˆi . This is because when yˆi is maximized, over the training parameters
ŵ, 1 − yˆi is minimized to approach 0. We, again, obtained the correct
correspondence.

The above setting nicely covers both possible situations for our binary
classification problem. Hence, for the ith sample, this setting can be
expressed in the following formula:

ŷi , if yi = 1
Pŵ (yi |xi ) = (11.17)
1 − ŷi , if yi = 0

Using a mathematical trick, Eq. (11.17) can be rewritten in a single piece of

Pŵ (yi |xi ) = ŷiyi (1 − ŷi )1−yi (11.18)

Finally, take a log for the probability above for all samples; our loss function
can be expressed as
m

L(y, ŷ) = − (yi log ŷi + (1 − yi ) log(1 − ŷi )), (11.19)
i=1

where the minus sign is added, because the loss function will be minimized
rather than optimized in the training process. Note also that log function
is again used to make the power function easier to deal with in gradient
432 Machine Learning with Python: Theory and Applications

computations. It is clear now that the above loss function is exactly the
measure of the binary cross-entropy of the distribution of the label with
respect to that of prediction, which are given in the last section of Chapter 4.
This loss function is called “log loss” or “binary cross-entropy”, or “negative
log likelihood”. It is a special case of cross-entropy, which can used for multi-
classification problems.

11.2.8 Remarks

Note that the loss functions discussed above are basic ones. There are
also problem-specified loss functions. For example, when we deal with
k-classification problems, cross-entropy loss may be a better choice (see the
next chapter). Also, when designing a facial recognition neural network, we
may use a special loss function called triplet loss.

11.3 A Simple Neural Network for Classification

With all the concepts set properly, let us build a feed forward net to perform
binary classifications. The neural network is schematically shown in Fig. 11.7.
This simple neural network has just one input and one output layer. It
uses a combination of affine mapping, logistic prediction function, and binary
cross-entropy loss. One may choose other types of combinations, depending
on the application problem. The training of the net requires a backward
process to update the learning parameters of weights and bias.

Figure 11.7: A simple neural network with just one input and one output layer for binary
classification: affine mapping, logistic prediction function, and binary cross-entropy loss.
 Loss Functions and Models for Classification 433

11.4 Example of Binary Classification Using Neural

Network with mxnet

11.4.1 Dataset for binary classification

We demonstrate here an example of binary classification provided by mxnet-

the-straight-dope (https://github.com/zackchase/mxnet-the-straight-dope)
(under the Apache License 2). This time, we use a real dataset taken from the
UCI repository (http://archive.ics.uci.edu/ml/datasets/). It was constructed
by Barry Becker in 1994, containing 14 features, including age, education,
occupation, sex, and native country. In the version hosted by National
Taiwan University (https://www.csie.ntu.edu.tw/%7Ecjlin/libsvmtools/
datasets/binary.html), the data have been re-processed to have 123 features.
The label is a binary indicator on whether the person corresponding to each
row made more (yi = 1) or less (yi = 0) than $50,000 income in the year of
1994. The dataset contains 30,956 training and 1,605 testing samples.
We first read the following in the dataset:

import numpy as np
import mxnet as mx
from mxnet import nd, autograd, gluon
import matplotlib.pyplot as plt

data_ctx = mx.cpu()
# Change this to mx.gpu(0) to train on an NVIDIA GPU
model_ctx = mx.cpu()

with open("datasets/a1a.train") as f:
train_raw = f.read()

with open("datasets/a1a.test") as f:
test_raw = f.read()

print(len(test_raw))
print(test_raw[0:216])

114816
-1 3:1 11:1 14:1 19:1 39:1 42:1 55:1 64:1 67:1 73:1 75:1 76:1 80:1 83:1
-1 3:1 6:1 17:1 27:1 35:1 40:1 57:1 63:1 69:1 73:1 74:1 76:1 81:1 103:1
-1 4:1 6:1 15:1 21:1 35:1 40:1 57:1 63:1 67:1 73:1 74:1 77:1 80:1 83:1
434 Machine Learning with Python: Theory and Applications

The first three lines of data are printed above. The first entry in each
row is the label value. The following tokens are the indexes of the non-zero
features. The number “1” is redundant. Let use the following code to process
the dataset.

def process_data(raw_data):
train_lines = raw_data.splitlines()
num_examples = len(train_lines)
num_features = 123
X = nd.zeros((num_examples, num_features), ctx=data_ctx)
Y = nd.zeros((num_examples, 1), ctx=data_ctx)
for i, line in enumerate(train_lines):
tokens = line.split()
label(int(tokens[0])+1)/2 #Change label:{-1,1} to {0,1}
Y[i] = label
for token in tokens[1:]:
index = int(token[:-2]) - 1
X[i, index] = 1
return X, Y

Verify the data by checking the shape of the arrays.

Xtrain, Ytrain = process_data(train_raw)

Xtest, Ytest = process_data(test_raw)

print('Xtrain:',Xtrain.shape)
print('Ytrain:',Ytrain.shape)
num_sample = len(Xtrain[0:])
print('num_sample=',num_sample)
print('Xtest:',Xtest.shape)
print('Ytest:',Ytest.shape)

Xtrain: (30956, 123)

Ytrain: (30956, 1)
num sample= 30956
Xtest: (1605, 123)
Ytest: (1605, 1)

It is confirmed that the dataset X has a total of m = 30,956 train samples,

each of which has p = 123 figures. Corresponding to each training sample is
a label y, with a total of 30,956 labels. The dataset for testing has a total of
1605 samples, each of which also has p = 123 figures. Corresponding to each
test sample is a label y with a total of 1,605 labels.
 Loss Functions and Models for Classification 435

The data value of each sample is given in either 1.0 or 0.0, corresponding
to “yes” or “no”. The same is true also for these labels.
Check the ratio of positive samples in the training and test datasets. This
shall give us an indication of the overall distribution of the training and test
data.

print(nd.sum(Ytrain)/len(Ytrain))
print(nd.sum(Ytest)/len(Ytest))
[0.24053495]
<NDArray 1 @cpu(0)>

[0.24610592]
<NDArray 1 @cpu(0)>

Let us instantiate a dataloader

batch_size = 64
train_data = gluon.data.DataLoader(gluon.data.ArrayDataset
(Xtrain, Ytrain), batch_size=batch_size, shuffle=True)
test_data = gluon.data.DataLoader(gluon.data.ArrayDataset
(Xtest, Ytest), batch_size=batch_size, shuffle=True)

11.4.2 Define loss functions

We write codes for loss functions discussed earlier for use and examinations.
Readers may try out these function by simply changing the function name
in the training. The trial outcomes by the author done earlier are given in
the comment lines.

def logistic(z): # Sigmoid function

return 1. / (1. + nd.exp(-z))

def cross_entropy(output,y): # This will not work well, try

yhat = logistic(output) # for k-classification
return - nd.nansum(y * nd.log(yhat+1e-6))

def log_loss(output, y): # readers may try different loss

yhat = logistic(output)
#return nd.mean((1-y*yhat)**2) # fast, not accurate: 0.241
#return nd.sum(nd.maximum(1.-(y*yhat+(1-y)*(1-yhat)),
# nd.zeros_like(y))) works: 0.837
436 Machine Learning with Python: Theory and Applications

return nd.nansum((1-y*yhat-(1-y)*(1-yhat))**2)
#work best: 0.849.

#return - nd.nansum( y * nd.log(yhat) + (1-y)

#* nd.log(1-yhat)) standard: 0.848

import numpy as np

# Define and initialize the neural network model

# When re-train the model from fresh, start from here,
# so that the learning parameters can be reset in a random.
net = gluon.nn.Dense(1)
net.collect_params().initialize(mx.init.Normal(sigma=1.),
ctx=model_ctx)

# Instantiate an optimizer
# Choices on optimizer and the learning rate are made here.

trainer = gluon.Trainer(net.collect_params(),'sgd',
{'learning_rate':0.01})

# Estimate the maximum epochs

max_epochs = int(num_sample/batch_size + 1)
# Decide the number of epochs, or use the max_epochs
epochs = 100 # max_epochs
print('num_sample=',num_sample,'epochs=',epochs,'max_epochs='
,max_epochs,' batch_size=',batch_size)
print('The initial loss in the order of',
int(-np.log(0.5)*num_sample))
loss_sequence = []
num_examples = len(Xtrain)

for e in range(epochs):
cumulative_loss = 0
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(model_ctx)
label = label.as_in_context(model_ctx)
with autograd.record():
output = net(data) # output=z, net=x.w+b
loss = log_loss(output, label)

loss.backward()
trainer.step(batch_size)
cumulative_loss += nd.sum(loss).asscalar()
 Loss Functions and Models for Classification 437

if e%20 == 0 or e == epochs-1:
print("Epoch %s, loss: %s" % (e, cumulative_loss ))
loss_sequence.append(cumulative_loss)
num sample= 30956 epochs= 100 max epochs= 484 batch size= 64
The initial loss in the order of 21457
Epoch 0, loss: 7092.165759086609

11.4.3 Plot the convergence curve of the loss function

# plot the convergence of the estimated loss function

%matplotlib inline
import matplotlib
import matplotlib.pyplot as plt
plt.figure(num=None,figsize=(8, 6))
plt.plot(loss_sequence)
plt.grid(True, which="both")
plt.xlabel('epoch',fontsize=14)
plt.ylabel('average loss',fontsize=14)

11.4.4 Computing the accuracy of the trained model

The convergence curve shows the process of the learning parameters getting
improvements during the training. It is not exactly the measure of the
accuracy of the model trained. To assess how good the trained model is,
we must use an independent set of the test data. This can be done using the
following code. It is also a good practice to also check how good the model
performs using the training data.

num_correct = 0.0
num_total = len(Xtrain)
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(model_ctx)
label = label.as_in_context(model_ctx)
output = net(data)
prediction = (nd.sign(output) + 1) / 2
num_correct += nd.sum(prediction == label)

print("Accuracy: %0.3f (%s/%s)" % (num_correct.asscalar()/

num_total,num_correct.asscalar(),num_total))
438 Machine Learning with Python: Theory and Applications

The accuracy looks good. It should be good, because we have used the
data to train the net. Therefore, this is only an indication that the training
itself is well done. Let us finally conduct a more objective assessment, using
the test data, which gives the indication of the accuracy and usefulness of
the trained model.

num_correct = 0.0
num_total = len(Xtest)
for i, (data, label) in enumerate(test_data):
data = data.as_in_context(model_ctx)
label = label.as_in_context(model_ctx)
output = net(data)
prediction = (nd.sign(output) + 1) / 2
num_correct += nd.sum(prediction == label)

print("Accuracy: %0.3f (%s/%s)" % (num_correct.asscalar()/

num_total, num_correct.asscalar(), num_total))

We now see the accuracy of the trained model after the training is done
for the specified epochs, using both the training and the test datasets. We
have also run the same code using 483 epochs (the max epochs), which allows
the model to see almost all the training samples, and the accuracy can reach
about 84% for the test dataset.
This model is simple, but it shall give one good experience in training
neural networks for similar types of problems.

11.5 Example of Binary Classification Using Sklearn

We now introduce a case study done by the scikit-learn team [2]. It uses
a total of 10 classifiers in scikit-learn on three synthetic datasets, and plots
decision boundaries together for easy comparison.

%matplotlib inline
print(__doc__)
# Code source: Gaël Varoquaux
# Andreas Müller
# Modified for documentation by Jaques Grobler
# License: BSD 3 clause
 Loss Functions and Models for Classification 439

import numpy as np
import matplotlib.pyplot as plt
from matplotlib.colors import ListedColormap
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.datasets import make_moons,
make_circles, make_classification
from sklearn.neural_network import MLPClassifier
from sklearn.neighbors import KNeighborsClassifier
from sklearn.svm import SVC
from sklearn.gaussian_process import
GaussianProcessClassifier
from sklearn.gaussian_process.kernels import RBF
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import
RandomForestClassifier, AdaBoostClassifier
from sklearn.naive_bayes import GaussianNB
from sklearn.discriminant_analysis import
QuadraticDiscriminantAnalysis

h = .02 # step size in the mesh

names = ["Nearest Neighbors", "Linear SVM", "RBF SVM",
"Gaussian Process", "Decision Tree", "Random Forest",
"Neural Net", "AdaBoost", "Naive Bayes", "QDA"]
classifiers = [
KNeighborsClassifier(3),
SVC(kernel="linear", C=0.025),
SVC(gamma=2, C=1),
GaussianProcessClassifier(1.0 * RBF(1.0)),
DecisionTreeClassifier(max_depth=5),
RandomForestClassifier(max_depth=5, n_estimators=10,
max_features=1),
MLPClassifier(alpha=1, max_iter=1000),
AdaBoostClassifier(),
GaussianNB(),
QuadraticDiscriminantAnalysis()]

X, y = make_classification(n_features=2, n_redundant=0,
n_informative=2,random_state=1,n_clusters_per_class=1)
rng = np.random.RandomState(2)
X += 2 * rng.uniform(size=X.shape)
linearly_separable = (X, y)
440 Machine Learning with Python: Theory and Applications

datasets = [make_moons(noise=0.3, random_state=0),

make_circles(noise=.2, factor=.5, random_state=1),
linearly_separable
]

figure = plt.figure(figsize=(27, 9))

i = 1
# iterate over datasets
for ds_cnt, ds in enumerate(datasets):
# preprocess dataset, split into training and test part
X, y = ds
X = StandardScaler().fit_transform(X)
X_train, X_test, y_train, y_test = \
train_test_split(X, y, test_size=.4, random_state=42)

x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5

y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5
xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
np.arange(y_min, y_max, h))

# just plot the dataset first

cm = plt.cm.RdBu
cm_bright = ListedColormap(['#FF0000', '#0000FF'])
ax = plt.subplot(len(datasets), len(classifiers) + 1, i)
if ds_cnt == 0:
ax.set_title("Input data")
# Plot the training points
ax.scatter(X_train[:, 0], X_train[:, 1], c=y_train,
cmap=cm_bright,edgecolors='k')
# Plot the testing points
ax.scatter(X_test[:,0],X_test[:,1],c=y_test, cmap=cm_bright,
alpha=0.6,edgecolors='k')
ax.set_xlim(xx.min(), xx.max())
ax.set_ylim(yy.min(),yy.max())
ax.set_xticks(())
ax.set_yticks(())
i += 1

# iterate over classifiers

for name, clf in zip(names, classifiers):
ax=plt.subplot(len(datasets),len(classifiers)+1,i)
clf.fit(X_train, y_train)
score = clf.score(X_test, y_test)
 Loss Functions and Models for Classification 441

# Plot the decision boundary. We assign a color to each

# point in the mesh [x_min, x_max]x[y_min, y_max].
if hasattr(clf, "decision_function"):
Z = clf.decision_function(np.c_[xx.ravel(),
yy.ravel()])
else:
Z = clf.predict_proba(np.c_[xx.ravel(),1]

# Put the result into a color plot

Z = Z.reshape(xx.shape)
ax.contour f(xx, yy, Z, cmap=cm, alpha=.8)

# Plot the training points

ax.scatter(X_train[:,0],X_train[:,1],c=y_train,
cmap=cm_bright,edgecolors='k')

# Plot the testing points

ax.scatter(X_test[:,0],X_test[:,1],c=y_test,cmap=cm_bright,
edgecolors='k',alpha=0.6)

ax.set_xlim(xx.min(), xx.max())
ax.set_ylim(yy.min(), yy.max())
ax.set_xticks(())
ax.set_yticks(())
if ds_cnt == 0:
ax.set_title(name)
ax.text(xx.max()-.3,yy.min()+.3,('%.2f'
%score).lstrip('0'),
size=15,horizontalalignment='right')
i += 1

plt.tight_layout()
plt.show()

Automatically created module for IPython interactive

environment
442 Machine Learning with Python: Theory and Applications

Figure 11.8: Comparison of the classification results using various classifiers. A study by
Sklearn team.
 Loss Functions and Models for Classification 443

Three randomly generated datasets are used in this study, and the
results are plotted in three row in Fig. 11.8. The methods used are
Nearest neighbors, Linear SVM, RBF SVM, Gaussian Process, Decision
Tree, Random Forest, MLP Neural Network, Ada Boost, Gaussian Naive
Bayes, and Quadratic Discriminant Analysis (QDA). The training data-
points are in solid colors, and the testing data-points are semi-transparent.
The number in the lower-right corner is the accuracy of the classification on
the test datasets. Although we did not discuss some of these classification
methods, the decision boundaries obtained by these different methods give
some intuitive indication of how these methods work. Readers may read the
documentation of Sklearn for more details.

11.6 Regression with Decision Tree, AdaBoost,

and Gradient Boosting

Decision tree, AdaBoost, and Gradient boosting can also be applied to

regression problems. Scikit Learn has provided a number of examples that
are quite easy to follow. Readers may take a look at these examples:
• Decision Tree Regression.
• Decision Tree Regression with AdaBoost [3].
• Gradient boosting regression.
• A study on real-time prediction of projectile penetration to laminates,
using both neural network and decision tree [4].

References

[1] R. Lorenzo, De Vito Ernesto, C. Andrea et al., Are loss functions all the same?,
Neural Computation, 16(5), 1063–1076, May 2004. https://doi.org/10.1162/0899766
04773135104.
[2] P. Fabian, V. Gae, G. Alexandre et al., Scikit-learn: Machine learning in Python,
Journal of Machine Learning Research, 12(85), 2825–2830, 2011. http://jmlr.org/
papers/v12/pedregosa11a.html.
[3] H. Drucker, Improving regressors using boosting techniques, Proc. 14th International
Conference on Machine Learning, 1997.
[4] W. Pushkar and G.R. Liu, Real-time prediction of projectile penetration to laminates
by training machine learning models with finite element solver as the trainer, Defence
Technology, 17(1), 147–160, 2021. https://www.sciencedirect.com/science/article/
pii/S2214914720303275.
MACHINE LEARNING
WITH PYTHON
 Chapter 12

Multiclass Classification

In the previous chapter, we discussed neural networks for binary

classification, useful for “yes” or “no” types of problems. It is also a basis
for handling many types of multi-classification problems, because one can
often change a multi-classification problem into a set of binary classification
problems. The techniques include one-to-rest, all-pairs, and their variation,
as we have discussed in the Chapter 6 on SVM. There are many well-
established modules for this type of purpose, such as Scikit learn. In
this chapter, we discuss a systematic means to handle multi-classification
problems, also known as k -classification problems, where we have in
general a dataset that has k classes.

12.1 Softmax Activation Neural Networks for

k-Classifications

For binary logistic regression problems, we often use the following activation
function for the neuron at the final layer to predict the output:
1 1 1
ŷ = σ(z) = −z
= = (12.1)
1+e 1+e−(x·w+b) 1 + e−(x·ŵ)
where σ stands for the sigmoid function. It squashes an arbitrarily real
number into a number between 0 and 1, which becomes a type of probability
output.
For k-classification problems, we create a neural network with k neurons
at the output layer. If there are p inputs xi (i = 1, 2, . . . , p), the input layer
shall have p neurons. The structure of the simplest neural network can be
shown in Fig. 12.1.

445
446 Machine Learning with Python: Theory and Applications

Figure 12.1: A simple one-layer network for k-classifications.

In order to have a probability type of output, we would need to use a

special activation function (instead of the sigmoid) to force the output layer
to produce a set (not just one) of probabilities for all these k classes. This
is the well-known softmax function defined as

ez
softmax(z) = k (12.2)
zj
j=1 e

where z can be a vector with k components zj (i = 1, 2, . . . , k), and the

computation of the exponentiates is all element-wise. In matrix form, z can
be given, for example, in the following form:

z = x W+ b = x Ŵ (12.3)
1×k 1×p p×k 1×k 1×(p+1) (p+1)×k

The jth component zj can be expressed as

zj = x Ŵ[j] (12.4)

Readers my refer to Chapter 5 for these notations. For image classification

problems, p is the total number of pixels in the image.
It is clear that the softmax function uses ez to convert all values to be
strictly positive followed by a normalization, so that all values can sum to 1.
 Multiclass Classification 447

The outputs at these k neurons on the final layer become a vector of

e(x Ŵ)
ŷ = softmax(x Ŵ) = k (12.5)
(x Ŵ[j])
j=1 e

The softmax activation function is widely used in k-classification, also

known as softmax regression or multinomial regression.

12.2 Cross-Entropy Loss Function for k-Classifications

In a k-classification problem, one of the k labels is 1 (the correct one or

the hot one) and the rest are all 0. Therefore, we can use the following
cross-entropy function (see Chapter 4) as the loss function:
⎛ ⎞
⎜ ⎟
1 ⎜ ⎟
m k m
⎜  ⎟ 1 
L(y, ŷ) = ⎜− yj log ŷj ⎟ = − (ys · log ŷs ) (12.6)
m s=1 ⎜ ⎟ m s=1
⎝ j=1 ⎠
 

ys ·log ŷs

where s stands for the sth sample in the dataset. The above loss function
essentially cares only about the loss related to the correct label. Note that
this cross-entropy loss function is still a scalar function.
Compared to the loss function used in the binary classification problems,
we can expect that the binary-cross-entropy loss function is a special case
of the k-classification loss function. To proof this, we simply let k = 2 in
Eq. (12.6), which gives
⎛ ⎞
m 2
1 ⎝
L(y, ŷ) = − yj log ŷj ⎠
m
s=1 j=1
s (12.7)
m
1
=− (yj=1 log ŷj=1 + yj=2 log ŷj=2)s
m s=1

Notice that index j is for the class, and s stands for the sth sample. Now,
in the case of binary classifications, we must have

yj=2 = 1 − yj=1, and ŷj=2 = 1 − ŷj=1 (12.8)

448 Machine Learning with Python: Theory and Applications

This is true for any sample. Substituting Eq. (12.8) into (12.7), dropping
j = 1 because it is the only one left, and then moving the index s inside, we
have
m
1 
L(y, ŷ) = − (ys log ŷs + (1 − ys ) log(1 − ŷs )), (12.9)
m
s=1

which is exactly the same equation for binary classification, as we seen in

Chapter 11 (except the normalization by m). This proof implies that if one
just uses two labels (k = 1, 2) and arranges the dataset in accordance with
a multi-classifier, one should be able to perform a binary classification.

12.3 Case Study 1: Handwritten Digit Classification

with 1-Layer NN

We are now ready to perform a demonstration of k-classification: to classify

handwritten digits. We have done this earlier using the Naive Bayes
Classifier. Now, we do the same, but using a simply neural network with
a single (input and output, no hidden) layer. We will discuss the detailed
procedure using the codes available at mxnet-the-straight-dope (under the
Apache License 2), with some minor changes. As usual, we import the
necessary libraries to begin with.

# Import necessary libraries

from __future__ import print_function
import numpy as np
import mxnet as mx
from mxnet import nd, autograd, gluon
mx.random.seed(1)

12.3.1 Set contexts according to computer hardware

We can set the compute context, using either model ctx or mx.gpu(0) of
your choice. In this demonstration, we use cpu.

data_ctx = mx.cpu()
model_ctx = mx.cpu() # model_ctx = mx.gpu() # alternative

12.3.2 Loading the MNIST dataset

We again use the real MNIST data for this demonstration. The MNIST
dataset contains a total of 70,000 images of handwritten digits (0–9), each
 Multiclass Classification 449

image is a 28 by 28 centrally cropped black&white photograph. We are to

train the k-classification model to classify handwritten digits in {0−9}.
We first use MXNet’s utility to get a copy of the MNIST dataset and
then preprocess each item in it. The data and label are casted to floats
and normalized to the range of [0, 1]. The 70,000 images are divided, in a
random fashion, into a training dataset with 60,000 and a test dataset with
10,000 images. Both x train and x test parts contain grayscale RGB codes
(from 0 to 255), while y train and y test parts contain labels (from 0 to
9), representing the actual number of the handwriting. To visualize these
numbers, we can use matplotlib.

def transform(data, label):

return data.astype(np.float32)/255, label.astype(np.float32)
# data type conversion & normalization
mnist_train=gluon.data.vision.MNIST(train=True,transform=transform)
#get the data for training
mnist_test=gluon.data.vision.MNIST(train=False,transform=transform)
#get the data for testing

Each item in the train (and test) dataset is a tuple of an image paired
with a label:

print(len(mnist_train))
image,labelmnist_train[100] # unpack the 100th data pair
print(image.shape, label) # check the shape and label
#print(image) # print the image in floats

60000
(28, 28, 1) 5.0

This shows that each image is formatted as a three-tuple (height,

width, channel). The channel here is one dimension, because the image is
black&white. For color images, the channel would have three dimensions
(red, green, and blue).
We next define variables for the numbers of inputs and outputs.

num_inputs = 784 # 28×28=784

num_outputs = 10 # 10 digits
num_examples = 60000 # number of training samples
450 Machine Learning with Python: Theory and Applications

Because matplotlib expects either (height, width) or (height, width,

channel) with RGB channels, we reshape the data by broadcasting the single
channel to 3.

im = mx.nd.tile(image, (1,1,3)) #This should be a 5

print(im.shape)

(28, 28, 3)

We shall now visualize the image to see whether the image tallies with
the label.

import matplotlib.pyplot as plt

plt.imshow(im.asnumpy())
plt.show()
print('label=',label)

Figure 12.2: A sample image of handwritten digit from the MNIST dataset.

label= 5.0

Let us load the training dataset of images in batches (for stochastic

gradient descent optimization) with a batch size specified. This is done in a
random fashion (by setting shuffle to True). We also load the test dataset
but without shuffling.
 Multiclass Classification 451

batch_size = 64
train_data=mx.gluon.data.DataLoader(mnist_train,batch_size,\
shuffle=True)
test_data=mx.gluon.data.DataLoader(mnist_test,batch_size,\
shuffle=False)

12.3.3 Set model parameters

Our model parameters should be as follows:

1. We need to do a 10-classification, thus k = 10.

2. We decided to use 64 training samples in a batch.
3. Because the image data are in a matrix of 28 × 28, we should flatten it
into a 1D row vector with 28 × 28 = 784 components. Therefore, p for
this case is 784.
4. The dimension of any of these samples is a row vector x with 784
dimensions.
5. Because we have 10 classes, the weight matrix W should be 784 × 10.
6. The prediction outputs of our model should be 10 probability values
for each of these 10 digits c to be the label digit for a given image x:
P (y = c | x).
7. Finally, the row vector for the bias b is 10 in dimension.

12.3.4 Multiclass logistic regression

Define softmax function for our 10-classification problem.

def softmax(y_linear):
exp=nd.exp(y_linear-nd.max(y_linear,axis=1).reshape((-1,1)))
norms = nd.sum(exp, axis=1).reshape((-1,1))
return exp / norms

sample_y_linear = nd.random_normal(shape=(2,10))
sample_yhat = softmax(sample_y_linear)
print(sample_yhat)
print(nd.sum(sample_yhat, axis=1))
452 Machine Learning with Python: Theory and Applications

[[0.08024886 0.04739863 0.02992393 0.08034719 0.03729992

0.01006166 0.3406885 0.2191301 0.04921847 0.10568278]
[0.02349378 0.02465977 0.06156655 0.04537923 0.11654365
0.04362861 0.5023322 0.02725256 0.1258423 0.02930132]]
<NDArray 2x10 @cpu(0)>

[1. 1.]
<NDArray 2 @cpu(0)>

This confirms that each of the rows sums to 1, as expected.

12.3.5 Defining a neural network model

We define now our 10-classification model.

def net(X):
y_linear = nd.dot(X, W) + b
yhat = softmax(y_linear)
return yhat

12.3.6 Defining the cross-entropy loss function

Because our k-classification problem is to predict probabilities, the relevant

loss function would be the cross-entropy function. In digit classification case,
label vector y always has a 1 (hot) in its entries. Hence, it is sometimes called
one-hot vector. The value corresponding to the correct label digit is set to 1
and the rest are set to 0. For example, if the label is 2, we set

y = [0, 1, 0, 0, 0, 0, 0, 0, 0, 0] (12.10)

Because we only care about the component in ŷ corresponding to 2, and

the cross-entropy counts only the hot 1. The minimization algorithm can
then drive ŷ to approach y, by updating the weights in W and biases in b
in the net.
The code definition of the cross-entropy is as follows:

def cross_entropy(yhat, y): # cross- entropy loss

return - nd.sum(y * nd.log(yhat+1e-6))
# return nd.sum((y - yhat)**2) # one may try this too
 Multiclass Classification 453

12.3.7 Optimization method

We use the stochastic gradient descent (SGD) for optimization.

def SGD(params, lr):

for param in params:
param[:] = param - lr * param.grad

12.3.8 Accuracy evaluation

We can compute the accuracy of a model with its current learning parame-
ters, by finding out the ratio of the number of correct answers with the total
number of tests. We use the following evaluation loop to do this:

def evaluate_accuracy(data_iterator, net):

numerator = 0.
denominator = 0.
for i, (data, label) in enumerate(data_iterator):
data = data.as_in_context(model_ctx).reshape((-1,784))
label = label.as_in_context(model_ctx)
label_one_hot = nd.one_hot(label, 10)
output = net(data)
predictions = nd.argmax(output, axis=1)
numerator += nd.sum(predictions == label)
denominator += data.shape[0]
return (numerator / denominator).asscalar()

For our randomly initialized model, roughly one tenth of all examples
may belong to each of the 10 classes. We thus expect an accuracy of 0.1.
Now, we are ready to train the model.

12.3.9 Initiation of the model and training execution

# Initialize the learning parameters randomly

W=nd.random_normal(shape=(num_inputs, num_outputs),ctx=model_ctx)
b=nd.random_normal(shape=num_outputs,ctx=model_ctx)

params = [W, b]
print(W.shape,b.shape)
454 Machine Learning with Python: Theory and Applications

# Let MXNet know that the gradients with respect to each of

# these parameters will be needed during training.

for param in params:

param.attach_grad()

# Estimate the accuracy of the un-trained model

print("Accuracy of untrained model:",
evaluate_accuracy(test_data,net))

import time # To time the performance

epochs = 10 #20
learning_rate = .005

print('Parameter setting: epochs =',epochs, ' learning_rate =',\

learning_rate,' batch_size=',batch_size)

start_t = time.process_time()
for e in range(epochs):
cumulative_loss = 0
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(model_ctx).reshape((-1,784))
label = label.as_in_context(model_ctx)
label_one_hot = nd.one_hot(label, 10)
with autograd.record():
output = net(data)
loss = cross_entropy(output, label_one_hot)
loss.backward()
SGD(params, learning_rate)
cumulative_loss += nd.sum(loss).asscalar()

test_accuracy = evaluate_accuracy(test_data, net)

train_accuracy = evaluate_accuracy(train_data, net)
if e%3 == 0 or e == epochs-1:
print("Epoch %s. Loss: %s, Train_acc %s,
Test_acc %s" %
(e,cumulative_loss/num_examples,train_accuracy,
test_accuracy))

t_elapsed = (time.process_time() - start_t)

print('One-layer neural network, training time =',
f'{t_elapsed}','s')
 Multiclass Classification 455

(784, 10) (10,)

Accuracy of untrained model: 0.1152
Parameter setting: epochs = 10 learning rate = 0.005 batch size= 64
Epoch 0. Loss: 1.3851381640116374, Train acc 0.8584833,
Test acc 0.8635
Epoch 3. Loss: 0.4689176872412364, Train acc 0.898, Test acc 0.8941
Epoch 6. Loss: 0.39240473754405975, Train acc 0.9094, Test acc 0.9035
Epoch 9. Loss: 0.3527593985239665, Train acc 0.9142333,
Test acc 0.9077
One-layer neural network, training time = 1696.734375 s

12.3.10 Prediction with the trained model

We pick 10 samples randomly from the test dataset, and then use our trained
model to make a prediction.

# Define the function to do prediction

def model_predict(net,data):
output = net(data)
return nd.argmax(output, axis=1)

# let us sample 10 random data-points from the test set

plt.figure(figsize=(15, 60))

sample_data = mx.gluon.data.DataLoader(mnist_test,10,shuffle=True)
start_t = time.process_time()
for i, (data, label) in enumerate(sample_data):
data = data.as_in_context(model_ctx)
print(data.shape)
im = nd.transpose(data,(1,0,2,3))
im = nd.reshape(im,(28,10*28,1))
imtiles = nd.tile(im, (1,1,3))

plt.imshow(imtiles.asnumpy())
plt.show()
pred=model_predict(net,data.reshape((-1,784)))
print('model predictions are:', pred)
break

t_elapsed = (time.process_time() - start_t)

print('One-layer neural network, prediction time=',
f'{t_elapsed}','s')

(10, 28, 28, 1)

456 Machine Learning with Python: Theory and Applications

Figure 12.3: Predicated digits from the trained one-layer NN, using the testing dataset
of MNIST.

model predictions are:

[6. 8. 6. 4. 4. 1. 9. 9. 1. 7.]
<NDArray 10 @cpu(0)>
One-layer neural network, prediction time= 0.5 s

Summary of this study: The accuracy is about 90% after five epochs of
training, and the time taken on the author’s laptop for the training is about
500 s. The time taken for the prediction of 10 digits is only about 0.55 s.
When using 10 epochs, the test accuracy increases to 91%, and the time
taken for the training is 1,100 s (about double). When using 50 epochs, the
test accuracy increases to 92%. When using 100 epochs, the test accuracy
increases to 92.3%. Further training using this model may not make much
difference. To improve further, one may need to increase the learning ability
of the net to capture more nonlinear features for the images of handwritten
digits. Based on the Universal Prediction Theory, this can be done using
more hidden layers known as the multilayer perceptron (MLP, see the next
chapter).

12.4 Case Study 2: Handwritten Digit Classification with

Sklearn Random Forest Multi-Classifier

For comparison, let us do the same digit prediction using the gradient
boosting classifier. It is one of the powerful techniques for classification
problems. Readers may read an online article by Harshdeep Singh at towards
datascience. We will not discuss this technique in detail, but will simply
use the model and codes provided by scikit learn (or Sklearn) [1] for this
comparison study. We slightly modified the code to allow readers to view
more detailed differences in the results for different parameter settings.

import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
%matplotlib inline
 Multiclass Classification 457

# Read-in and take a look at the data

data=pd.read_csv('mnist.csv') # the mnist data already in a
# csv file under the local direction
data.head() # take a look at the first few lines
label 1x1 1x2 1x3 1x4 1x5 1x6 1x7 1x8 1x9 ... 28x19 28x20 \
0 5 0 0 0 0 0 0 0 0 0 ... 0 0
1 0 0 0 0 0 0 0 0 0 0 ... 0 0
2 4 0 0 0 0 0 0 0 0 0 ... 0 0
3 1 0 0 0 0 0 0 0 0 0 ... 0 0
4 9 0 0 0 0 0 0 0 0 0 ... 0 0

28x21 28x22 28x23 28x24 28x25 28x26 28x27 28x28

0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0
2 0 0 0 0 0 0 0 0
3 0 0 0 0 0 0 0 0
4 0 0 0 0 0 0 0 0

5 rows x 785 columns

# Take look at the image of a digit.

digit_array = data.iloc[4,1:].values
digit_image = digit_array.reshape(28,28).astype('uint8')
plt.imshow(digit_image)

<matplotlib.image.AxesImage at 0x1eea6883dd8>

Figure 12.4: A sample image of handwritten digit in the MNIST dataset.

458 Machine Learning with Python: Theory and Applications

# Split the data into train and test datasets.

df_x=data.iloc[:,1:]
df_y=data.iloc[:,0]
x_train, x_test, y_train, y_test = train_test_split(df_x, df_y,\
test_size=0.25, random_state=4)
y_train.head() # Take a look at it
50965 8
1315 3
57499 4
11420 2
1328 5
Name: label, dtype: int64

# Use sklearn RandomForestClassifier() to train the model.

rf=RandomForestClassifier()
#(n_estimators=200,min_samples_split=4)
rf.fit(x_train,y_train)
# Predict the digits in the test dataset.
predicted_digits =rf.predict(x_test)

# Print out the results and image

print('First 10 true digits in the test set:',y_test.values[0:10])
fig, figs = plt.subplots(1, 10, figsize=(15, 15))
for i in range(10):
digit_array=x_test.iloc[i,:].values
figs[i].imshow(digit_array.reshape(28,28).astype('uint8'))
figs[i].axes.get_xaxis().set_visible(False)
figs[i].axes.get_yaxis().set_visible(False)

plt.show()
print('Predicated digits for the first 10 in the test set:',
predicted_digits[0:10])
First 10 true digits in the test set: [0 6 5 6 2 8 9 5 0 9]

Figure 12.5: Predicated handwritten digits using the Sklearn Random forest and the
testing dataset of MNIST.
 Multiclass Classification 459

Predicated digits for the first 10 in the test set:

[0 6 5 6 2 7 9 5 0 9]

# compute the accuracy for the whole test set

print('Total number of digits in the test set:',
len(predicted_digits))
correct = 0
for i in range(len(predicted_digits)):
if predicted_digits[i] == y_test.values[i]:
correct += 1

print('Total number of correctly predicted_digitsicted digits:',

correct)
print('The accuracy is:',100*correct/len(predicted_digits),'%')

Total number of digits in the test set: 17500

Total number of correctly predicted digitsicted digits: 16964
The accuracy is: 96.93714285714286 %

Summary of this study: The result is better than the one-layer

neural network. Without discussing the details, further tests have given the
following findings:

• When n estimators = 50, and min samples split=2 are used, the accuracy
is: 96.11%.
• When n estimators = 50, and min samples split=2 are used, the accuracy
is: 96.81%.
• When n estimators = 100, and min samples split=2 are used (this is the
default setting), the accuracy is: 96.96%.
• When n estimators = 100, and min samples split=4 are used, the accuracy
is: 96.91%.
• When n estimators = 200, and min samples split=2 are used, the accuracy
is: 97.15%.
• When n estimators = 400, and min samples split=2 are used, the accuracy
is: 97.14%.
• When n estimators = 200, and min samples split=4 are used, the accuracy
is: 97.01%.
• When n estimators = 400, and min samples split=4 are used, the accuracy
is: 97.14%.
460 Machine Learning with Python: Theory and Applications

12.5 Case Study 3: Comparison of Random Forest,

Extra-Forest, and Gradient Boosting for Multi-Classifier

We now introduce a study on comparison between the random-forest, extra-

forest, and gradient boosting methods for multi-classification problems, using
the same MNIST digit prediction and the modules provided by the scikit
learn team [1].

#from sklearn.datasets import load_digits

from sklearn.ensemble import RandomForestClassifier
from sklearn.ensemble import ExtraTreesClassifier
from sklearn import ensemble
from sklearn import metrics
from sklearn.model_selection import train_test_split
from matplotlib import pyplot as plt
import numpy as np
import pandas as pd
import time
%matplotlib inline

# Read-in and take a look at the data

data=pd.read_csv('mnist.csv')
#data.head()

# Split the data into train and test datasets.

df_x=data.iloc[:,1:]
df_y=data.iloc[:,0]
x_train,x_test,y_train,y_test=train_test_split(df_x, df_y,\
test_size=0.25, random_state=4)
y_train.head() # Take a look at it

50965 8
1315 3
57499 4
11420 2
1328 5
Name: label, dtype: int64
 Multiclass Classification 461

# Using the classifiers in sklearn to train the model, and timeit

# %%timeit is not working, use time
n_e=10 #n_e = 100: very slow for the GradientBoosting Classifier
m_s=2
print('Parameters:n_estimators=',n_e,' min_samples_split=',m_s)

rf=RandomForestClassifier(n_estimators=n_e,min_samples_ split=m_s)
start_t = time.process_time()
rf.fit(x_train,y_train)
t_elapsed = (time.process_time() - start_t)
print('RandomForestClassifier, training time =',f'{t_elapsed}','s')

ef = ExtraTreesClassifier(n_estimators=n_e, min_samples_split=m_s)
start_t = time.process_time()
ef.fit(x_train,y_train)
t_elapsed = (time.process_time() - start_t)
print('ExtraTreesClassifier, training time =',f'{t_elapsed}','s')

gf = ensemble.GradientBoostingClassifier(n_estimators=n_e,\
min_samples_split=m_s)
start_t = time.process_time()
gf.fit(x_train,y_train)
t_elapsed = (time.process_time() - start_t)
print('GradientBoostingClassifier,training time=',\
f'{t_elapsed}','s')

Parameters:n estimators= 10 min samples split= 2

RandomForestClassifier, training time = 4.5625 s
ExtraTreesClassifier, training time = 3.65625 s
GradientBoostingClassifier,training time= 187.71875 s

The above test conditions are as follows:

• n estimators = n e = 100, min samples split = m s= 2 for all classifiers.

• All other parameters are defaults set by the corresponding classifiers.

The test results are as follows:

• RandomForestClassifier, training time = 40.625 s, which is about 10 times

faster than the one-layer neural network.
• ExtraTreesClassifier, training time = 39.141 s.
• GradientBoostingClassifier, training time = 1958.8 s, which is about four
times slower than the one-layer neural network.
462 Machine Learning with Python: Theory and Applications

When the test conditions are set to the following:

• n estimators = n e = 10, min samples split = m s= 2 for all classifiers.

• All other parameters are defaults set by the corresponding classifiers.

The test results are as follows:

• RandomForestClassifier, training time = 5.2969 s.

• ExtraTreesClassifier, training time = 4.8594 s.
• GradientBoostingClassifier, training time = 211.97 s.

Time scalability is roughly linear with n estimators.

# Predict the digits in the test dataset, and time it.

print('Parameter setting:n_estimators=',n_e,
'min_samples_split=',m_s)

start_t = time.process_time()
rf_predicted_digits =rf.predict(x_test)
t_elapsed = (time.process_time() - start_t)
print('RandomForestClassifier, prediction time=',
f'{t_elapsed:.5f}','s')

start_t = time.process_time()
ef_predicted_digits =ef.predict(x_test)
t_elapsed = (time.process_time() - start_t)
print('ExtraTreesClassifier, prediction time =',
f'{t_elapsed:.5f}','s')

start_t = time.process_time()
gf_predicted_digits =gf.predict(x_test)
t_elapsed = (time.process_time() - start_t)
print('GradientBoostingClassifier,prediction time=',
f'{t_elapsed:.5f}','s')

Parameter setting:n estimators= 10 min samples split= 2

RandomForestClassifier, prediction time= 0.14062 s
ExtraTreesClassifier, prediction time = 0.09375 s
GradientBoostingClassifier,prediction time= 0.17188 s
 Multiclass Classification 463

The test conditions for the training are as follows:

• n estimators = n e = 100, min samples split = m s= 2 for all classifiers.

• All other parameters are defaults set by the corresponding classifiers.

The test results on prediction time of the classifiers are as follows:

• RandomForestClassifier, prediction time = 1.40625 s.

• ExtraTreesClassifier, prediction time = 1.12500 s.
• GradientBoostingClassifier, prediction time = 1.50000 s.

When the test conditions for the training are as follows:

• n estimators = n e = 10, min samples split = m s= 2 for all classifiers.

• All other parameters are defaults set by the corresponding classifiers.

The test results on prediction time of the classifiers are as follows:

• RandomForestClassifier, prediction time = 0.20312 s.

• ExtraTreesClassifier, prediction time = 0.20312 s.
• GradientBoostingClassifier, prediction time = 0.21875 s.

It is clear that the time for prediction is much less than linear with
n estimators, and they all are very fast in general.

print('Total number of digits in the test set:',

len(rf_predicted_digits))
rf_accuracy = metrics.accuracy_score(y_test, rf_predicted_digits)
ef_accuracy = metrics.accuracy_score(y_test, ef_predicted_digits)
gf_accuracy = metrics.accuracy_score(y_test, gf_predicted_digits)
print('The rf_accuracy is:',f'{rf_accuracy:.5f}','%')
print('The ef_accuracy is:',f'{ef_accuracy:.5f}','%')
print('The gf_accuracy is:',f'{gf_accuracy:.5f}','%')

Total number of digits in the test set: 17,500

The rf accuracy is: 0.94657 %
The ef accuracy is: 0.94829 %
The gf accuracy is: 0.83806 %

The conditions for the training are as follows:

• n estimators = n e = 100, min samples split = m s= 2 for all classifiers.

• All other parameters are defaults set by the corresponding classifiers.
464 Machine Learning with Python: Theory and Applications

The test accuracy results on prediction of the classifiers are as follows:

• Total number of digits in the test set: 17,500.

• The rf accuracy is: 0.97023%.
• The ef accuracy is: 0.97160%.
• The gf accuracy is: 0.94651%.

When the following conditions are used for the training:

• n estimators = n e = 10, min samples split = m s= 2 for all classifiers.

• all other parameters are defaults set by the corresponding classifiers.

The test accuracy results on prediction of the classifiers become as follows:

• Total number of digits in the test set: 17,500.

• The rf accuracy is: 0.94577%, which is about 4% higher than the neural
network with a single layer.
• The ef accuracy is: 0.95011%.
• The gf accuracy is: 0.83806%, which is about 6% lower than the neural
network with a single layer.

# Print out the results and image

print('True digits for the 10 first in the test set:',
y_test.values[0:10])
print('rf_predicated digits for first 10 in testset:',
rf_predicted_digits[0:10])
print('ef_predicated digits for first 10 in testset:',
ef_predicted_digits[0:10])
print('gf_predicated digits for first 10 in testset:',
gf_predicted_digits[0:10])
True digits for the 10 first in the test set: [0 6 5 6 2 8 9 5 0 9]
rf predicated digits for first 10 in testset: [0 6 5 6 2 9 9 3 0 9]
ef predicated digits for first 10 in testset: [0 6 5 6 2 3 9 5 0 9]
gf predicated digits for first 10 in testset: [0 6 5 6 2 7 9 8 0 9]

12.6 Multi-Classification via TensorFlow

Due to rapid development over the past decades, many Python packages
have already been developed for k-classification problems. TensorFlow is one
such excellent package. Readers may take a look at the TensorFlow tutorial
problems of classification images of clothing, at https://www.tensorflow.org/
tutorials/keras/classification. After going through this chapter, readers shall
not have much problem in understanding this TensorFlow tutorial.
 Multiclass Classification 465

12.7 Remarks

Note that the performance of a module can depend on a list of factors,

including settings in computer hardware, datasets, condition of the computer
system, and even on when it is run. The setting parameters used in the
comparison tests presented above may not be entirely fair for some classifiers.
The test outcome is therefore highly subjective. When developing large-scale
machine learning models, for which computational resources are a concern,
one may use a smaller-scale model (such as setting n estimators at a smaller
number) and quickly test the models for their datasets and then choose a
proper one for the large-scale project.

Reference
[1] P. Fabian, V. Gae, G. Alexandre et al., Scikit-learn: Machine learning in Python,
Journal of Machine Learning Research, 12(85), 2825–2830, 2011. http://jmlr.org/
papers/v12/pedregosa11a.html.
MACHINE LEARNING
WITH PYTHON
 Chapter 13

Multilayer Perceptron (MLP)

for Regression and Classification

In the previous chapter, we have seen examples of using a simple (input

and output layers, only one layer has learning parameters) neural network
for k-classification of handwritten digits. We found that there is room for
improvements, and we have shown that other classifiers like the random
forest and the extra-forest performed better than the one-layer neural
network. This chapter will show that a neural network with more layers (by
adding in some hidden layers) is one of the ways to make an improvement.
This type of neural networks is often known as multilayer perceptron (MLP)
[1–5].
The fundamental theoretical basis for MLPs is the Universal Prediction
Theory discussed in Chapter 5. It uses a chain of stacked affine transfor-
mations wrapped with nonlinear activation functions grows the dimension
of hypothesis space, increases the learning parameters, giving as much as
needed learning ability. It is regarded as a universal model with unlimited
learning abilities, at least in theory, that can be trained and applied for
all types of problems [1–3], including classification, regression, and inverse
analysis of mechanical problems [6–9].

13.1 The General Architecture and Formulations of MLP

13.1.1 The general architecture

The difference in the configuration of an MLP from the one-layer neural
networks is that an MLP has a number of hidden layers in addition to the
two (input and output) layers. It is the base network for many other types

467
468 Machine Learning with Python: Theory and Applications

Figure 13.1: An MLP with multiple hidden layers with a chain of stacked affine
transformations. All the neurons are connected with weights (dense, or fully connected),
and a bias is also added (not shown) to each of the neurons in all the layers (except the
input layer). Here cross entropy loss is used for classification, but loss function discussed in
Chapter 12 may also be used. When the MLP is used for regression, L2 loss or any other
loss function discussed in Chapter 10 may be used.

of modern configurations of neural networks. The typical configurations of

an MLP is shown in Fig. 13.1.
The MLP shown in Fig. 13.1 has nh hidden layers with neurons in gray
color. Together with the input layer of p green neurons and output layer of k
red neurons, the net has a total of NL = n + 2 layers. Only NL = n + 1 layers
has learning parameters Ŵ. It is thus called NL -layer network. Beginning
from the input layer, each layer feeds leftward (forward) into the next layer
sequentially, until the output layer. This a typical architecture for the MLP
networks for p → k mapping.
We may have an arbitrary number of hidden layers, each of which can have
an arbitrary number of neurons. Each neuron in a layer is fully connected to
neurons in the next layer, known as fully connected (unless we use special
techniques, such as dropouts, to make the connection sparse, mainly for
regularization purposes). A fully connected layer is also called dense layer.
The use of more hidden layers leads to an increase in the complexity of
the neural network and hence the learning capacity. In principle, a dense
MLP can have a large number of training parameters to model complicated
problems.
 Multilayer Perceptron (MLP) for Regression and Classification 469

13.1.2 The xw+b formulation

At the input layer, we have a input vector of training samples with p features:

x(0) = x = [x1 , x2 , . . . , xp ] (13.1)

At the output layer, we can have a single neuron or multiple neurons,

depending on the need of the problem to be modeled. In general, let us have
k neurons.
The optimization algorithms needed to find the minimizers (of all the
learning parameters) for the loss function are in principle the same as
these neural networks studied earlier. One shall expect increasing difficulties
during the training of such deepnets, mostly because of the vanishing
gradient problems. The Relu activation function becomes a frequent choice
for the neurons in the hidden layers to overcome such problems, which will
be discussed in detail in the next section.
Recall that in the k-classification problem, we used the softmax as the
prediction function at the output layer of a neural network. It is defined
as
(NL ) (NL −1) Ŵ(NL ) )
ez e(x
ŷ = softmax(z(NL ) ) =  (NL )
= k (NL −1) Ŵ(NL ) [i])
(13.2)
(x
i=1 e
k zi
i=1 e

(N )
where z(NL ) is an affine transformed vector with k components zi L (i =
1, 2, . . . , k), and the computation of the exponentiates is element-wise. The
dimensions of all these matrices can be seen in the following form:

z(NL ) = x(NL −1) W(NL ) + b(NL ) = x(NL −1) Ŵ(NL ) (13.3)

1×k 1×p p×k 1×k 1×(p+1) (p+1)×k

For an MLP network, x(NL −1) is the output from the last hidden layer and
can be written as a vector of

(NL −1) (NL −1)

x(NL −1) = [x1 , x2 , . . . , x(N
nh
L −1)
] (13.4)

where the substrate nh is the number of neurons in the last hidden layer of
the MLP.
470 Machine Learning with Python: Theory and Applications

For an MLP, in the feed-forward process, the values that the neurons
receive and produce are
x(0) = x = [x1 , x2 , . . . , xp ] feature at the input layer
(0) (1) (1)
x(1) = φ(1) (x
 W  + b ) feature at the 1st hidden layer
z(1)
x(2) = φ (x W(2) +
(2)

(1)
 b(2)) feature at the 2nd hidden layer
z(2)
...
(i−1) (13.5)
x(i) = φ(i) (x
 W((i)) ((i))
 + b ) feature at the ith hidden layer
z(i)
...
ŷ = x(NLi )
(NL − 1)
= φ(NLi ) (x
 W(NLi )
 + b
(NLi )
) prediction at the output layer
z(NLi )

Here, the number in the parentheses in the superscript stands for the layer
number. The terms above the curly brackets are the affine transformation
functions z(i) (i = 1, 2, . . . , NL ), which can change from layer to layer and
bring in information for the past layers. After fitted to the activation
function, it becomes new independent “features” for the next layer. The
dimension of the new feature space changes accordingly.
The above chain equations for the network flow can also be expressed as

ŷ = φ(NL ) (· · · φ(2) (φ(1) (xW(1) + b(1) )W(2) + b(2) ) · · · W(NL ) + b(NL ) ) (13.6)

We now see explicitly that our MLP model is indeed a giant function
with k components in the feature space Xp and controlled (parameterized)
by the training parameters in WP , as mentioned in Section 1.5.5. When these
parameters W(i) (i = 1, 2, . . . , NL − 1) are tuned, one gets k giant functions
over the feature space Xp . On the other hand, this set of k giant functions can
also be viewed as differentiable functions of these parameters for any given
data-point in the dataset, which can be used to form a differentiable loss
function using the corresponding k labels given in the dataset. The training
is to minimize such a loss function for all the data-points in the dataset, by
updating the training parameters to become minimizers.
This feed-forward network creates a chain “reaction”, and such a nice
flowing structure is made use of in computing the gradients in the so-
called backward propagation or backprop or BP [10], using the chain rule
of differentiation. The backprop process can be achieved via the autograd
discussed in detail in Chapter 8. Each layer in an MLP has its own set
of learning parameters (in the weight matrix and bias vector). As the layer
 Multilayer Perceptron (MLP) for Regression and Classification 471

number grows, the total number of learning parameters can grow fast. Notice
also that the activation function used in each layer can be in theory different,
although in many current practices, the hidden layer often uses the same type
of activation function, typically Relu and tanh.
For the last output layer, the activation function is often chosen differently
based on the application problem the network is built for. For many
engineering problems of the regression type, we may simply use a linear
activation function, meaning that we do nothing for the last layer, and output
whatever is received.

ŷ = x(NL ) = z(NL ) = x(NL −1) W(NL ) + b(NL ) = x(NL −1) Ŵ(NL ) (13.7)

In this case, the mean square error (or L2) loss function is often used and
minimized.
For classification problems, we may use the sigmoid or tanh function for
binary classification. In general, one can use the softmax as the prediction
function, as given in Eq. (13.2). Note that we have proved in the previous
chapter that a binary classification problem can be treated as a special case
of k-classification problem.

13.1.3 The xw formulation, use of affine transformation

weight matrix
Using the affine transformation weight matrix W defined in Chapter 5, we
have for the ith layer,
 
(j) 1 b(j)
W = , j = 1, 2, . . . , NL (13.8)
0 W(j)

We can then perform a chain of ATA transformations. The feed-forward

process an MLP can have the following neat formulation.
n0
x(0) = [1, x1 , x2 , . . . , xp ] ∈X at the input layer

(1) n1
x(1) = φ(1) x(0) W ∈X at the 1st hidden layer

(2) n2
x(2) = φ(2) x(1) W ∈X at the 2nd hidden layer
... (13.9)

(j) nj
x(j) = φ(j) x(j−1) W ∈X at the ith hidden layer
...

ŷ = φ(NL ) x(NL −1) Ŵ(NL ) ∈ YnNL prediction at the output layer

472 Machine Learning with Python: Theory and Applications

where nj is the pseudo-dimension of the affine space in the j layer, or the

number of neurons for the interlayer features plus 1. For the input layer, n0 =
p + 1, and for the output layer, nNL = k, where the connection terminates
and an affine transformation is no longer needed for these k output neurons.
Notice that in the xw formulation, we assume the linear (i.e., no)
activation function is applied to constant 1. Under this condition, we can
easily prove that

  
(j) b(j)1
φ(j) x(j−1) W = φ(j) x(j−1)
W(j)0
  
(j)
(j−1)
 1 b(j)
=φ 1 x
0 W(j)
  (13.10)
= φ(j) 1 x(j−1) W(j) + b(j)
 
= 1 φ(j) x(j−1) W(j) + b(j)

= 1 x(j) = x(j)

In the 3rd line in Eq. (13.10), we moved φ(j) inside the array, because we
assumed the constant 1 does not subject to activation and the fact that
φ(j) operation is element-wise. It is clear that Eq. (13.10) achieves the same
results as in Eq. (13.5), but with an extra constant 1 component in the
array. The constant 1 is automatically generated when moving to the next
layer. This is because an affine transformation stays in the affine space, as
demonstrated in Chapter 5. The affine transformation weight matrix Ŵ(j)
performs the transformation from one affine space to another. Therefore,
when our xw formulation is used, there is no need to inject the bias b to each
neuron, which is one major difference from the xw+b formulation.
Note also that at the output layer, affine transformation ends and we use
only Ŵ(NL ) as discussed in Chapter 5, producing predictions naturally back
to a vector label space Y (k) , where a loss function can now be constructed
for ”terminal” control, using given labels.
Using the affine transformation matrix, Eq. (13.6) becomes neatly to,

(1) (2)
ŷ = φ(NL ) · · · φ(2) φ(1) x W W · · · Ŵ(NL ) (13.11)

Assuming the number of (total) neurons in the jth layer is nj , the dimension
(j)
of the affine transformation weight matrix W is nj−1 × nj , offers a natural
match in matrix computations for chained ATAs.
 Multilayer Perceptron (MLP) for Regression and Classification 473

13.1.4 MLP configuration with affine transformation

weight matrix
Using the affine transformation weight matrix, our MLP can have a config-
uration shown in Fig. 13.2.
The general formula for calculating the total number of learning param-
eters of an MLP becomes,

P = [n0 (n1 − 1)] + [n1 (n2 − 1)] + · · · + [ni−1 (ni − 1)]

        
layer 1 layer 2 layer i
(13.12)
+ · · · + [nNL −1 (nNL − 1)]
  
layer NL

The reason for subtracting 1 in the parenthesis is because the first column
(j)
in the affine transformation weight matrix W contains all constants (1 or
zeros) that are not trainable, as shown in Eq. (13.8). Equation (13.12) is an
alternative equation to Eq. (5.32) that is based on the xw+b formulation.
This number is usually reported in the summary report when an MLP model
is created in a module. The vector of trainable parameters ŵ is in space WP ,
as discussed in Section 1.5.5.

Figure 13.2: An MLP with multiple hidden layers with a chain of stacked affine
transformations for a n0 → k mapping. When affine transformation weight matrix is used,
the constant feature 1 in the top neuron of each layer is automatically generated. The
network flows naturally without injection of biases, and the dimensions of all these weight
matrices between layers match naturally. Here cross entropy loss is used for classification,
but loss function discussed in Chapter 12 may also be used. When the MLP is used for
regression, L2 loss or any other loss function discussed in Chapter 10 may be used.
474 Machine Learning with Python: Theory and Applications

13.1.5 Space evolution process in MLP

We discussed at the beginning in Chapter 1 about spaces in machine learning
models. We are now ready put the detailed connections in the following
equation.
(N −1)
W
(1)
W
(2) W L
n0 ×n1 n1 ×n2 n(N −2) ×n(N −1)
x(0) =⇒ x(1) =⇒ x(2) x(N L −1)
L L
n n
,..., =⇒ n
p+1 1 2 NL −1
∈X ∈X ∈X ∈X
(13.13)
Ŵ(NL )
n(N −1) ×n(N )
φ( ) L(ŷ(ŵ),y)
x(NL)
L L
=⇒ n
−→ ŷ ⇐⇒ y
∈X (NL )
∈Yk ∈Yk

Equation (13.13) shows clearly the chained structure for a p → k mapping.

The mapping is controlled by the learning parameters ŵ ∈ WP , through a
gigantic loss function L(ŷ(ŵ), y) chosen based on type of the problems to be
modeled. The discussion on how ŵ relates the weight matrices in Eq. (13.13)
are discussed in Section 1.5.5.
We emphasize again that the loss function is continuous and differentiable
with respect to all the learning parameters, because all the prediction and
activation functions used in all the layers are continuous and differentiable
with respect to the local arguments. In addition, the chained structure of
MLP allows the Autograd to work effectively through the layers to compute
the gradient, by the well-known chain rule of differentiation. Therefore, the
minimization techniques discussed in Chapter 9 can be used for finding
the minimizers in terms of these learning parameters. The only subtlety is
when the Relu function is used. It is continuous, piecewise differentiable, but
not differentiable at the origin. The gradient jumps at the origin, as shown
graphically in Section 7.6. We thus would need to compute its sub-gradient,
that is, the average of the gradients on the both sides of the origin, and use
it in the minimization process, as discussed at the end of Section 7.6.

13.2 Neurons-Samples Theory

The most often asked question in the machine learning community is on the
relationship between the number neurons/layers in an NN and the number
of samples (or data-points) in the dataset needed to train the NN. However,
no concrete answers were available, until now. The author has searched for
such a convincing answer but failed. In the past years, the author studied
various NNs, which has led the Neurons-Samples Theory [13] presented in
this section.
 Multilayer Perceptron (MLP) for Regression and Classification 475

13.2.1 Affine spaces and the training parameters

used in an MLP
Using Eq. (13.12), we can calculate the total number of trainable parameters
P for a given MLP. The P number does relate to the sample number m of the
dataset, but only loosely. The direct relationship is given by our Neurons-
Samples Theory, which uses a P ∗ -number that is the total pseudo-dimension
of the affine spaces involved in a machine learning model. The formula for
an MLP is given as follows.

P ∗ = n0 + n1 + · · · + nj−1 + · · · + nNL −1 (13.14)

     
layer 1 layer 2 layer j layer NL

The P ∗ -number increases with the number of layers and the number of
neurons in the previous layers. In other words, the P ∗ -number is also the
total number of neurons in the MLP, including neurons in the input layer
but excluding the neurons in the output layer where affine transformation is
no longer needed.
Deriving Eq. (13.14) requires a careful analysis and deep understanding
on the relationship between the pseudo-dimension of the affine space
nj−1 , dataset X(j−1) in the affine space, and its corresponding affine
(j)
transformation weight matrix W in the jth layer.
Optimal learning parameters in a layer: Consider a given MLP, from
which we isolate a nj−1 → 1 neural network (an ATU discussed in Section
5.2) with an interlayer input of nj−1 neurons and a single output neuron
at the jth layer, as shown in Fig. 5.2. The single output neuron is the ith
neuron in the layer. Using Eq. (9.12), the optimal solution for the learning
parameters ŵi∗ for that neuron can be written as follows.
 
 
X(j−1) X(j−1) ŵi∗ = X(j−1) yi (13.15)

where X(j−1) denotes the interlayer dataset with m data-points in the affine
space, with the original dataset at the input layer as X(0) . And yi is an
assumed label vector corresponding to the ith neuron in the layer. It is a
vector for these m data-points in the dataset. We do not usually know the
content of yi for these hidden layers. Fortunately, we do not need to know,
because its content is not relevant in this analysis.
Now, for the (i + 1)th neuron in the same jth layer, we shall have
 
 ∗ 
X(j−1) X(j−1) ŵi+1 = X(j−1) yi+1 (13.16)
476 Machine Learning with Python: Theory and Applications

where yi+1 an assumed label vector corresponding to the (i + 1)th neuron

in the same layer.
Comparison of Eqs.(13.15) and (13.16) finds that the existence of (unique)
solutions ŵi∗ and ŵi+1
∗ all depend on the same interlayer dataset X(j−1) :

whether or not [X(j−1) X(j−1) ] is invertible. The terms in the right hand-side
of these two equations are irrelevant. This reveals a very important fact that
ŵi∗ for all the neurons in a given layer, depend only on the same dataset
with m data-points. An n(j−1) → nj net is the same as an n(j−1) → 1 net in
this regard. In other words, all these ŵi∗ live in the same hypotheses space
Wn(j−1) , where n(j−1) is the pseudo-dimension of the affine space for the jth
layer (with n(0) = p + 1 for the input layer).
From another view point, we know that the affine space for the jth layer
n
is X (j−1) , the corresponding hypotheses space can only be Wn(j−1) .
Note that ŵi∗ may be a different vector for different i, but they are all

just vectors living in the column-space of matrix [X(j−1) X(j−1) ]. When the

column-space of matrix [X(j−1) X(j−1) ] is of full rank (when m is sufficiently
large), it has zero null-space, and ŵi∗ will be unique. Therefore, the pseudo-
dimension of the affine space for the jth layer should be the number directly
related to m for a layer. The optimal learning parameters for the jth layer
∗
should be: P(j−1) = n(j−1) .
Optimal learning parameters in MLPs: The above discussion is applicable
to each individual layer in an MLP. However, when multiple layers are
connected together to form the MLP (using nonlinear activation functions
in all the hidden layers), these interlayer inputs are linearly independent
from layer to layer, meaning the ranges of variations of ŵi in a layer is
determined only by the pseudo-dimension of the affine space in that layer,
which is nj−1 . In addition, the values in ŵi of a layer will be influenced
by those in other layers. Therefore, the total optimal training parameters is
effectively increased, which shall sum up the dimensions of the affine spaces of
all layers, leading to Eq. (13.14). This again shows the importance of using
nonlinear activation functions in the hidden layers for increasing optimal
training parameters in an MLP, and hence its learning ability.

13.2.2 Neurons-Samples Theory for MLPs

We are now ready to state the Neurons-Samples Theory.

Neurons-Samples Theory: The number of samples m in the dataset for
training an MLP capable of producing a unique solution must be no-less
 Multilayer Perceptron (MLP) for Regression and Classification 477

than the P ∗ -number, which is the sum of the dimensions of all the affine
spaces used in 1 ∼ NL layers.

m ≥ P∗ (13.17)

In other words, the P ∗ -number is the lower bound of m.

Alternatively, this theorem can be stated as: The number of samples m in
the dataset for training an MLP capable of producing a unique solution must
be no-less than the total number of neurons in the MLP. In this statement,
the total number of neurons include the neuron for constant 1 on the top of
each layer. It includes the neurons in the input layer, but excludes those in
the output layer. It is defined as the P ∗ -number in Eq. (13.14).
Based on our analysis, we note the following remarks:

1. The P ∗ -number is essentially the total number of neurons in the net. In

other words, the number of neurons is the number of data-points for an
ideal model.
2. The P ∗ -number is the sum of the dimensions of all the affine spaces used
in the net.
3. The P ∗ -number is independent of the number of the output neurons.
4. Equation (13.14) can also be used to estimate the needed number of layers
NL for an MLP:
P∗ m
NL = ≤ (13.18)
n n
where n is the average number of neurons used in 0 ∼ (NL − 1) layers.

We mention that Eq. (13.14) is the most important and essential equation
for estimating the minimum number of data-points for training an ML model
based on affine transformations, including the standard MLP, RNN and
CNN.
Note that in the openly available ML modules, the learning parameters
used in an MLP are the total number P given in Eq. (13.12). In all these
MLP or deepnet algorithms, P parameters are indeed allowed to vary
independently. However, because of the special construction of the MLP, ŵi
for all the neurons in the jth layer lives only in the same hypothesis-space
Wn(j−1) , which is only a small sub-space of Wn(j−1) ×(n(j) −1) , as discussed in
this section. Therefore, P ∗ given by Eq. (13.14), and not Eq. (13.12), relates
directly the number of samples in the dataset. The P ∗ -number is the lower
bound of m.
478 Machine Learning with Python: Theory and Applications

Note that the above analysis is based on regression formulations. For

classification problems, there are some subtleties, because the loss function
may use probability. In such cases, a correct prediction may not necessarily
need a perfect match in values for the predicted probability and that of the
given label. However, the principle of our above analysis shall apply.

13.3 Nonlinear Activation Functions for the Hidden Layers

For hidden layers in MLP, one has to use some nonlinear activation functions
to take the advantage of multilayer effects. If the whole network uses only
linear function, it would be equivalent to using the simple one-layer network,
because an output in a hidden layer will be linearly dependent on that of
the previous layer. This can be seen more clearly in the derivation below,
using the xw formulation:
(1) (2)
ŷ = (· · · ((x W )W ) · · · · · Ŵ(NL ) )
(1) (2)
(13.19)
=xW ·W · · · · · Ŵ(NL ) = xWcombined

where Wcombined is a product of weight matrices:

(1) (2)
Wcombined = W ·W · · · · · Ŵ(NL ) (13.20)
which is nothing but a single weight matrix for a one-layer network. The use
of such hidden layers does not increase the hypothesis space dimension and
the independent learning parameters. Therefore, there is no point at all in
using the linear activation function in any layer in an MLP. Using the linear
activation function in a layer of an MLP is effectively the same as removing
the layer from the MLP.

13.4 General Rule for Estimating Learning Parameters

in an MLP

Based on the Neurons-Samples Theory, we can now determine the number of

layers and the number of neurons in each layer. In practice, we shall consider
also the type of the problem to be modeled. Below is a general rule.
For a given dataset with m data-points, the upper bound of the P ∗ -
number used in an MLP should be m.
P∗ ≤ m (13.21)
where P ∗ is computed using Eq. (13.14). For the simplest linear regression
model with a single output neuron, P ∗ = p + 1, and Eq. (13.21) is the
 Multilayer Perceptron (MLP) for Regression and Classification 479

equality part of Eq. (9.8) in Chapter 9. To estimate the number of layers

needed, Eq. (13.7) may be used.
The dataset quality is defined as its representativeness to the underlaying
problem to be modeled, including correctness, size, data-point distribution
over the features space, and noise level. This guideline is very simple, because
it depends only on the size of the dataset. This assumes that (1) the m
data-points in the dataset are all linearly independent and (2) the training
parameters in the model are linearly independent following the Universal
Prediction Theory (see Chapter 5). For an MLP, if P ∗ < m, a trained
MLP may not pass the data-points, and generalizes better. If P ∗ = m,
the data-points may be reproduced by the MLP when it is fully trained. If
P ∗ > m, a fully trained MLP is capable of producing the data-points that
may have noise, leading possibly to overfitting (see Chapter 14) and may
not be used in general. Note in the practice that we may use MLP with
P ∗ > m and may not observe overfitting, because the MLP has not fully
trained.
It is worth mentioning again that m is only the upper bound for P ∗ . In
practice, P ∗ can be much smaller than m.

13.5 Key Techniques for MLP and Its Capability

We have discussed the Universal Approximation Theorem earlier. It states

that a network with a single hidden layer is capable of approximating any
function to a desired accuracy, as along as the hidden layer is sufficiently
“wide” with enough number of neurons. Based on our discussion on the
Universal Prediction Theory, increasing the layers can further increase the
learning ability of MLP. There is thus a question on whether a wide
(more neurons in a layer) net or a deeper (more layers) net should be
used. The general understanding is that: (1) when the number of layers
increases, one can significantly reduce the width of the net; (2) it has been
found that one may be able to approximate more complicated functions
more “compactly” if we use deeper neural networks compared to wider ones,
meaning that deepnets can use a smaller number of neurons in total. This
is because of the nonlinearity introduced in the MLP. Therefore, it is more
capable of modeling nonlinear behavior, and also it can produce decision
boundaries that may be highly nonlinear.
However, compared with wide nets, MLP is found more difficult to train
because of the well-known vanishing (or explosion) gradients of the loss
function. Therefore, in the past, wide nets were used more, until recent
480 Machine Learning with Python: Theory and Applications

breakthrough on techniques or algorithms for training deepnets. The major

techniques may include (but are not limited to) the following:

1. Relu activation function. Because the Relu function is linear with a

gradient of 1 (in the positive region), it allows the affine transformed
information in neurons in a layer to reach farther unaffected in the depth
of the net. There will be no squashing like in the sigmoid-type activation
functions near z = 1 where the gradients would become nearly zero. In
addition, the Relu function vanishes in the negative region and hence still
offers the needed nonlinearity for deepnets. Such a cut-off may also offer
some regularization effects. Moreover, the vanishing in the negative region
is for both the z function value and its local gradients together, which
eliminates other possible complications. It is the author’s understanding
that simply omitting the negative half of the information in an MLP
should not give rise to many problems, based on the Neurons-Samples
Theory. The number of the training parameter P in an MLP is usually
many times more than the P ∗ -number. The MLP has plenty of room to
adjust the learning parameters to find the minimizers, even if the negative
half of the information is discarded.
2. Sparse connection. This technique reduces the “density” of the training
parameters, such as the convolution neural networks that use filters
leading to sparse connections and hence a drastic reduction of the learning
parameters (weights and biases) in a layer. When the learning parameters
are not so crowded, the numerical conditioning improves. The legitimacy
of discarding some of the connections is also rooted in the Neurons-
Samples Theory, and the successive P number in an MLP.
3. Skip connection (or short cut). This is another technique that enables
the network to go further in depth. It is used in the development of the
so-called ResNet, which will be discussed more in Chapter 15. Essentially,
it allows the activated values of neurons in one layer to feed to the next-
next layer (skip one layer) before the activation function is applied. This
well mitigates the vanishing of gradients in deepnets.
4. Long short-term memory or LSTM [11]. This technique is useful
for overcoming the gradient vanishing problem. It is widely used in the
recurrent neural network (RNN, Chapter 16) that often has very deep
connections for handling time sequential inputs, such as in language
translation models.

Because of the nicely chained stacked structure of the MLP, simple

affine transformation, activation functions with analytical formula for their
 Multilayer Perceptron (MLP) for Regression and Classification 481

differentials, and the Autograd algorithms, there is no additional difficulty

for us to perform feed forward computation and backward propagation for
computing the gradients. This is done practically in the same way as we
have done for wide nets in the earlier chapters. Because the vanishing (or
explosion) gradients problem are mitigated effectively, one can now build up
and successfully train very deep neural networks that can have the needed
complexity for modeling complicated problems. This is why deepnets were
developed very fast in the past few years.

13.6 A Case Study on Handwritten Digits Using MXNet

Let us now build an MLP with two hidden layers and one output layer for
the same task of identification of handwritten digits. We use the code (with
slight modifications) provided at mxnet-the-straight-dope under the Apache-
2.0 License, which allows us to go through the process step by step. We will
use the same MNIST dataset that is familiar to us.

13.6.1 Import necessary libraries and load data

from __future__ import print_function

import mxnet as mx
import numpy as np
from mxnet import nd, autograd, gluon
# Set contexts according to computer hardware
ctx = mx.gpu() if mx.test_utils.list_gpus() else mx.cpu()
data_ctx = ctx
model_ctx = ctx
# Load the MNIST dataset
num_inputs = 784
num_outputs = 10
batch_size = 64
num_examples = 60000

def transform(data, label):

return data.astype(np.float32)/255, label.astype(np.float32)

train_data = gluon.data.DataLoader(mx.gluon.data.vision.
MNIST(train=True, transform=transform),
batch_size, shuffle=True)
test_data = gluon.data.DataLoader(mx.gluon.data.vision.
MNIST(train=False, transform=transform),
batch_size, shuffle=False)
482 Machine Learning with Python: Theory and Applications

13.6.2 Set neural network model parameters

# Here we set the model with two-hidden layers

# One can easily modify the network and try
num_hidden = 256
weight_scale = .01
# Allocate parameters for the first hidden layer
W1 = nd.random_normal(shape=(num_inputs, num_hidden),
scale=weight_scale, ctx=model_ctx)
b1 = nd.random_normal(shape=num_hidden,
scale=weight_scale, ctx=model_ctx)
# Allocate parameters for the second hidden layer
W2 = nd.random_normal(shape=(num_hidden,
num_hidden), scale=weight_scale,ctx=model_ctx)
b2 = nd.random_normal(shape=num_hidden,
scale=weight_scale, ctx=model_ctx)
# Allocate parameters for the output layer
W3 = nd.random_normal(shape=(num_hidden, num_outputs),
scale=weight_scale, ctx=model_ctx)
b3 = nd.random_normal(shape=num_outputs,
scale=weight_scale, ctx=model_ctx)
params = [W1, b1, W2, b2, W3, b3]
# Allocate now the space for the gradients of each parameter.
for param in params:
param.attach_grad()

13.6.3 Softmax cross entropy loss function

When computing the jth components of the output for a k-classification

problem, we use the following softmax function:

ezj
ŷj = k (13.22)
zi
i=1 e

where zj is the jth element of the input z in Eq. (13.2), and j = 1, 2, . . . , k.

We may have a numerical problem in computing these exponentials,
because the argument zi can be both very large and small. Hence, its
exponential ezi can be extremely large or small, leading to possible over-
or underflow in the computer that has limited precision. The solution to
 Multilayer Perceptron (MLP) for Regression and Classification 483

this is to use the so-called LogSumExp trick (https://en.wikipedia.org/wiki/

LogSumExp).
As we have shown in the previous chapter, we know that log function
is monotonic and will not affect the location of extremes of an argument
function that is positive (and ŷj ). We thus take logarithm to both sides of
Eq. (13.22), which gives
   k   k 
ezj  
zj zi zi
log(ŷj ) = log k = log(e ) − log e = zj − log e
zi
i=1 e i=1 i=1
(13.23)

Note that ezi is still there in the last term. We next compute the last
term using
 k  ⎛ ⎞
 k
 ∗
log ezi = z ∗ + log⎝ ezi −z ⎠ (13.24)
i=1 i=1,i=i∗

∗
where we simply factor out ez from the first sum with z ∗ = max(z1 ,
z2 , . . . , zk ) with i∗ being the index for z ∗ .
It is clear now that all the numbers involved in the computation are now
bounded. In these computations, we need only z as inputs. This formula
is known as softmax cross-entropy, because both ideas of softmax and log
cross-entropy are used. Using mxnet, it is defined as follows:

def softmax_cross_entropy(yhat, y):

return -nd.nansum(y*nd.log_softmax(yhat), axis=0, exclude=True)

13.6.4 Define a neural network model

def relu(z):
return nd.maximum(z, nd.zeros_like(z))

def net(X):
h1_linear = nd.dot(X, W1) + b1 # 1st hidden layer
h1 = relu(h1_linear) # Relu activation
h2_linear = nd.dot(h1, W2) + b2 # 2nd hidden layer
h2 = relu(h2_linear)
yhat_linear = nd.dot(h2, W3) + b3 # output layer
return yhat_linear
484 Machine Learning with Python: Theory and Applications

13.6.5 Optimization method

Here, we choose to use the stochastic gradient descent (SGD) method.
Readers may try other methods.
def SGD(params, lr):
for param in params:
param[:] = param - lr * param.grad

13.6.6 Model accuracy evaluation

The following function is defined for model accuracy evaluation.
def evaluate_accuracy(data_iterator, net):
numerator = 0.
denominator = 0.
for i, (data, label) in enumerate(data_iterator):
data=data.as_in_context(model_ctx).reshape((-1, 784))
label = label.as_in_context(model_ctx)
output = net(data)
predictions = nd.argmax(output, axis=1)
numerator += nd.sum(predictions == label)
denominator += data.shape[0]
return (numerator / denominator).asscalar()

13.6.7 Training the neural network and timing the training

import time # timing the performance

num_inputs = 784
num_outputs = 10
epochs = 10 # readers may change this and try
learning_rate = .001 # readers may change this and try
smoothing_constant = .01 # readers may change this and try
print('Parameter setting: epochs=',epochs,' num of neurons in
hidden layers =', num_hidden,'\n learning_rate =',
learning_rate, 'batch_size=',batch_size,
'smoothing_constant=',smoothing_constant)
start_t = time.process_time()

for e in range(epochs):
cumulative_loss = 0
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(model_ctx).reshape((-1,
num_inputs))
 Multilayer Perceptron (MLP) for Regression and Classification 485

label = label.as_in_context(model_ctx)
label_one_hot = nd.one_hot(label, num_outputs)
with autograd.record():
output = net(data)
loss = softmax_cross_entropy(output, label_one_hot)

loss.backward()
SGD(params, learning_rate)
cumulative_loss += nd.sum(loss).asscalar()

test_accuracy = evaluate_accuracy(test_data, net)

train_accuracy = evaluate_accuracy(train_data, net)
if e%3 == 0 or e == epochs-1:
print("Epoch %s. Loss: %s, Train_acc %s, Test_acc %s" %
(e,cumulative_loss/num_examples,train_accuracy,
test_accuracy))

t_elapsed = (time.process_time() - start_t)

print('3-layer neural network, training time =',f'{t_elapsed}','s')

Parameter setting: epochs= 10 num of neurons in hidden layers = 256

learning rate = 0.001 batch size= 64 smoothing constant= 0.01
Epoch 0. Loss: 1.2388756741285325, Train acc 0.8823, Test acc 0.8849
Epoch 3. Loss: 0.1648619899849097, Train acc 0.95958334,
Test acc 0.9578
Epoch 6. Loss: 0.08768035567005475, Train acc 0.9797, Test acc 0.973
Epoch 9. Loss: 0.056695277426143484, Train acc 0.98615,
Test acc 0.9746
3-layer neural network, training time = 2798.078125 s

When parameters are set as epochs = 10, learning rate = 0.001,

batch size = 64, and smoothing constant = 0.01, we found the following:
• at the 5th epoch, Loss: 0.12915, Train acc 0.96988, Test acc 0.9654
• at the 10th epoch, Loss: 0.056414, Train acc 0.98768, Test acc 0.9757.
The 3-layer neural network (with two hidden layers each with 256
neurons) needs a training time of 1,653 s, and produces test accuracy of about
97.6%. Compared to the 1-layer net without a hidden layer (in the previous
chapter) that took 1100 s to train and with 91% accuracy, the accuracy is
as much as 5.6% higher, which is quite a significant improvement. This is
achieved by capturing the nonlinear features of the handwritten digits. Note
that these numbers can change depending on computer system and business
of the computer.
486 Machine Learning with Python: Theory and Applications

13.6.8 Prediction with the model trained

We use some randomly selected data-points from the test set to perform
predictions.

%matplotlib inline
import matplotlib.pyplot as plt
# Define the function to do prediction
def model_predict(net,data):
output = net(data)
return nd.argmax(output, axis=1)

samples = 10
mnist_test=mx.gluon.data.vision.MNIST(train=False,
transform=transform)
# Sample randomly "samples" data-points from the test set
sample_data=mx.gluon.data.DataLoader(mnist_test,samples,
shuffle=True)
plt.figure(figsize=(10, 60))
for i, (data, label) in enumerate(sample_data):
data = data.as_in_context(model_ctx)
im = nd.transpose(data,(1,0,2,3))
im = nd.reshape(im,(28,samples*28,1))
imtiles = nd.tile(im, (1,1,3))
plt.imshow(imtiles.asnumpy())
plt.show()
pred=model_predict(net,data.reshape((-1,784)))
print('model predictions are:', pred)
print('true labels :', label)
break

Figure 13.3: Predicted digits using the 3-layer MLP on the testing dataset of MNIST.

model predictions are:

[0. 7. 8. 1. 9. 2. 6. 2. 2. 3.]
<NDArray 10 @cpu(0)>
true labels :
[0. 7. 8. 1. 9. 2. 6. 2. 2. 3.]
<NDArray 10 @cpu(0)>
 Multilayer Perceptron (MLP) for Regression and Classification 487

One can also try predictions using the training dataset. It should reflect
how well the model is trained.
mnist_test=mx.gluon.data.vision.MNIST(train=True,
transform=transform)
# Sample randomly "samples" data-points from the test set
sample_data=mx.gluon.data.DataLoader(mnist_test,samples,
shuffle=True)
plt.figure(figsize=(10, 60))
for i, (data, label) in enumerate(sample_data):
data = data.as_in_context(model_ctx)
im = nd.transpose(data,(1,0,2,3))
im = nd.reshape(im,(28,samples*28,1))
imtiles = nd.tile(im, (1,1,3))
plt.imshow(imtiles.asnumpy())
plt.show()
pred=model_predict(net,data.reshape((-1,784)))
print('model predictions are:', pred)
print('true labels :', label)
break

Figure 13.4: Predicted digits using the 3-layer MLP on the training dataset of MNIST,
just to see how good the training is.

model predictions are:

[1. 6. 7. 1. 4. 7. 4. 8. 3. 5.]
<NDArray 10 @cpu(0)>
true labels :
[1. 6. 7. 1. 4. 7. 4. 8. 3. 5.]
<NDArray 10 @cpu(0)>

13.6.9 Remarks
With two hidden layers (three layers with Ws), each containing 256 neurons,
the MLP model has achieved over 97% accuracy, compared to the 1-layer
net used in the previous chapter, and we had about 6% improvement.
488 Machine Learning with Python: Theory and Applications

When parameters are set as epochs = 10, learning rate = 0.001,

batch size = 64, and smoothing constant = 0.01, this 3-layer neural network
needs training time = 1,653 s, which is about 3 times slower than the random
forest model used in the previous chapter. The accuracy is about 0.5% higher
than the random forest model.
The same handwritten Digits MNIST and the Fashion MNIST recogni-
tion can also be easily done using the Tensorflow and Keras modules. Readers
may refer to the following reference materials for more details:

• For Digits MNIST: https://www.youtube.com/watch?v=wQ8BIBpya2k.

• For Fashion MNIST: https://www.tensorflow.org/tutorials/keras/classifi
cation.

13.7 Visualization of MLP Weights Using Sklearn

We now present a case study done by the scikit-learn team [12] on visua-
lization of MLP weights on the MNIST dataset. To have the model run
faster, we use only one hidden layer, and train it only for a small number of
iterations.

13.7.1 Import necessary Sklearn module

import matplotlib.pyplot as plt

from sklearn.datasets import fetch_openml
from sklearn.neural_network import MLPClassifier
import time

13.7.2 Load MNIST dataset

%matplotlib inline
# Load data from https://www.openml.org/d/554
X, y = fetch_openml('mnist_784', version=1, return_X_y=True)
X = X / 255.
# rescale the data, use the traditional train/test split
X_train, X_test = X[:60000], X[60000:]
y_train, y_test = y[:60000], y[60000:]
 Multilayer Perceptron (MLP) for Regression and Classification 489

13.7.3 Set an MLP model

hls= 50 # hidden_layer_sizes
mx_iter=8 # Maximum number of iterations for optimization
alpha=1e-2 # L2 penalty parameter for L2 regularization
lri=.1 # initial learning rate
mlp = MLPClassifier(hidden_layer_sizes=(hls,), max_iter=mx_iter,
alpha=alpha, solver='sgd', verbose=10,
random_state=1, learning_rate_init=lri)
print('Parameter setting: epochs=',epochs,'max_iter=',mx_iter,
' learning_rate_init =',lri)

Parameter setting: epochs= 10 max iter= 8 learning rate init = 0.1

13.7.4 Training the MLP model and time the training

start_t = time.process_time()
mlp.fit(X_train, y_train)
print("Training set score: %f" % mlp.score(X_train, y_train))
print("Test set score: %f" % mlp.score(X_test, y_test))
t_elapsed = (time.process_time() - start_t)
print('One-layer neural network, training time=',
f'{t_elapsed}','s')

Iteration 1, loss = 0.32484982

Iteration 2, loss = 0.15960803
Iteration 3, loss = 0.12264513
Iteration 4, loss = 0.10188910
Iteration 5, loss = 0.08910487
Iteration 6, loss = 0.08259557
Iteration 7, loss = 0.07394721
Iteration 8, loss = 0.06810929

Training set score: 0.987083

Test set score: 0.970700
One-layer neural network, training time= 27.78125 s

13.7.5 Performance analysis

When parameters are set as epochs = 10, max iter = 10, learning rate init =
0.1, at the 10th epoch, we obtain in the following:
Training set score: 0.9868, and Testset score: 0.9700.
490 Machine Learning with Python: Theory and Applications

This one-layer neural network needs only 30 s to train, which is about

30% faster than the random forest model. The accuracy is about the same
as the random forest model we used in the previous chapter.

13.7.6 Viewing the weight matrix as images

Sometimes, looking at the parameters learned in a trained neural network
can provide some insight into the learning behavior. For example, if the
weights look unstructured, these weights may not be very useful. If very
large weights are observed, the regularization may be too low, or the learning
rate may be too high. In the following, we plot weights of some of the first
layers in the MLPClassifier trained on the MNIST dataset. Let us look at
weights in the form of image. The input data in the MNIST dataset consist
of handwritten digits in an image with 28 × 28 pixels. Therefore, it has 784
(pixel) features in the dataset. The weight matrix in the first layer net shall
have a shape of (784, hidden layer sizes[0]). We reshape the single column
of 784 weights back to a 28 × 28 pixel image, so that we can show it.

#plt.figure(figsize=(50, 100))
fig, axes = plt.subplots(5, 10, figsize=(20, 10))
# use global min/max to ensure all weights are shown on
# the same scale
print(mlp.coefs_[0].shape,mlp.coefs_[1].shape)
vmin, vmax = mlp.coefs_[0].min(), mlp.coefs_[0].max()
for coef, ax in zip(mlp.coefs_[0].T, axes.ravel()):
ax.matshow(coef.reshape(28, 28), cmap=plt.cm.gray,
vmin=.5 * vmin, vmax=.5 * vmax)
ax.set_xticks(())
ax.set_yticks(())

plt.show()
(784, 50) (50, 10)

13.8 MLP for Nonlinear Regression

MLP is also a powerful tool regression. As we have discussed, MLP always

uses nonlinear activation functions in all these hidden layers. Therefore,
MPL regression is nonlinear and can handle complicated problems. Here,
we introduce a case study done by the scikit-learn team [12]. It studies
 Multilayer Perceptron (MLP) for Regression and Classification 491

Figure 13.5: Images of the weight matrix in the layer of the one-layer MLP trained on
the MNIST dataset.

the dependence between the target function and a set of “target” features,
marginalizing over the values of all other features (the complement features).
For possible visualization of the results, the target features should be one or
two. We thus choose the important target features to plot.
The study uses the California housing dataset [3]. The dataset has a total
of 8 features, and 4 of them are found relatively more influential to the house
price. We will build MLP models to reveal these influences. For comparison
purposes, we will also build a model of gradient boosting regressor to do the
same study.

# import necessary modules

%matplotlib inline
print(__doc__)
from time import time
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D

from sklearn.model_selection import train_test_split

from sklearn.preprocessing import QuantileTransformer
from sklearn.pipeline import make_pipeline
from sklearn.inspection import partial_dependence
from sklearn.inspection import plot_partial_dependence
from sklearn.experimental import enable_hist_gradient_boosting
from sklearn.ensemble import HistGradientBoostingRegressor
492 Machine Learning with Python: Theory and Applications

from sklearn.neural_network import MLPRegressor

from sklearn.tree import DecisionTreeRegressor
from sklearn.datasets import fetch_california_housing

Automatically created module for IPython interactive environment

13.8.1 California housing data and preprocessing

# get the dataset

cal_housing = fetch_california_housing()
X = pd.DataFrame(cal_housing.data, columns=cal_housing.
feature_names)
y = cal_housing.target
print(' data-points:',y.shape,'\n Some sample sof labels:', y)
# check the raw label data
data-points: (20640,)
Some sample sof labels: [4.526 3.585 3.521 ... 0.923 0.847 0.894]

y -= y.mean() # gradient boosting

# with 'recursion' method does not account for initial
# estimator (here the average target, by default)
X_train, X_test, y_train, y_test = train_test_split(X, y,
test_size=0.1, random_state=0) #split the data
X.head()
MedInc HouseAge AveRooms AveBedrms Population
AveOccup Latitude Longitude
0 8.3252 41.0 6.984127 1.023810 322.0
2.555556 37.88 -122.23
1 8.3014 21.0 6.238137 0.971880 2401.0
2.109842 37.86 -122.22
2 7.2574 52.0 8.288136 1.073446 496.0
2.802260 37.85 -122.24

3 5.6431 52.0 5.817352 1.073059 558.0

2.547945 37.85 -122.25
4 3.8462 52.0 6.281853 1.081081 565.0
2.181467 37.85 -122.25

There are a total of eight features, which means that the regression is over
an eight-dimensional space, and hence is it not possible to view all of them in
such a high-dimensional space. We thus plot in price-dependence curve one
 Multilayer Perceptron (MLP) for Regression and Classification 493

by one, or plot contour plots two by two. The target features for the plots are
4 relatively important features: median income (MedInc), average occupants
per household (AvgOccup), median house age (HouseAge), and average rooms
per household (AveRooms). We have added in also two less important features
“AveBedrms” and “Population” to the plots, for comparison purposes.
Note that this tabular dataset has very different value ranges for these
features. Because neural networks can be sensitive to features with varying
scales, preprocessing these numeric features is critical.

print(' data-points:',y.shape,'\n Some samples of labels:', y)

# check if zero centered
data-points: (20640,)
Some samples of labels: [ 2.45744183 1.51644183 1.45244183
... -1.14555817 -1.22155817 -1.17455817]

print(X_train.shape, X_test.shape, y_train.shape, y_test.shape)

(18576, 8) (2064, 8) (18576,) (2064,)

13.8.2 Configure, train, and test the MLP

Let us fit an MLPRegressor and compute single-variable partial dependence
plots.

print("Training MLPRegressor...")
tic = time ()
# a pipeline to scale the numerical input
# features and tuned the neural network size and learning rate.
mlp = make_pipeline(QuantileTransformer(),
MLPRegressor(hidden_layer_sizes=(50, 50),
learning_rate_init=0.01, early_stopping=True))
#configure the MLP
mlp.fit(X_train, y_train) #Train the MLP
print("Done in {:.3f}s".format(time() - tic))
print("Test R2 score: {:.2f}".format(mlp.score(X_test,
y_test))) # Check the performance
Training MLPRegressor...
Done in 20.779s
Test R2 score: 0.80

One can increase the complexity of an MLP for stronger learning ability.
However, to avoid overfitting, the complexity of the MLP should be in
accordance with the number of the data-points in the training set.
494 Machine Learning with Python: Theory and Applications

13.8.3 Compute and plot the partial dependence

Let us now compute and plot the partial dependence using the MLP.

print('Compute and plot the partial dependence...')

tic= time()
# If compute the 2-way PDP (5, 1), it would be much slower
# with the brute method.
features=['MedInc','AveOccup', 'HouseAge', 'AveRooms',
('AveOccup', 'HouseAge'),('AveBedrms', 'MedInc')]
#Select features to plot

plot_partial_dependence(mlp,X_train,features,n_jobs=3,
grid_resolution=20)
print("done in {:.3f}s".format(time() - tic))
fig = plt.gcf()
fig.suptitle('Partial dependence of house value on\n
non-location features''for the California\n
housing dataset, with MLPRegressor')
fig.subplots_adjust(wspace=0.4,hspace=0.4)

Compute and plot the partial dependence...

done in 60.972s

Figure 13.6: Results on price-feature dependence obtained using the MLP regressor in
Sklearn for the California housing dataset.
 Multilayer Perceptron (MLP) for Regression and Classification 495

The tick marks on the x-axis in the above plots represent the deciles of
the feature values in the dataset. We can observe the following findings:

• The median house price is approximately linearly dependent on the median

income (top left plot).
• The house price drops with the increase of the average occupants per
household (top middle plot).
• The house age does not have a strong influence on the house price, nor do
the average rooms per household.

The contour plots show the interactions of two target features on the
house price. For example, for average occupancy greater than two, the house
price is nearly independent of the house age. For average occupancy less than
two, there is some dependence.

13.8.4 Comparison studies on different regressors

For comparison purposes, we also perform regressions on the same dataset

using the GradientBoostingRegressor and the DecisionTreeRegressor at the
Scikit Learn package, and plot these results.

13.8.5 Gradient boosting regressor

Now, we train a model of gradient boosting regressor.

print("Training GradientBoostingRegressor...")
tic = time()
gbr = HistGradientBoostingRegressor()
gbr.fit(X_train, y_train)
print("done in {:.3f}s".format(time() - tic))
print("Test R2 score: {:.2f}".format(gbr.score(X_test, y_test)))

Training GradientBoostingRegressor...
done in 1.600s
Test R2 score: 0.85

For this tabular dataset, the Gradient Boosting Regressor is found both
significantly faster to train and more accurate on the test dataset, compared
to the MLP. It is also relatively easier to tune the hyperparameters, and the
defaults usually work well.
496 Machine Learning with Python: Theory and Applications

print('Computing partial dependence plots...')

tic = time()
features = ['MedInc', 'AveOccup', 'HouseAge', 'AveRooms',
('AveOccup', 'HouseAge'),('AveBedrms', 'MedInc')]
plot_partial_dependence(gbr, X_train, features,
n_jobs=3, grid_resolution=20)
print("done in {:.3f}s".format(time() - tic))
#fig = plt.figure(figsize=(45, 29))
fig = plt.gcf()
fig.suptitle('Partial dependence of house value on\n
non-location features' 'for the California\n
housing dataset, with Gradient Boosting')
fig.subplots_adjust(wspace=0.4, hspace=0.4)
Computing partial dependence plots...
done in 0.530s

Figure 13.7: Results on price-feature dependence obtained using the gradient boosting
regressor in Sklearn for the California housing dataset.

fig, ax = plt.subplots(figsize=(8, 4))

ax.set_title("Gradient Boosting Regressor")
gbr_disp=plot_partial_dependence(gbr,X_train,
["MedInc", "AveOccup"],ax=ax)
 Multilayer Perceptron (MLP) for Regression and Classification 497

Figure 13.8: Results using the gradient boosting regressor in Sklearn for the California
housing dataset.

fig, ax = plt.subplots(figsize=(8, 4))

ax.set_title("Multi-layer Perceptron")
mlp_disp = plot_partial_dependence(mlp, X, ["MedInc", "AveOccup"],
ax=ax, line_kw={"c": "red"})

Figure 13.9: Results using the MLP in Sklearn for the California housing dataset.

We observe that the MLP results are smooth.

498 Machine Learning with Python: Theory and Applications

13.8.6 Decision tree regressor

We next train a decision tree regressor.

print("Training DecisionTreeRegressor...")
tic = time()
tree = DecisionTreeRegressor()
tree.fit(X_train, y_train)
print("done in {:.3f}s".format(time()- tic))
print("Test R2 score: {:.2f}".format(tree.score(X_test, y_test)))

Training DecisionTreeRegressor...
done in 0.243s
Test R2 score: 0.61

fig, ax = plt.subplots(figsize=(8, 4))

ax.set_title("Decision Tree Regressor")
tree_disp=plot_partial_dependence(tree,X,["MedInc",
"AveOccup"],ax=ax, line_kw={"c": "m"})

Figure 13.10: Results using the decision tree regressor in Sklearn for the California
housing dataset.

fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(10, 5))

gbr_disp.plot(ax=[ax1, ax2], line_kw={"label":\
"Gradient Boosting Regressor"})
 Multilayer Perceptron (MLP) for Regression and Classification 499

mlp_disp.plot(ax=[ax1, ax2], line_kw={"label":\

"Multi-layer Perceptron", "c": "red"})
tree_disp.plot(ax=[ax1, ax2], line_kw={"label":
"Decision Tree Regressor", "c": "m"})
ax1.legend()
ax2.legend()

<matplotlib.legend.Legend at 0x18e35abb128>

Figure 13.11: Comparison of results using three regressors in Sklearn for the California
housing dataset.

Figure 13.11 shows that MLP prediction curves are much smoother than
the gradient boosting and the decision tree regressors.
Note that all the test results depend heavily on settings of the hyper-
parameters used to set up the model, datasets, computer system, and even
on when it is run. The setting parameters used in the above tests may not
be fair for some models. Thus, the test results are subjective.

References

[1] D.E. Rumelhart and J.L. McClelland, Learning internal representations by error
propagation, in Parallel Distributed Processing: Explorations in the Microstructure
of Cognition: Foundations, pp. 318–362, MIT Press, Cambridge, MA, 1987.
[2] J. Orbach, Principles of neurodynamics: Perceptrons and the theory of brain
mechanisms, Archives of General Psychiatry, 7(3), 218–219, 09 1962. https://doi.org/
10.1001/archpsyc.1962.01720030064010.
500 Machine Learning with Python: Theory and Applications

[3] T. Hastie, R. Tibshirani and J. Friedman, Elements of Statistical Learning Ed. p. 2,

Springer, New York, 2009.
[4] G.R. Liu, FEA-AI and AI-AI: Two-way deepnets for real-time computations for both
forward and inverse mechanics problems, International Journal of Computational
Methods, 16(08), 1950045, 2019.
[5] G.R. Liu, S.Y. Duan, Z.M. Zhang et al., Tubenet: A special trumpetnet for explicit
solutions to inverse problems, International Journal of Computational Methods,
18(01), 2050030, 2021. https://doi.org/10.1142/S0219876220500309.
[6] G.R. Liu and X Han, Computational Inverse Techniques in Nondestructive Evalua-
tion., Taylor and Francis Group, New York, 2003.
[7] D. Shuyong, W. Li, G.R. Liu et al., A technique for inversely identifying joint-
stiffnesses of robot arms via two-way TubeNets, Inverse Problems in Science &
Engineering, 29, 3041–3061, 2021.
[8] D. Shuyong, H. Zhiping, G.R. Liu et al., A novel inverse procedure via creating
tubenet with constraint autoencoder for feature-speace dimension-reduction, Inter-
national Journal of Applied Mechanics, Revised, 2021.
[9] L. Shi, F. Wang, G. Liu et al., Two-way tubeNets uncertain inverse methods for
improving positioning accuracy of robots based on interval, The 11th International
Conference on Computational Methods (ICCM2020), 2020.
[10] R. David, H. Geoffrey and W. Ronald, Learning representations by back-propagating
errors, Nature, 323(6088), 533–536, October 1986.
[11] H. Sepp and S. Jürgen, Long short-term memory, Neural Computation, 9(8), 1735–
1780, 1997. https://doi.org/10.1162/neco.1997.9.8.1735.
[12] P. Fabian, V. Gae, G. Alexandre et al., Scikit-learn: Machine Learning in Python,
Journal of Machine Learning Research, 12(85) 2825–2830, 2011. http://jmlr.org/
papers/v12/pedregosa11a.html.
[13] G.R. Liu, Neurons-Samples Theorem, International Journal of Computational Meth-
ods, 19(9), in-printing, 2022.
 Chapter 14

Overfitting and Regularization

14.1 Why Regularization

Regularization in machine learning is closely related to the so-called “over-

fitting” that happens often in practice in supervised training. Overfitting
depends on mainly two aspects: number of the training data-points m and
the dimension of the hypothesis space WP (the complexity of the neural
network or other machine learning model). A quality model should have just
enough complexity to capture the important features from the dataset and
not be sensitive to the noise in the data. If a model merely reproduces all or
most of the data-points in the training dataset, it may become overfitting.
Such a model is of little practical use because it cannot be generalized for
deployment. A rough general guideline may be given as follows:

m ≥ P∗ (14.1)

where P ∗ is calculated using Eq. (13.14) in Chapter 13. When m < P ∗ ,

the prediction by a fully trained model may, in theory, pass through all the
data-points, the labels are exactly produced, and the loss is zero. There may
be chances of overfitting and some type of regularization is needed.
For Eq. (14.1) to be an effective guide, we assume that the dataset is of
reasonably good quality (defined at the end of Section 5.9). For example,
if the dataset has too many parallel data-points, the number of the data-
points does not mean much. In such cases, m should count only these linearly
independent samples. In other words, there may still be overfitting, even if
Eq. (14.1) is satisfied, due to the insufficient quality of the dataset.

501
502 Machine Learning with Python: Theory and Applications

Figure 14.1: Fitting to data-points. Left: one data-point, infinite number of lines can be
used to fit this data; Right: two data-points, one straight line fits perfectly to these data-
points, but there are infinite numbers of higher-order curves can fit these two data-points.

Figure 14.1 illustrates schematically simple cases of overfitting. Consider

the simplest case with only one data-point. If one would use a linear line
(or any other types of curves) to fit perfectly the data-point, we shall have
an infinite number of lines. Any of these lines is unlikely predictive to any
other test data-point. To produce a robust predictive line, one would need
more data-points to generate a predictive line. A regularization technique
might be of help to obtain a unique line, but one would need to have some
information on the feature of the data, if the data are so scare.
Similarly, when there are two data-points, one straight line perfectly fits
these data-points, but there are infinite numbers of higher-order curves that
can fit these two data-points. All these high-order curves can be an overly
fitted prediction because it can predict test data-points incorrectly.
Overfitting can also occur in ML models for classifications. Figure 14.2
shows possible cases for given data-points in the 2D space. The orange line
gives an overly simplified boundary, the green line gives an overly fitted
boundary, and the black line gives a better fitted boundary for the given
data-points and possibly for the test data-points not used in training the
model.
Therefore, care is needed when using such an overly fitted machine
learning model for practical problems, because when applying it to other
data-points outside of the training dataset, or when the data-points have
noises, the performance is not going to be reliable. For a given dataset, we
may be able to train a model with extremely high accuracy, even with zero
loss. Such a model, however, cannot be used in general for prediction for
 Overfitting and Regularization 503

Figure 14.2: Fitting to find decision boundaries for classification problems. Orange line:
overly simplified boundary; Green line: overly fitted boundary; Black line: better fitted
boundary. Modified based on image from Wikimedia Commons by Chabacano (https://
en.wikipedia.org/wiki/Overfitting) under the CC BY-SA 4.0 license.

real-life problems. It is thus of importance to conduct objective tests: test

the trained model using a sufficient number of randomly selected data (from
the entire dataset), that was not used in training the model.
To overcome overfitting, various types of regularization techniques are
used. However, introducing regularization into a model gives some of
your biases to the features of your model. It thus unavoidably introduces
additional errors to the model. Hence, one should use it with care. To have
a regularization work well, one should also pay attention to the following
matters:

• Always be aware of the number of trainable parameters in the model. It

is the number of degrees of freedom of your model. When this number is
large, the model has a chance to be overly fitted.
• In contrast, the number of (independent) training examples plays a critical
role. When this number is too small compared to the P ∗ -number in the
model, the model has a chance to be overly fitted.
• Consider the number and quality of the training datasets you have and
then build a model accordingly, making sure that the number of the quality
samples is sufficiently large. The P ∗ -number is only a lower bound. In
504 Machine Learning with Python: Theory and Applications

practice, m can be much larger than the P ∗ -number. This lets the data
do the job. Use the regularization as a preventive tool or as a last resort.

Commonly used regularization techniques include the following:

1. Early stop during training: This is done by using a randomly chosen

independent validation dataset to monitor the training. When the
validation error comes across the training errors, consider stopping the
training.
2. Dropout: Simply randomly drop neurons along with their connections
from the net during training to limit the overfitting ability.
3. Tikhonov regularization that has a well-established mathematical
foundation.

The early-stop technique is straightforward and easy to implement. The

dropout technique needs to determine the ratio of neurons to be dropped
out. This will depend on the Neurons-Samples Theory and how many
successive training parameters there are in the model. A trial-and-error type
of approach may be required for the model to be trained. We will discuss
in great detail the Tikhonov regularization that has a clear mathematical
foundation.

14.2 Tikhonov Regularization

Tikhonov regularization is a widely used regularization technique. It is

named after the Russian scientist, Andrey Tikhonov. It is typically used for
regularizing the solutions for ill-posed problems, such as the inverse problems
in engineering [1]. It is also known as ridge regression in statistics. It is widely
used in machine learning for regularizing the training parameters.
Let us revisit the general formulation for linear regressions that we
presented at the beginning of Chapter 9, where we try to find solution for
ŵ by solving the following equation:

ŷ = Xŵ (14.2)

For given dataset X and label y, we built a loss function: L(y, ŷ(ŵ)) =
[y − Xŵ] [y − Xŵ]. We then went through an analytical minimization
process, and the solution is found in the following matrix form:

T T
ŵ∗ = (X X)−1 X y (14.3)
 Overfitting and Regularization 505

If one is careful, there will naturally be three important questions to ask:

T
1. Is X X invertible? What if not?
T
2. Assuming X X is invertible, and we do get the solution, is the solution
usable for a test dataset?
T
3. Assuming X X is invertible, and we do get the solution, how does the
solution behave when the test dataset is noisy?

We are interested in providing answers to these questions, and will see

where the Tikhonov regularization method can be of help.
When the Tikhonov regularization method (the simplest form) is used,
our formulation starts from re-writing the loss function:

L(y, ŷ(ŵ), λ) = [y − Xŵ] [y − Xŵ] + λ [ŵ − ŵ0 ] [ŵ − ŵ0 ] (14.4)

The first term on the right-hand side is the data loss. The 2nd term
is added as a regularization loss. λ is the regularization parameter that
is a pre-specified positive real number. Vector ŵ0 is also pre-specified for
the preferred the parameter ŵ. ŵ0 is often set to zero, because we usually
do not know what is preferred. If we have some knowledge on ŵ0 we can
put it in, and we shall have the shifted regularization. Hence, ŵ0 is also a
regularization parameter (more on this in the example studies).
Adding of the regularization term means that if the parameter ŵ departs
from the preferred vector, there will be penalty to the loss function. The
farther it departs, the more the penalty is scaled by the regularization
parameter. Therefore, a compromise needs to be arrived at to fit the data
and also with least regularization penalty.
We now go through the same analytical minimization process to find a
regularized solution that is the minimizer to the above loss function:

min L(y, ŷ(ŵ), λ) = min([y − Xŵ]T [y − Xŵ] + λ[ŵ − ŵ0 ] [ŵ − ŵ0 ]) (14.5)
ŵ ŵ

It is obvious that this function is also quadratic in ŵ. This time, we

can be quite sure that there is a unique ŵ∗ that minimizes the modified loss
function, regardless of the data X. This is because at least the regularization
term is strictly positive, as long as the λ is not zero. Such a regularized
solution can be found by setting the derivative of objective function to 0:

∂L(y, ŷ(ŵ), λ) T
= −2X (y − Xŵ∗ ) + 2λ[ŵ∗ − ŵ0 ] = 0 (14.6)
∂ ŵ
506 Machine Learning with Python: Theory and Applications

which gives a regularized solution:

∗ T T
ŵR = [X X + λI]−1 {X y − λŵ0 } (14.7)

which is a find of close-form solution.

Now, even if, for some reason (too small dataset, data-points are closely
T
correlated, etc.), matrix X X is singular or close to singular, λI is strictly
positive-definite regardless. In addition, by adjusting λ, one can make its
“positivity” sufficiently strong so that the matrix in the square blankets
is always invertible, to ensure the existence of a regularized solution. This
answers the first question.
Use of norms to add terms into the loss function is convenient because
the loss function needs to be a positive functional and norms by definition
satisfy the positivity requirement. Possible norms are
 P
 1q

norm = |wi |q (14.8)
i=0

where P is the total number of learning parameters. The term in the

parentheses in Eq. (14.8) multiplied by a regularization parameter is added
into the loss function. Figure 14.3 gives contours of the functions using these
norms defined in the foregoing equation for different q values.
In our modified loss function Eq. (14.4), we used the L2 norm for the
regularization term, which is quadratic and popular in current practices. It
has a good convex property, and is easy in formulation and implementation.
One may choose to use other norms, such as the L1 norm (called Lasso),
to regularize the absolute values of the training parameters. The L1 norm
function is still convex, but its derivatives are not continuous.
When the loss function is modified with the regularization term, we would
need to find the minimizer for the combined loss function. This situation is

Figure 14.3: Contourfs of the norms in simple two dimension.

 Overfitting and Regularization 507

Figure 14.4: Contours of both the data loss function and the regularization loss function.

schematically drawn in Fig. 14.4. The combined minimizer should be at the

kissing point of the contour of the data loss function and that of the regular-
ization loss function. When the quadratic regularization function is used, the
kissing point is usually with multiple nonzero feature parameters, unless the
data loss minimizer is located exactly on one of the feature axes, which is not
likely to happen. When the L1 norm is used, its contours have sharp vortexes.
Therefore, the oval contour of the data loss function can kiss at a vortex, even
if the data loss minimizer is not located exactly on any of the feature axes.
Because of this, the combined minimizer tends to select some of the features
and make the other features zero, creating a sparse model (with many zero
parameters) in the minimization process. Therefore, L1 regularization can
also be used for making the machine learning model sparse.
In Eq. (14.4), we have applied the same λ equally to all the parameters
in the above formulation. This does not have to be this way. If you know the
features of your model well, you may choose different λ values for individual
parameters. In this case, λ should be a diagonal matrix. In Python, one can
just define a vector to store these λs and use its element-wise products (as
will be shown in the example codes below) to apply different λ to different
features.
508 Machine Learning with Python: Theory and Applications

In some practices, regularizations are only applied to the weights not the
biases. This is equivalent to using a diagonal λ matrix, but set zero for these
diagonal terms corresponding to the biases. We will demonstrate this in the
following example too.
We are now ready to test out this method using a very simple example,
so that we can see through the process and feel how the regularization is at
work.

14.2.1 Demonstration examples: One data-point

Case 1: Consider the simplest possible one-dimensional problem with only

one data-point (x1 , y1 ). We would like to find a linear regression y = xW + b
based on this data-point. In this case, we shall have ŵ = [W 0 , W 1 ] = [b, W ] .
Clearly, the linear line has two model parameters: W and b. Thus, we can
have an infinite number of lines that satisfy the requirement if you use
Eq. (14.3). This can be clearly seen using Part I of the code attached
below, by setting xdata = np.arange(2, 5, 6), which gives only one data-
point x1 = 2.; with x0 = 1, we then have X = [1 x1 ] = [1., 2.].
Let also y = [2.].
Note that our code is written for multiple data-points. One can use just
one data-point to see what we get, and then try out using multiple data-
points.

import numpy as np
import random

# Part I code: No regularization, one data-point,

# two regression parameters, b and w

xdata = np.arange(2, 5, 6) # for 1 data-point

#xdata = np.arange(1, 5, 2) # for 2 data-points
#xdata = np.arange(1, 7, 2) # for 3 data-points

lxd = len(xdata)
One = np.ones(lxd)
print('Length of xdata=',lxd, ' Value of xdata=', xdata)
X = np.stack((One, xdata), axis=-1) # form X matrix
print('X=',X,X.shape)
 Overfitting and Regularization 509

np.random.seed(8)
y=xdata.T+np.random.randn(lxd)*0. # Label y, 1 data-point
#y=xdata.T+np.random.rand(lxd)*2. # Label y, >=3 data-points,
# using random number so that [x_i, y_i] will not be inline.
#y=2.*xdata.T+np.random.randn(lxd)*2.
# form y, for >= 3 data-points,
# using random number so that [x_i, y_i] will not be inline.
print('y=',y,y.shape)

XTy = np.dot(X.T,y) # Compute X.T * y

print('XTy=',XTy,XTy.shape)

XTX = np.matmul(X.T,X) # Compute X.T * X

print('XTX=',XTX,XTX.shape)

cd = np.linalg.cond(XTX) # c=np.linalg.solve(XTX,XTy) will not

# work for 1 data-point. Check condition number

print('Condition number=',cd,' Rank=',np.linalg.matrix_rank(XTX))

# check the rank also
if lxd >= 2:
w = np.linalg.solve(XTX,XTy) # only for data-points >=2.
print('w*=', w, w.shape)

Length of xdata= 1 Value of xdata= [2]

X= [[1. 2.]] (1, 2)
y= [2.] (1,)
XTy= [2. 4.] (2,)
XTX= [[1. 2.]
[2. 4.]] (2, 2)
Condition number= 4.804857307547117e+16 Rank= 1

It is clear that when only one data-point is used, the condition number
 
of the X X matrix is very large, and the rank of the X X matrix is only 1.
It is not solvable for ŵ∗ = [W ∗0 , W ∗1 ] = [b∗ , W ∗ ] . This means that there are
infinite number of [W 0 , W 1 ] = [b, W ] that can be the solution, which honor
the data.
We now use Eq. (14.7) to compute regularized solutions for this linear
regression, using different settings for the regularization parameter λ.
We used Part II of our code given below to perform the computation and
Part III to plot the results. Here, we shall have three cases to be studied, by
510 Machine Learning with Python: Theory and Applications

setting different values for lb0 and lw0 in vector L0=np.array([lb0, lw0]) in
the codes below. For all these cases, we will get a unique solution because
of the effects of the regularization. However, the solution behavior will also
be different, affected by the setting of the regularization parameters for b
and/or for W . Let us first examine Case 2.

# Part II code: Regularized linear regression, 2 parameters, b, w

# The code works for multiple data-points. Try one point first.
# For multiple data-points, use proper settings in Part I code.

w0 = np.array([.0,.0]) # Case 2,3,4; pre-specified parameters: b,w

#w0 = np.array([.4,.8]) # Case 5; with setting

# L0=np.array([1.,1.])
#w0 = np.array([.6,1.0]) # Case 5-1: 2 data, L0=np.array([1.,1.])
#w0 = np.array([.45341,1.]) # Case 5: 2 data, L0=np.array([1.,1.])
#w0 = np.array([.0,1.]) # Case 5: 2 data, L0 = np.array([1.,1.])
#w0 = np.array([1.8138,.99788]) # Case 5: 3 data,
# L0=np.array([1.,1.])
#w0 = np.array([0.65, 2.0]) # Case 5: 3 data, L0=np.array([1.,1.])

print('w0=',w0,w0.shape)

L0 = np.array([1.,0.]) # Case 2; # define base lambdas for b and w

# [lb0, lw0]) # One may try 3 cases: regularize b
# only set [1.,0]; [0.,1]; [1.,1.]

#L0 = np.array([0.,1.]) # Case 3; # regularize w only

#L0 = np.array([1.,1.]) # Case 4; # regularize both b and w

# Case 5; with w0 = np.array([.4,.8])

print('Lambda for b w =',L0, L0.shape)
L = 0.5 * L0 # set lambda value
I = np.identity(2)
print('I=',I,' L*I=',L*I) # regularization matrix lambda

K = XTX+L*I # regularized system matrix

XTy0 = XTy+L*w0 # regularized y vector

print('XTy=',XTy,XTy.shape,'w0=',w0,'XTy0=',XTy0)

crd = np.linalg.cond(K) # condition number of regularized

# system matrix. Should be small.
 Overfitting and Regularization 511

print('crd=',np.linalg.cond(K),' Rank=', np.linalg.matrix_rank(K))

# Check rank
wr = np.linalg.solve(K,XTy0) # regularized solution
print('wr=',wr)

w0= [0. 0.] (2,)

Lambda for b w = [1. 0.] (2,)
I= [[1. 0.]
[0. 1.]] L*I= [[0.5 0. ]
[0. 0. ]]
XTy= [2. 4.] (2,) w0= [0. 0.] XTy0= [2. 4.]
crd= 13.048362028288917 Rank= 2
wr= [0. 1.]

Note in the code above, that the input XTy is a 1D numpy array, and
the result is also a 1D array, which is not the convention of linear algebra, as
discussed in Section 2.9.13. One can also purposely define a column vector
(a 2D array with only on column) following the convention of linear algebra,
and get the solution. The solution will be the same, but is in a column vector.
Use of 1D numpy array will be more efficient. We have also seen an example
in Section 3.1.11.
Case 2: Using the same data as case 1 (one data-point x0 = 1, x1 =
2., X = [1 2.], y = [2.]) but set the regularization parameters as L0 =
np.array([1.0, 0.]), which regularizes only the bias b.
We write the following code to compute and plot the results.

# Part III: plot the regression linears.

%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
from matplotlib import rc

def y_hat(X, w): # function for the predicated regression

# line for given solution wr*
return np.dot(X, w) # X = [1 x1] ; wr* = [b* w*]^T

x = np.arange(0, 9, 1.) # points on x-axis

lx = len(x)
One = np.ones(lx)
512 Machine Learning with Python: Theory and Applications

print('Length of x=',lx, x)
Xm = np.stack((One, x), axis=-1) #build up X using points x.

#Lmda = np.arange(0.1, 1.5, 0.4) #control the range of lambda

#Lmda = np.arange(0.01, 15., 4)
Lmda = np.arange(0.1, 20., 6)

I = np.identity(2)
print('w0=', w0, w0.shape)
print('Lambda base for b, and w =',L0, L0.shape)
print('X=',X,X.shape)
print('y=',y,y.shape)

plt.figure(figsize=(4.5, 2.9),dpi=100)
for L in Lmda:
lmd = L*L0
K = XTX+lmd*I #regularized system matrix.
XTy0 = XTy +lmd*w0 #regularized y vector.
wr = np.linalg.solve(K,XTy0)
print('Lambda='+str(f"{L:.2f}"), ' wr=', wr,XTy.shape,\
w0.shape,XTy0.shape,L.shape)
plt.plot(x,y_hat(Xm, wr),label="Lambda="+str(f"{L:.2f}"))
# compute the predicted lines and plot out.

plt.scatter(xdata, y, s=35., c='r', alpha=0.5)

plt.xlabel('x')
plt.ylabel('$\hat{y}$')
plt.title('$\^{w}_0=$['+str(f"{w0[0]:.3f}")+','+\
str(f"{w0[1]:.3f}")+'], lb0='+str(L0[0])+',lw0 ='+str(L0[1]))
plt.grid(color='r', linestyle=':', linewidth=0.5)
plt.legend()
plt.show()

Length of x= 9 [0. 1. 2. 3. 4. 5. 6. 7. 8.]

w0= [0. 0.] (2,)
Lambda base for b, and w = [1. 0.] (2,)
X= [[1. 2.]] (1, 2)
y= [2.] (1,)
Lambda=0.10 wr= [0. 1.] (2,) (2,) (2,) ()
Lambda=6.10 wr= [0. 1.] (2,) (2,) (2,) ()
Lambda=12.10 wr= [0. 1.] (2,) (2,) (2,) ()
Lambda=18.10 wr= [0. 1.] (2,) (2,) (2,) ()
 Overfitting and Regularization 513

Figure 14.5: Simple linear regression for one data-point, with bias regularized at various
levels.

In this case, we regularize b only and leave W free (from regularization).

The results are plotted in Fig. 14.5. We shall observe the following
points:

• This solution would produce a zero solution for b, because it is regularized,

and some value for W , because it is allowed to change freely.
• As we can see, it is a line passing through the origin with a slope. This is
because the solution for b is zero, and W (responsible for the slope) is not.
• The regression line passes through the one data-point. This is because we
regularize one parameter and the other parameter is free. Therefore, we
have a total of two conditions: one is passing the data-point, and one is
zero bias b. Both conditions determine uniquely two regression parameters,
b and W .
• The magnitude of the λ has no effect on the solution, because of the setting
of this case with two conditions and two parameters. The amount of the
regularization is immaterial.
• It is clear that the Tikhonov regularization helped us obtain a unique
solution for the problem (Case 1) that has an infinite number of solutions.
This is why the Tikhonov regularization technique is also used to obtain
pseudo-inversion of an un-invertible matrix.

Using Part II and III of our code again, we now study Case 3:
Case 3: Use the same dataset as Case 2, but set L0 = np.array([0.0,
1.0]), which regularizes W 0 only.
514 Machine Learning with Python: Theory and Applications

Figure 14.6: The regression line for Case 3: only W is regularized, but the results are not
affected by the regularization parameter λ (these 4 lines are overlapping). b is set free from
regularization.

In this case, b is left free (from regularization). The results are shown in
Fig. 14.6. We shall observe the following points:

• This solution would produce a zero solution for W (hence the slope will be
zero), because it is regularized, and some value for b, because it is allowed
to change freely.
• As we can see, it is a horizontal line, as expected. This is because the
solution for W is zero and it is responsible for the slope.
• The line is no longer passing the origin, because b is no longer zero.
• The regression line still passes through the one data-point. This is because
we regularize one parameter and the other parameter is free. Therefore, we
have a total of two conditions: one is passing the data-point, and one is zero
slope W . Both conditions determine uniquely two regression parameters, b
and W .
• The magnitude of the λ has, again, no effect on the solution, because of
the setting with two conditions and two parameters. The amount of the
regularization is immaterial.
• It is clear that the Tikhonov regularization helped us again obtain a unique
solution for the problem (Case 1) that has an infinite number of solutions.
But, the solution is different because of the different way in which the
regularization is applied.

Using Part II and III of our code again, we now study Case 4 where we
regularize both b and W .
 Overfitting and Regularization 515

Figure 14.7: The regression line for Case 4: both b and W are all regularized.

Case 4: Use the same dataset as Case 2, but set L0 = np.array([1.0, 1.0])
in code Part II.
In this case, both b and W are all regularized, as shown in Fig. 14.7. We
shall observe the following points:

• This solution would produce a nonzero solution for both b and W , and
hence the intersection and the slope will all be nonzero. This is because
both cannot change freely.
• The regression line is no longer passing through the one data-point. This is
because we regularize both parameters. Therefore, we have a total of three
conditions: one is passing the data-point, one is a nonzero bias b, and one
is nonzero slope W . Because we have only two regression parameters b and
W , none of these conditions will be satisfied. The solution is a compromise.

• This time, the magnitude of the λ matters. When it is very small (0.1 for
example), the line is nearly passing through the data-point. Otherwise,
the line will not pass the data-point, for the same reason mentioned in the
above point.
• It is clear that the Tikhonov regularization helped us again obtain a unique
solution for the problem (Case 1) that has an infinite number of solutions.
But, the solution is different, depending on the level of the regularization
parameters used.

Figure 14.8 plots regression lines using small regularization parameters.

This re-emphasizes that the level of regularization is controlled by the
regularization parameter. Thus, setting proper a regularization parameter
is important.
516 Machine Learning with Python: Theory and Applications

Figure 14.8: The regression line for Case 4: both b and W are all regularized with a small
regularization parameter.

Figure 14.9: The regression line for Case 5: both b and W are all regularized, with respect
to the preferred solution set at the true solution of (0.4, 0.8).

Using our codes, we found that when we use a very small Lambda=0.01
for Case 4 where both b and W are all regularized, the resulting regularized
solution becomes [0.3992016 0.79840319]. We can then guess that the true
solution should be [0.4, 0.8]. Let us now study the following case.
Case 5: Set L0 = np.array([1.0, 1.0]) and ŵ0 = [0.4, 0.8] in the Part II
code.
In this case, both b and W are all regularized with respect to a preferred
solution that is the true solution. The results are plotted in Fig. 14.9.
We shall observe the following points:

• This solution becomes the true solution for both b and W , meaning that if
we know the true solution, we can put it in the regularization term, and
 Overfitting and Regularization 517

Figure 14.10: The regression line for Case 5: both b and W are all regularized, but with
respect to the preferred solution set at the true solution of (0.4, 0.8).

our code will reproduce it. The regression line passes through the data-
point. Figure 14.10 illustrates schematically what has happened in this
case.
• As shown in Fig. 14.10, these minimizers are for both the data loss and
the regularization loss, and thus the combined loss is located at the same
place (the true solution).
• In reality, it is not possible to know where the true solution for complex
problems is (if we do, we will not need any model). However, if we have
some but not exact knowledge about the true solution, using this shifted
regularization may be effective, because the trained model can be more
representative to the data, and yet biased to the knowledge. Because of
this, ŵ0 can also be a regularization parameter.
• Notice that ŵ0 is a vector, and hence the specification on it can be done
per parameter/feature.
• The magnitude of the λ does not matter in this case. This is because our
data loss and regularization loss are essentially the same for this simple
problem. When the data loss differs from the regularization loss (which is
true for practical complex problems), the magnitude of the λ shall have
effects.

14.2.2 Demonstration examples: Two data-points

With the understanding established for the simple examples using one data-
point, we can now extend the study to cases with two data-points, which shall
enforce our understanding of the regularization effects and the relationship
518 Machine Learning with Python: Theory and Applications

Figure 14.11: The regression line for Case 2 with 2 data-points: only b is regularized, and
W set free from regularization.

with the data. We shall do this in reference to all these 5 cases we have
studied for one data-point.
For Case 1, when we have two data-points, and still using the linear
regression model in one-dimensional space, we shall have a unique solution
as long as these two data-points are not coinciding. This is because the rank

of the system matrix X X is 2. The solution will be passing exactly these
two data-points. When these two data-points are X = [[1. 1.], [1. 3.]] and
y = [1.4534 3.4534] (note randomly generated), the true (un-regularized) is
ŵ∗ = [0.45341172, 1.] .
For Case 2 using two data-points, we also set L0 = np.array([1.0, 0.]).
In this case, we regularize b only and leave W free (from regularization). The
solution is plotted in Fig. 14.11.
We shall observe the following points:

• The regression line does not pass through these two data-points unless the
regularization parameter is set to zero. This is because we regularize one
parameter (and the other parameter is free). Therefore, we have a total
of three conditions: two for passing these two data-points, and one is for
zero bias b. The regularized solution is a compromise.
• When the regularization parameter is very small (0.01 for this case, blue
line in Fig. 14.11), the solution line closely passes these two points and
the result is close to Case 1 using two data-points.
• When the regularization parameter is large, the line is turned to give a
nearly zero bias b while minimizing the data loss (the sum of the distances
between these two data-points and the regression line).
 Overfitting and Regularization 519

Figure 14.12: The regression line for Case 3: only W is regularized, and b set free from
regularization.

• The magnitude of the λ has effects on the solution. The larger the λ, the
closer the b to zero.
• The Tikhonov regularization helped us obtain a compromise solution. It
uniquely depends on the data and the regularization parameter.

Case 3 for two data-points: Set L0 = np.array([0.0, 1.0]).

In this case, we regularize W only and leave b free (from regularization),
while the data loss is for both data-points. The results are plotted in
Fig. 14.12.
We shall observe the following points:

• The regression lines do not pass through these two data-points unless the
regularization parameter is set to zero. This is because we regularize one
parameter (and the other parameter is free). Therefore, we have a total
of three conditions: two for passing these two data-points, and one is for
small weight W . The regularized solutions are compromises.
• When the regularization parameter is very small (0.01 for this case, blue
line in Fig. 14.12), the solution line closely passes these two data-points
and the result is close to Case 1 using two data-points.
• When the regularization parameter is large, the line is turned to give
a small slope controlled by W while minimizing the data loss (the
sum of the distances between these two data-points and the regression
line).
• The magnitude of the λ has effects on the solution. The larger the λ, the
closer the slope W to zero.
520 Machine Learning with Python: Theory and Applications

Figure 14.13: The regression line for Case 4: both b and W are all regularized.

Case 4 with two data-points: Set L0 = np.array([1.0, 1.0]) in code Part

II, and using two data-points.
In this case, both b and W are all regularized. The result from our code
is plotted as follows.
We shall mention the following points:

• The regression line does not pass through these two data-points, unless the
regularization parameter is set to zero. This is because we regularize two
parameters. Therefore, we have a total of four conditions: two for passing
these two data-points, and two for small weight W . The regularized solution
is a compromise.
• When the regularization parameter is very small (0.01 for this case, blue
line in Fig. 14.13), the solution line closely passes these two data-points
and the result is close to Case 1 using two data-points.
• When the regularization parameter is large, the line is turned to give a
small slope controlled by W , and also shifted downward to give a smaller
intersection controlled by b, while minimizing the data loss (the sum of
the distances between these two data-points and the regression line).
• The magnitude of λ has effects on the solution. The larger the λ, the
smaller the intersection and the closer the slope W to zero.

From the study of Case 1 with two data-points, we obtained the true (un-
regularized) solution ŵ∗ = [0.4534, 1.] . We can then use the true solution
for a shifted regularization.
 Overfitting and Regularization 521

Figure 14.14: The regression line for Case 5 with two data-points: both b and W are all
regularized, with respect to the preferred solution set at the true solution of [0.4534, 1.0].

Case 5 with two data-points: Set L0 = np.array([1.0, 1.0]) and ŵ0 =

[0.4534, 1.0] in Part II code.
In this case, both b and W are all regularized, with respect to a preferred
solution that is the true solution. The results are plotted in Fig. 14.14.
We shall observe the following point:

• This solution becomes the true solution for both b and W , meaning that if
we know the true solution, we can put it in the regularization term, and
our code will reproduce it. The regression line passes through these two
data-points, regardless of the level of regularization.

14.2.3 Demonstration examples: Three data-points

We can now extend the study to cases with three data-points.

For Case 1, when we have three data-points, and are still using the linear
regression model in one-dimensional space, we shall have a unique solution,
as long as these three data-points are not coinciding. This is because the

rank of the system matrix X X become 2. However, the solution will not be
passing these three data-points exactly, because it is a least-square solution,
ensuring only the distances between these points to the linear line being
minimum. When these three data-points are X = [[1. 1.], [1. 3.], [1. 5.]]
and y = [2.7469 4.9379 6.7389] (randomly generated), the true least-square
solution (un-regularized) is ŵ∗ = [1.8138, 0.9979] .
522 Machine Learning with Python: Theory and Applications

Figure 14.15: The regression line for Case 2 with 3 data-points: only b is regularized, and
W set free from regularization.

For Case 2 using three data-points, we also set L0 = np.array([1.0,0.]).

In this case, we regularize b only and leave W free (from regularization). The
solution is plotted in Fig. 14.15.
We shall observe the following points:

• The regression lines do not pass through these three data-points, even if
the regularization parameter is set to zero. This is because we have a total
of four conditions/equations: these three data-points and one for zero bias
b. Therefore, the regularized solution is a compromise.
• When the regularization parameter is very small (0.01 for this case, blue
line in Fig. 14.15), the solution line is close to the un-regularized true
least-square solution.
• When the regularization parameter is large, the line is turned to give a
nearly zero bias b while minimizing the data loss (the sum of the distances
between these three data-points and the regression line).
• The magnitude of the λ has effects on the solution. The larger the λ, the
closer the b to zero.
• The Tikhonov regularization helped us obtain a compromise solution. It
uniquely depends on the data and the regularization parameter.

Case 3 for three data-points: Set L0 = np.array([0.0,1.0])

In this case, we regularize W only and leave b free (from regularization),
while the data loss is for all the data-points. The results are plotted in
Fig. 14.16.
 Overfitting and Regularization 523

Figure 14.16: The regression line for Case 3 with three data-points: only W is regularized,
and b set free from regularization.

We shall observe the following points:

• The regression lines do not pass through these three data-points, even if
the regularization parameter is set to zero. This is because we have a total
of four conditions/equations: these three data-points and one on slope W .
Therefore, the regularized solution is a compromise.
• When the regularization parameter is very small (0.01 for this case, blue
line in Fig. 14.16), the solution line is close to the un-regularized true
least-square solution.
• When the regularization parameter is large, the line is turned to give a
small slope controlled by W while minimizing the data loss (the sum of the
distances between these three data-points and the regression line).
• The magnitude of λ has effects on the solution. The larger the λ, the closer
the slope W to zero.
Case 4 with three data-points: Set L0 = np.array([1.0, 1.0]) in Part II
code.
In this case, both b and W are all regularized. The result from our code
is plotted as follows.
We shall mention the following points:
• The regression lines do not pass through these three data-points, even if
the regularization parameter is set to zero. This is because we have a total
of five conditions/equations: these three data-points, and one for small b
and the other for small slope W . Therefore, the regularized solution is a
compromise.
524 Machine Learning with Python: Theory and Applications

Figure 14.17: The regression line for Case 4 with three data-points: both b and W are all
regularized.

• When the regularization parameter is very small (0.01 for this case, blue
line in Fig. 14.17), the solution line is close to the un-regularized true
least-square solution.
• When the regularization parameter is large, the line is turned to give a
small slope controlled by W , and also shifted downward to give a smaller
intersection controlled by b, while minimizing the data loss.
• The magnitude of λ has effects on the solution. The larger the λ, the
smaller the intersection and the closer the slope W to zero.

From the study of Case 1 with three data-points, we obtained the true
(un-regularized) solution ŵ∗ = [1.8138, 0.9979] . We can then use the true
solution for a shifted regularization.

Case 5 with three data-points: Set L0 = np.array([1.0, 1.0]) and ŵ0 =

[1.8138, 0.9979] in Part II code.
In this case, both b and W are all regularized, with respect to a preferred
solution that is the true solution. The results are plotted in Fig. 14.18.
We shall observe the following point:

• This solution becomes the true least-square solution for both b and
W , meaning that if we know the true solution, we can the put it in

the regularization term, and our code will reproduce it. The regression
line passes through these three data-points, regardless of the level of
regularization.
 Overfitting and Regularization 525

Figure 14.18: The regression line for Case 5 with three data-points: both b and W are
all regularized, with respect to the preferred solution set at the true solution of [1.8138,
0.9979] .

14.2.4 Summary of the case studies

These studies show that the Tikhonov regularization has the following major
effects:

• For small datasets, it allows one to make use of knowledge of the data to
make up the shortfall in information to obtain a unique solution.
• For excessive datasets, it offers means to control the magnitudes of the
training parameters to prevent overfitting.
• For noisy datasets, the regularization offers means for mitigating the error.
Based on our study on the universal approximation theory, we found
that the weights and biases are responsible for producing the pulses for
approximating a function (see Section 7.3). We have also demonstrated
that the larger the weights, the steeper the pulse, hence offering more
capability in approximating more complex functions. If the weight values
are sufficiently large, it may even pick up the noises, which generally
have higher variation in frequencies. When the Tikhonov regularization is
applied, the values of the weights are generally reduced as shown explicitly
in these case studies. Therefore, the model will be less sensitive to noise.
• All these are done via the use of regularization parameters including the
following:
(1) Parameter λ or λ matrix of parameters to control the norms of all or
partial or individual features.
526 Machine Learning with Python: Theory and Applications

(2) Shifting parameter ŵ0 that can also be specified for individual
features, to allow the regularization loss to be more in alignment with
the data loss.
• Regularization parameters can be set differently based on the types of the
features.
• Other norms may be used for accounting the regularization loss.

14.3 A Case Study on Regularization Effects using MXNet

Based on the above detailed study using the simple examples, we shall have a
good understanding of how a regularization may work for a machine learning
model. We are ready for a study on its effects on a real machine learning
model. We now revisit the handwritten digit classification problem on the
MNIST dataset, which was studied earlier.
Note that our earlier models do not overfit because we used all 60,000
training examples, which is far more than the P ∗ -number and sometimes
even the P -number in those models. For example, we used a one-layer net
that has P = 7,850 parameters: 784 × 10 = 7,840 weights plus 10 biases, and
P ∗ = 785 (which is the pseudo-dimension of the affine space for the single
layer). Thus, m  P ∗ and m  P .
Let us do a case study to examine what might happen if we use fewer
training samples, say 2,000. For the one-layer net, we have m > P ∗ but
m  P . This study will show in detail the effects of the regularization tech-
nique on mitigating the over-fitting. It is done following the study and using
the code at mxnet-the-straight-dope (https://github.com/zackchase/mxnet-
the-straight-dope/blob/master/chapter02 supervisedlearning/regularization-
scratch.ipynb), under the Apache-2.0 License.

from __future__ import print_function

import mxnet as mx
import mxnet.ndarray as nd
from mxnet import autograd
import numpy as np
ctx = mx.cpu()
mx.random.seed(1)
# for plotting purposes
 Overfitting and Regularization 527

%matplotlib inline
import matplotlib
import matplotlib.pyplot as plt

14.3.1 Load the MNIST dataset

mnist = mx.test_utils.get_mnist()
num_examples = 2000
batch_size = 64
train_data = mx.gluon.data.DataLoader(mx.gluon.data.ArrayDataset(
mnist["train_data"][:num_examples], mnist["train_label"]
[:num_examples].astype(np.float32)), batch_size, shuffle=True)

test_data = mx.gluon.data.DataLoader(mx.gluon.data.ArrayDataset(\
mnist["test_data"][:num_examples], mnist["test_label"]
[:num_examples].astype(np.float32)), batch_size, shuffle=False)

14.3.2 Define a neural network model

We use a simple linear model with the softmax as activation as the prediction
function. We set up the following neural network model:

W = nd.random_normal(shape=(784,10)) #Allocate model parameters

b = nd.random_normal(shape=10)
params = [W, b]

for param in params:

param.attach_grad() # request for gradient

def net(X):
y_linear = nd.dot(X, W) + b # Define the affine mapping
yhat = nd.softmax(y_linear, axis=1)
return yhat # return the prediction function

14.3.3 Define loss function and optimizer

We use the cross entropy, which we have discussed in detail earlier, as the
loss function.
528 Machine Learning with Python: Theory and Applications

def cross_entropy(yhat, y):

return - nd.sum(y * nd.log(yhat), axis=0, exclude=True)

def SGD(params, lr):

for param in params:
param[:] = param - lr * param.grad

14.3.4 Define a function to evaluate the accuracy

We use the following function to calculate the average loss function as a

measure of the accuracy of the model during training:

def evaluate_accuracy(data_iterator, net):

numerator = 0.
denominator = 0.
loss_avg = 0.
for i, (data, label) in enumerate(data_iterator):
data = data.as_in_context(ctx).reshape((-1,784))
label = label.as_in_context(ctx)
label_one_hot = nd.one_hot(label, 10)
output = net(data)
loss = cross_entropy(output, label_one_hot)
predictions = nd.argmax(output, axis=1)
numerator += nd.sum(predictions == label)
denominator += data.shape[0]
loss_avg=loss_avg*i/(i+1)+nd.mean(loss).asscalar()/(i+1)
return (numerator / denominator).asscalar(), loss_avg

14.3.5 Define a utility function plotting convergence curve

The following function is used to plot out the change of the loss functions
against the iterations during the minimization process:

def plot_learningcurves(loss_tr,loss_ts, acc_tr,acc_ts):

xs = list(range(len(loss_tr)))
f = plt.figure(figsize=(12,6))
fg1 = f.add_subplot(121)
fg2 = f.add_subplot(122)
 Overfitting and Regularization 529

fg1.set_xlabel('epoch',fontsize=14)
fg1.set_title('Comparing loss functions')
fg1.semilogy(xs, loss_tr)
fg1.semilogy(xs, loss_ts)
fg1.grid(True,which="both")
fg1.legend(['training loss', 'testing loss'],fontsize=14)

fg2.set_title('Comparing accuracy')
fg1.set_xlabel('epoch',fontsize=14)
fg2.plot(xs, acc_tr)
fg2.plot(xs, acc_ts)
fg2.grid(True,which="both")
fg2.legend(['training accuracy', 'testing accuracy'],
fontsize=14)

14.3.6 Train the neural network model

We now train the neural network model to its fullest learning ability.

epochs = 1000
moving_loss = 0.
niter=0
loss_seq_train = []
loss_seq_test = []
acc_seq_train = []
acc_seq_test = []

for e in range(epochs):
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(ctx).reshape((-1,784))
label = label.as_in_context(ctx)
label_one_hot = nd.one_hot(label, 10)
with autograd.record():
output = net(data)
loss = cross_entropy(output, label_one_hot)
loss.backward()
SGD(params, .001)
# Keep a moving average of the losses
niter +=1
moving_loss=.99*moving_loss+.01*nd.mean(loss).asscalar()
est_loss = moving_loss/(1-0.99**niter)
530 Machine Learning with Python: Theory and Applications

test_accuracy, test_loss = evaluate_accuracy(test_data, net)

train_accuracy, train_loss=evaluate_accuracy(train_data, net)
# save them for later
loss_seq_train.append(train_loss)
loss_seq_test.append(test_loss)
acc_seq_train.append(train_accuracy)
acc_seq_test.append(test_accuracy)

if e % 300 == 299:
print("Epoch %s. Train Loss:%.2e, Test Loss:%.2e,
Train_acc:%.2e, Test_acc:%.2e" %
(e+1,train_loss, test_loss,train_accuracy,
test_accuracy))

# Plotting the learning curves

plot_learningcurves(loss_seq_train,loss_seq_test, acc_seq_train,
acc_seq_test)

Epoch 300. Train Loss:1.61e-01, Test Loss:1.16e+00,

Train acc:9.64e-01, Test acc:7.75e-01
Epoch 600. Train Loss:6.72e-02, Test Loss:1.15e+00,
Train acc:9.90e-01, Test acc:7.83e-01
Epoch 900. Train Loss:3.73e-02, Test Loss:1.16e+00,
Train acc:9.98e-01, Test acc:7.85e-01

Figure 14.19: A case of over-fitting when the one-layer net is used for a reduced dataset
of MNIST that has only 2,000 samples.
 Overfitting and Regularization 531

14.3.7 Evaluation of the trained model: A typical case

of overfitting

In the above example, it is seen that at the 1,000th epoch, the MLP model
achieves 100% accuracy on the training data (the loss function is still quite
far from zero, partially because m > P ∗ ). However, it classifies only ∼79%
of the test data accurately. This is a typical case of overfitting. Clearly, such
a “well” trained model is not possible to perform well on the test dataset or
dataset in the real world the model did not seen during the training.
If the sample is reduced to 1,000 with all other settings unchanged, this
MLP model would achieves 100% accuracy on the training data at the 700th
epoch, but it classifies only ∼75% of the test data accurately.

14.3.8 Application of L2 regularization

We now apply the L2 regularization by just adding in a penalized L2 norm

of the parameters to the original loss function. We use the following function
to do so:

def l2_penalty(params):
penalty = nd.zeros(shape=1)
for param in params:
penalty = penalty + nd.sum(param ** 2)
return penalty

14.3.9 Re-initializing the parameters

We re-initialize the parameters randomly to erase these in the trained model.

for param in params:

param[:] = nd.random_normal(shape=param.shape)

14.3.10 Training the L2-regularized neural network model

epochs = 1000
moving_loss = 0.
l2_lambda = 0.1 #0.3 # regularization parameter
niter=0
532 Machine Learning with Python: Theory and Applications

loss_seq_train = []
loss_seq_test = []
acc_seq_train = []
acc_seq_test = []

for e in range(epochs):
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(ctx).reshape((-1,784))
label = label.as_in_context(ctx)
label_one_hot = nd.one_hot(label, 10)
with autograd.record():
output = net(data)
loss = nd.sum(cross_entropy(output, label_one_hot))+\
l2_lambda*l2_penalty(params) # regularization
# loss added
loss.backward()
SGD(params, .001)

# Keep a moving average of the losses

niter +=1
moving_loss=.99*moving_loss+.01*nd.mean(loss).asscalar()
est_loss = moving_loss/(1-0.99**niter)

test_accuracy, test_loss = evaluate_accuracy(test_data, net)

train_accuracy,train_loss= evaluate_accuracy(train_data, net)

# save them for later

loss_seq_train.append(train_loss)
loss_seq_test.append(test_loss)
acc_seq_train.append(train_accuracy)
acc_seq_test.append(test_accuracy)
if e % 300 == 299:
print("Epoch %s. Train Loss:%.2e, Test Loss:%.2e,
Train_acc:%.2e, Test_acc:%.2e" %
(e+1,train_loss,test_loss,train_accuracy, test_accuracy))

# Plotting the learning curves

plot_learningcurves(loss_seq_train,loss_seq_test,acc_ seq_train,
acc_seq_test)
print('Regularization parameter=',l2_lambda)

Epoch 300. Train Loss:1.93e-01, Test Loss:5.19e-01,

Train acc:9.65e-01, Test acc:8.38e-01
Epoch 600. Train Loss:1.91e-01, Test Loss:5.03e-01,
Train acc:9.67e-01, Test acc:8.45e-01
 Overfitting and Regularization 533

Epoch 900. Train Loss:1.91e-01, Test Loss:5.03e-01,

Train acc:9.67e-01, Test acc:8.46e-01
Regularization parameter= 0.1

14.3.11 Effect of the L2 regularization

When we use 2,000 training samples, addition of the L2 regularization
resulted in the following effects, compared to the unregulated overfitted case:
• When the regularization parameter is set at 0.1 with all other settings
unchanged, at the 1,000th iteration, the accuracy of our MLP model on the
test dataset increased from 78.6% to 84.5%, a 5.7% increase. The accuracy
on the training dataset is reduced to 94.7% percent from 100%. The reg-
ularized model is more robust and shall generalize better for unseen data.
• When the regularization parameter is set at 0.3, the accuracy on the test
dataset increased from 78.6% to 84.4%, a 5.8% increase. The accuracy
on the training dataset is reduced to 93.6% percent from 100%. This
regularized model is about unchanged. This suggests that a regularization
parameter in between 0.1 and 0.3 may be a proper choice for this problem.
When only 1,000 training samples are used in the training (with all the
other settings unchanged in the above code), the findings are as follows,
compared to the corresponding unregulated overfitted case:
• For the regularization parameter set at 0.1, the accuracy of our MLP
model on the test dataset increased from 75% to 83.3%, an 8.3% increase.

Figure 14.20: The L2 regularization has reduced the gap between the training and testing
accuracy. One-layer net using the reduced dataset of MNIST that has only 2,000 samples.
534 Machine Learning with Python: Theory and Applications

The accuracy on the training dataset is reduced to 99.1% percent from

100%. The regularized model is more robust and shall generalize better
for unseen data.
• When the regularization parameter is set at 0.3, the accuracy on the test
dataset increased from 75% to 84.9%, a 9.9% increase. The accuracy on the
training dataset is reduced to 96.7% percent from 100%. This regularized
model is even more robust for unseen data compared to the case of using a
regularization parameter of 0.1. This may suggest that when the number of
samples becomes small, the regularization parameter should be increased
accordingly.

Ideally, we would want a model having the same level of accuracy for both
the training and test datasets. Tuning the regularization parameter can help
to achieve this goal, but it can often be a challenging task. Readers may give
it a try using the above code.
We finally note that the choice of the regularization parameter is an issue.
Experiences may help, or trial and error may be needed to train a robust
model. Below is another case study.

14.4 A Case Study on Regularization Parameters Using Sklearn

We now introduce another case study done by the scikit-learn team [2] for
classification to investigate the influence of the regularization parameters on
the predicted decision boundary. We have made minor changes to the original
code to fix issues possibly caused by module version differences. We have also
done some tuning on the parameters. This study is on the effects of the regu-
larization parameter “alpha” (equivalent to the lambda in our formulation).
This study uses synthetic datasets for classification, and the plots will show
different decision boundaries produced when using different alpha values.

%matplotlib inline
print(__doc__)
# Author: Issam H. Laradji, License: BSD 3 clause
import numpy as np
from matplotlib import pyplot as plt
from matplotlib.colors import ListedColormap
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.datasets import make_moons,make_circles,
make_classification
from sklearn.neural_network import MLPClassifier
 Overfitting and Regularization 535

h = .02 # step size in the mesh

alphas = np.logspace(-5, 3, 5)
names = ['alpha ' + str(i) for i in alphas]

classifiers = []
for i in alphas:
classifiers.append(MLPClassifier(solver='lbfgs', alpha=i,
max_iter=1500,random_state=1, hidden_layer_sizes=[100, 100]))

X, y = make_classification(n_features=2, n_redundant=0,
n_informative=2, random_state=0, n_clusters_per_class=1)
rng = np.random.RandomState(2)
X += 2 * rng.uniform(size=X.shape)
linearly_separable = (X, y)

datasets = [make_moons(noise=0.2, random_state=0),

make_circles(noise=0.1, factor=0.5, random_state=1),
linearly_separable]

figure = plt.figure(figsize=(17, 9))

i = 1
# iterate over datasets
for X, y in datasets:
# preprocess dataset, split into training and test part
X = StandardScaler().fit_transform(X)
X_train,X_test,y_train,y_test=train_test_split(X,y,
test_size=.4)

x_min, x_max = X[:, 0].min() - .5, X[:, 0].max() + .5

y_min, y_max = X[:, 1].min() - .5, X[:, 1].max() + .5
xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
np.arange(y_min, y_max, h))
# just plot the dataset first
cm = plt.cm.RdBu
cm_bright = ListedColormap(['#FF0000', '#0000FF'])
ax = plt.subplot(len(datasets), len(classifiers) + 1, i)
# Plot the training points
ax.scatter(X_train[:,0],X_train[:, 1], c=y_train )
# and testing points
ax.scatter(X_test[:,0],X_test[:,1], c=y_test, alpha=.4)
ax.set_xlim(xx.min(), xx.max())
ax.set_ylim(yy.min(), yy.max())
ax.set_xticks(())
ax.set_yticks(())
i += 1
536 Machine Learning with Python: Theory and Applications

# iterate over classifiers

for name, clf in zip(names, classifiers):
ax=plt.subplot(len(datasets),len(classifiers)+1,i)
clf.fit(X_train, y_train)
score = clf.score(X_test, y_test)
# Plot the decision boundary. Assign a color to each
# point in the mesh [x_min, x_max]x[y_min, y_max].
if hasattr(clf, "decision_function"):
Z = clf.decision_function(np.c_[xx.ravel(),
yy.ravel()])
else:
Z = clf.predict_proba(np.c_[xx.ravel(),
yy.ravel()])[:,1]

# Put the result into a color plot

Z = Z.reshape(xx.shape)
ax.contourf(xx, yy, Z, cmap=cm, alpha=.8)

# Plot also the training points

ax.scatter(X_train[:,0],X_train[:,1], c=y_train,
cmap=cm_bright,edgecolors='black')
# and testing points

ax.scatter(X_test[:,0], X_test[:,1], c=y_test,

cmap=cm_bright,alpha=0.4, edgecolors='black')

ax.set_xlim(xx.min(), xx.max())
ax.set_ylim(yy.min(), yy.max())
ax.set_xticks(())
ax.set_yticks(())
ax.set_title(name)
ax.text(xx.max()-.3,yy.min()+.3,
('%.2f' % score).lstrip('0'), size=15,
horizontalalignment='right')
i += 1

figure.subplots_adjust(left=.02, right=.98)
plt.show();

Automatically created module for IPython interactive environment

Figure 14.21 shows the effects of the regularization parameters used.

 Overfitting and Regularization 537

Figure 14.21: Effects of regularization parameters used in an MLPclassifier, based on a

study done by Sklearn team.
538 Machine Learning with Python: Theory and Applications

It is found that a larger alpha (that suppresses large weights) results in

less sharper decision boundaries, which helps to avoid overfitting. On the
other hand, a smaller alpha leaves larger room for the weights, resulting
in sharper decision boundaries, which helps to mitigate underfitting. This
study shows again that it is important to choose a proper regularization
parameter, which can often be a challenge.

References
[1] G.R. Liu and X. Han, Computational Inverse Techniques in Nondestructive Evaluation,
Taylor and Francis Group, New York, 2003.
[2] P. Fabian, V. Gae, G. Alexandre et al., Scikit-learn: Machine learning in Python,
Journal of Machine Learning Research, 12(85), 2825–2830, 2011. http://jmlr.org/
papers/v12/pedregosa11a.html.
 Chapter 15

Convolutional Neural Network (CNN)

for Classification and Object Detection

Convolutional neural network (CNN) or ConvNet [1–6] is one of the most

powerful tools in machine learning, especially for processing multi-
dimensional images and image-based classifications. It is currently widely
used, and effective modules are made available at many packages/libraries,
such as Scikit learn, PyTouch, TensorFlow, and Keras. The CNN uses
quite a number of special techniques, which have very much different net
configurations, compared to the neural networks discussed in the previous
chapters. CNNs still use affine transformations, but these are specially
designed with spatial filters to capture the spatial features of images. These
details are often hidden in the black boxes of the modules available at
various libraries. This chapter offers a concise description of the key ideas and
techniques used in CNN and then a demonstration of CNN. More detailed
information can be found in the rich open literature and the original papers
given for each CNN network. An excellent online course by Andrew Ng is
also available.
Our discussion starts from a brief description of the general structure of
CNN, some landmark CNN net, a demonstration of image classification, and
finally some examples of object detection using CNN.

15.1 Filter and Convolution

As discussed in Chapter 5, the most fundamental hypothesis made in a

machine learning model is the affine transformation. For different types of
problems, one naturally needs to design proper affine transformation units.
CNN works particularly well for image processing related problems, because

539
540 Machine Learning with Python: Theory and Applications

the key techniques used are rooted in computer vision-related techniques.

Filtering in a convolutional manner is one of those. Consider an idealized
“image” with 6 × 6 pixels, as shown in Fig. 15.1.

Figure 15.1: A 3 × 3 filter (in the middle of the figure) acting like an edge detector that
captures the vertical edge line in the middle of the 6 × 6 pixel image as a 4 × 4 bright
strap, when the filter slides over and within the 6 × 6 pixel image.

The filter (sometimes called kernel) in used in Fig. 15.1 is a 3 × 3 2D

matrix in the red frame in the middle of the figure. It has a 1-vector on the
left, a 0-vector in the middle, and a -1-vector on the right. The convolution
operator is denoted by “*” that performs the following operation with the
elements of the 2D sub-matrix in the blue frame in the 6 × 6 pixel image,
with the red frame sliding pixel by pixel over and within the 6 × 6 image.
In many machine learning modules, it is termed conv2D.
3 
 3
xij × wij = value in a green box corresponding to the filter center
i=1 j=1 (15.1)

The above computation is a “kind” of “dot-product” of the blue matrix

and the red matrix (of the same shape): summation of all the element-wise
products in these two matrices. Such a sliding (convolution) “2D-matrix-
dot-product” produces a 4 × 4 pixel image. It captures the vertical edge line
in the middle of the 6 × 6 pixel image as a 4 × 4 image with a bright strap
 Convolutional Neural Network (CNN): Classification and Object Detection 541

in the middle. Therefore, it acts like a vertical edge detector. We can expect
the following:

• If we rotate the red filter matrix, we shall have a horizontal edge detector.
• If we rotate the red filter matrix in an arbitrary angle, we shall have an
inclined edge detector.
• One can have many other types of filters, such as the “Sobel” filter and
“Scharr” filter.

Essentially, the filter is a matrix with desirable values in its entry.

Changing these values changes the functionality of the filter. In a CNN,
the key here is that we are not using this kind of manually designed filter.
Instead, we will make the network learn these values in the filter by feeding
it with a large number of images, as will be discussed in the next section.
Here, let us also note the following:

• In the above case, the filter slides per-pixel, or the “stride” is 1. One may
have multiple strides, such as two strides per sliding.
• The filter size can also change, and 5 × 5 (but usually an odd number, in
order to always have a central pixel) is also an often used filter size.
• The above convolution results in a reduction in size in the filtered image.
To avoid this, one may use a technique call “padding”: extending the
original 6 × 6 pixel image by filling in zeros to extend pixels on its
boundaries, and make the original image bigger, so that the filtered image
can have the same size.

Figure 15.2 is a picture made (https://github.com/vdumoulin/conv

arithmetic) by Dumoulin and Visin [7] under the MIT License.

Figure 15.2: A 3 × 3 filter (gray) sliding over a 5 × 5 pixel image (blue) with stride=1
and padding=1 (padding part is not colored). It generates a cyan image of the same (5 ×
5) size as the blue image.
542 Machine Learning with Python: Theory and Applications

The settings for different strides and paddings are shown in Fig. 15.3.

Figure 15.3: In the upper row, from left to right: f=3,s=1,p=0; f=4,s=1,p=2;
f=3,s=1,p=1. In the bottom row, from left to right: f=3,s=2,p=0; f=3,s=2,p=1;
f=3,s=2,p=1.

where f stands for filter, s for stride, and p for padding. Readers may
take a look at cartoons and pictures for more cartoons to get a better
understanding.

15.2 Affine Transformation Unit in CNNs

In a CNN, we do not use the handcrafted filters. Instead, we let elements

in the filter matrix be trainable parameters (the w and b). This may be
one of the most important tricks in CNN: we let the sample-based training
eventually decide what kind of filters to use, in order to get the correct label
from the training samples. This parametrization of the filter matrix creates
a special ATU for CNNs. A CNN model for one channel of covn2d in a
convolution layer is schematically illustrated in Fig. 15.4.
We mention the following key points for CNN models:

• The same set of weights and bias (total of 10 learning parameters in

this case) is used for filtering the entire image, regardless of the size of
the image. This is known as parameter sharing. Therefore, the training
parameters in a CNN are usually much smaller compared to the densely
 Convolutional Neural Network (CNN): Classification and Object Detection 543

Figure 15.4: Schematic of a typical affine transformation unit (ATU) wrapped with an
activation function. Learning parameters, weights wi , and bias b, in a single channel of
a conv2D using a 3 × 3 filter in one layer of CNN. φ stands for an activation function
(usually the Relu). For a 3 × 3 filter, we have a total of 9 weights, and one bias, resulting
in a hypothesis space W10 for the ATU. For this 6 × 6 pixel image with 1 stride and no
padding, the output is a 4 × 4 image. This output image may be subjected to a “pooling”
and then to another layer of convolution.

connected neural networks (such as the standard MLP). This leads to a

drastic reduction of learning parameters. Based on the Neurons-Samples
Theory given in Section 13.2.2, we know the number of the learning
parameters in an MLP can be huge, because P is usually much larger
than P ∗ . Hence, there is plenty room for reduction.
• A CNN is less prone to overfitting because of the reduction in learning
parameters.
• The parameter sharing works particularity well for image processing,
because the same filter is capable of capturing the features (such as edges)
in different parts of the same image.
• The parameter sharing leads to a sparse connection: each set of the
learning parameters connects only to very small part of the input image at
a time. This is also very much different from the densely connected neural
networks.
• One often uses multiple channels (it is also termed depths in the literature)
in one convolution layer. The number of learning parameters will then
increase proportionally with the increasing channel number. This enables
the training of different kinds of filters (one channel is a different filter
with its own w) in a CNN convolution layer, to capture different features
in the image.
544 Machine Learning with Python: Theory and Applications

15.3 Pooling

Another often used technique in a CNN is pooling. It uses a similar sheet

like the filter (but no parameters involved), called pooling filter, to cover a
part of the filtered image. There are two major types of pooling operations:
max pooling and average pooling. Max pooling simply takes the maximum
pixel value among all the pixels that the pooling filter covers, as shown in
Fig. 15.5.
Average pooling is very similar to max pooling, but it takes the average
of the covered pixels, instead of the maximum, as shown in Fig. 15.6.
For pooling operations, we note the following:

• There is no training parameter introduced in the pooling operation. Hence,

it may be treated as a part of one CNN layer.
• It seems that max pooling is capable of getting out the outstanding
features from the image, and hence it is used more often than average
pooling.

Figure 15.5: Max pooling in a CNN: a 2 × 2 pooling filter covers 2 × 2 pixels in a filtered
4 × 4 image. It takes the maximum pixel value from the 2 × 2 pixels of each of these four
covers, and forms a 2 × 2 new image.

Figure 15.6: Average pooling in a CNN: a 2 × 2 pooling filter covers 2 × 2 pixels in a

filtered 4 × 4 image. It gets the average pixel value of these covered 2 × 2 pixels. This is
done for each of these four covers, resulting in a 2 × 2 new image.
 Convolutional Neural Network (CNN): Classification and Object Detection 545

Figure 15.7: Up sampling in a CNN: an image with 2 × 2 pixels is expanded to a new

one with 4 × 4 pixels. In this case, the value of a pixel in the original image is copied four
times in the corresponding locations in the enlarged image.

• One may also take a different number of strides, when covering the image.
• One may also use a different size of covers.

15.4 Up Sampling

Another often used technique in a CNN is up sampling (or up pooling).

It enlarges images by expanding one pixel to four with the same value, as
shown in Fig. 15.7. This is used when there is a need to enlarge an image
such as in the decoder part of an Autoencoders (that will be discussed in
Chapter 17).
Note that there are quite a number of other similar operations in image
processing. With the basic ones introduced above, we are now ready to
configure a CNN network.

15.5 Configuration of a Typical CNN

The typical configuration of a CNN is shown in Fig. 15.8. An image in

black-white or in RGB is first fed to a set of filters (F) consisting of a
number (6–64) of filter channels, each having its own set of weights and a
bias. The filter size may be 3 × 3, 5 × 5 or another odd number in one
direction. The filtered image may be subjected to a possible pooling (P)
operation that is usually of size 2 × 2. This reduces (in half) both the height
and width, but the number of the channels is kept the same. The pooling
performs fixed operations: either max pooling or average pooling. It does
not introduce any additional training parameters. Another set of filters may
be applied, which further increases the number of filter channels. If another
546 Machine Learning with Python: Theory and Applications

Figure 15.8: Configuration of a CNN: an RGB image is first fed to a set of filters (F)
and is then subjected to a possible pooling (P) operation. Another set of filters may be
applied, followed by another possible pooling. An F-P pair may consist of a CNN layer.
The output of the final CNN layer may be fed to few normal dense layers producing a
prediction.

pooling is applied, it reduces the height and width again, but keeps the
number of channels unchanged. Usually, an F-P pair may consist of a CNN
layer, and we can have quite a large number of CNN layers. The output of
the final CNN layer may (or may not) be fed to few normal dense layers,
and then produces a prediction. The following operation is essentially the
same as those in an MLP.
The network in Fig. 15.8 has 1 input layer, 2 CONV layers, 1 hidden
dense layer, and 1 output dense layer. In practice, many CNNs have a large
number of CONV layers.

15.6 Some Landmark CNNs

With the building blocks we have discussed so far, one can build a really
complex neural network for very complicated tasks of computer vision and
image-based object detection. The development in this area is very fast
and many useful special networks have already been proposed. This section
discusses briefly some of the widely used landmark CNN networks. These
networks were built upon the basic configurations of dense layers, CONV
layers, and their variations and combinations. Many new special techniques
and tricks have been invented to improve the learning capacity and efficiency.
In fact, there is practically no limit to the types of neural networks that
we can build, but some work well and some may not for some problems.
Having an overall knowledge of these successful landmark networks can be
very useful before one starts to build a new network. The source codes of
many of these networks are made openly available at GitHub, which can be
effectively made use of based on the specified license there.
We will not be able to cover all of the good landmark CNN networks,
but readers can find more in the open sources with the leads given in the
 Convolutional Neural Network (CNN): Classification and Object Detection 547

following discussions, often by simply googling with a keyword. The online

courses by Andrew Ng also provide very clear discussions on this topic. The
intention of this section is to just provide an abstract discussion. To begin
with, we take a look at the LeNet-5.

15.6.1 LeNet-5

This famous CNN was proposed in 1998 by LeCun et al. [8]. The basic (not
exactly the same) structure is given in Fig. 15.9. It uses a black-white image
with a size of (32 × 32 × 1) as the input, the sigmoid activation function for
the hidden layers, and the softmax at the output.
The detailed numbers related to the configuration of the LeNet-5 network
are listed in Fig. 15.10.

Figure 15.9: LeNet-5 for handwritten digit recognition using black-white images. It has
1 input layer, 2 CONV layers, 2 hidden dense layers, and 1 output dense layer.

Figure 15.10: Data size and learning parameters for LeNet-5. It has a total of 61,706
training parameters. The CONV layers have very few training parameters, and most of
the learning parameters are due to the dense layers.
548 Machine Learning with Python: Theory and Applications

Figure 15.10 shows clearly how the number of data flowing forward across
the network and the number of learning parameters in each layer and in total
are calculated. Similar calculations can be used to examine other networks.

15.6.2 AlexNet
Another famous landmark CNN network was proposed in 2012 by Alex
Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. This network won
the 2012 ImageNet LSVRC-2012 competition. Its configuration is shown in
Fig. 15.11. The details can be found in the original paper [9]. It is a widely
used paper and can be easily found online. It has some similarities with
LeNet-5, but with 3 more CONV layers, and one more dense layer. These
3 dense layers have much more neurons. It was built for 1000-classification
on the ImageNet. The AlexNet is much large in scale. It also uses a number
of new techniques, including the Relu activation function to deal with the
vanishing gradient problem and dropout techniques for the first 2 fully
connected dense layers to mitigate over-fitting.

Reader may read the original paper to find more details. Good analysis
of the AlexNet is also given by Andrew Ng’s online lectures and Hao Gao’s
article.

Figure 15.11: AlexNet: It has 1 input layer, 5 CONV layers, 3 hidden dense layers, and
1 output dense layer. The total learning parameters are about 62.3 million, about 6% of
which are from the CONV layers, and 94% from the fully connected dense layers.
 Convolutional Neural Network (CNN): Classification and Object Detection 549

15.6.3 VGG-16

Another popular legend CNN network is VGG-16, which has a number

of improvements upon the AlexNet. It increases the layers from 8 to 16,
with nicely staged convolution and max pooling. Its configuration has the
following appealing uniformity:

• All the CONV layers are unified with the same 3 × 3 filter size, stride 1,
and 1 padding, followed by a max pooling.
• All the max pooling is with the same 2 × 2 size and stride 2.
• The height and width of the net block are roughly halved from one CONV
layer to the next.
• The number of filter channels always doubles from one CONV layer to the
next.

VGG-16 was proposed by K. Simonyan and A. Zisserman in their paper

[10]. VGG-16 was one of best models submitted to ILSVRC-2014. It achieved
92.7% accuracy on ImageNet with over 14 million images of 1,000 classes.
Interested readers may read the original paper and online articles, such
as the one at Nerohive. An implementation of VGG-16 is available at the
TensorFlow-vgg. It also has implementations at Keras and Pytorch.

15.6.4 ResNet

ResNet stands for residual neural network. It was proposed by He, Kaiming;
Zhang, Xiangyu; Ren, Shaoqing; and Sun, Jian in 2015 [11]. The major
difference for the ResNet from the nets discussed above, such as the VGG-16,
is the use of skip connection: the activation of one layer (CONV or dense)
can be connected to the next-next layer, by skipping one layer and then
applying the activation (often the Relu). This well mitigates vanishing of
gradients and enables the net going even deeper, and thus it made up to 152
layers. It has about 23 million learning parameters in total. The capability of
going deep is of importance for many visual recognition tasks. The ResNet
has obtained 28% relative improvement on the COCO object detection
dataset. It is the foundation for the ILSVRC & COCO 2015 competitions,
where the authors won the 1st place in the tasks of ImageNet detection,
ImageNet localization, COCO detection, and COCO segmentation.
550 Machine Learning with Python: Theory and Applications

The skip connection is a useful technique and can be implemented with

ease. For example, for a dense MLP, in the feed-forward process, the activated
values at a ResNet become (see, Eq. (13.5) in Chapter 13 for comparison)
x(0) = x = [x1 , x2 , ..., xp ] feature at input
x(1) = φ(1) (x(0) W(1) + b(1) ) 1st hidden layer
  
z(1)
x(2) = φ(2) (x(1) W(2) + b(2) ) 2nd hidden layer
  
z(2)
x(3) = φ(3) (x(2) W(3) + b(3) + x(0) W(1) ) 3rd hidden layer
  
z(3)
x(4) = φ(4) (x(3) W(4) + b(4) + x(1) W(2) ) 4th hidden layer
  
z(4)
···
x(n) = φ(n) (x(n−2) W(n−1) + b(n−1) + x(n−4) W(n−3) ) nth hidden layer
  
z(n)
ŷ = x(NL )
= φ(NL ) (x(NL −2) W(NL −1) + b(NL −1) + x(NL −4) W(NL −3) ) prediction at output
  
z(NL )
(15.1)

Here the number in the parentheses in the superscript stands for the layer
number. The terms above the curly brackets are the affine transformation
functions z(i) (i = 1, 2, . . . , NL − 1), which can change from layer to layer
and bring in information for the past layers, as new independent “features”
for the next layer after fed to an activation function. The dimension of the
new feature space changes accordingly. For layers that take inputs from the
previous layers, the dimensionality adds up. This may be the reason for
mitigating gradient vanishing issues.
Equation (15.1) assume that the skipping is applied to all the layers
starting from the 3rd hidden layer. One may choose not to skip some layers.
In that case, the term in blue in the corresponding layer should be removed.
It is also assumed that the skips are only over one layer. When multiple-layer
skipping is performed, Eq. (15.1) can be easily adjusted accordingly. Note
that the activation function used in each layer can be in theory different,
although we typically use Relu for all the layers in a ResNet. Also, the
activation in any layer zl can be replaced by that of a CONV layer. This
similarly applies to a net with mixed dense and CONV layers.
A more detailed description of the ResNet can be found in the original
paper and also an article at the website towards data science, where a
detailed architecture of the ResNet is given, in comparison with the VGG-
19 (a later version of VGG-16) and the so-called plain network (without
skipping).
 Convolutional Neural Network (CNN): Classification and Object Detection 551

Readers may also take a look at a clean Keras implementation of residual

network.

15.6.5 Inception

The Inception network is another milestone in CNN development. It was

proposed in 2014 by Christian Szegedy and Wei Liu and seven other co-
authors [12], where the detailed architecture is given. It has been improved
over the years, and the current version is Inception v4 and the Inception-
ResNet.
Inception set a new state of the art for classification and detection in the
ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC14). It
was also called GoogLeNet. The following key ideas are used in the Inception
network compared to the networks discussed above:

• It uses the so-called 1×1 convolution across the channels. This allows a
configuration of a densely connected 1D network across the channel within
a CONV layer, termed as network-in-network. The 1×1 convolution was
proposed by Min Lin, Qiang Chen, and Shuicheng Yan in 2014 [13].
• The use of network-in-network configuration enables the reduction in
channel numbers within a CONV layer, which leads to drastic saving in
computation.
• It constructs Inception Module that is packaged with multiple convo-
lutional operations, including 1×1, 3×3, 5×5, and max pooling all in a
CONV layer, which allows the net to learn which are the best operations
and filter (or kernel) sizes to use in the CONV layer. This is made possible
and efficient with the help of 1×1 convolutions.
• The Inception modules become new building blocks for constructing a
deep Inception network.

Readers may find more details from the original papers and an introduc-
tion article at towards data science.

15.6.6 YOLO: A CONV net for object detection

The development of CNN is getting faster and faster, driven by the

competitions in the related industry, such as self-driven cars. For many real
applications, such as in self-driven cars, we must have a fast and reliable
real-time object detection based on images. YOLO is one of the outstanding
deepnets for such tasks. YOLO stands for “You Only Look Once” [14], and
was proposed by Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali
552 Machine Learning with Python: Theory and Applications

Farhadi. The current version is YOLO-v3. Here, we summarize a number of

the key techniques that are used in the YOLO network:

• It applies a single CONV neural network to a full image given to perform

the detection, which eliminates the need for scanning over the image and
processes the sub-image frame by frame. This allows very fast predictions.
• Different types of predictions needed for the objects, including the types of
the objects (classes), locations of the objects, and (box) sizes of the objects,
are all labeled, so that the network can be trained for fast detections at
one go (You Only Look Once).
• It divides the given image into regions and predicts the probabilities of
the presence of classes and the bounding boxes for each region.

There are many other techniques and tricks developed and implemented
in YOLO. Interested readers may read the original paper and online articles,
such as the one at towards data science. One may also visit the YOLO
project website, where one finds a benchmark study on the performance of
object detection on the COCO dataset. One may also follow up with the
latest development there.
This course has prepared readers with quite a large amount of funda-
mentals on machine learning. We have so many building blocks now, and we
have seen some of the outstanding implementations. It is the time to try to
put these building blocks together, to solve some problems at hand. When
the network becomes deep, there are many different ways to add in different
types of connections, and to put things together for complicated tasks. We
shall also continue the study and invent new techniques and tricks. It may
also be productive to follow up with current work of the front-runners in
this very active area of development. There are many novel techniques and
tricks developed every day, and some of them can be useful not only for
projects that you may working on but also for developing new ideas if one
is interested in research in the related areas. The author and collaborators
have also conducted some study in the related area [6].
Finally, let us look at a CNN network proved by the TensorFlow.

15.7 An Example of Convolutional Neural Network

This tutorial demonstrates training a simple CNN to classify CIFAR

images. Because this tutorial uses the Keras Sequential API (Application
Programming Interface), creating and training a CNN model will take just
 Convolutional Neural Network (CNN): Classification and Object Detection 553

a few lines of code. The material is Licensed under the Apache License,
Version 2.0.

15.7.1 Import TensorFlow

import tensorflow as tf
from tensorflow.keras import datasets, layers, models
import matplotlib.pyplot as plt

15.7.2 Download and preparation of a CIFAR10 dataset

The CIFAR10 dataset contains 60,000 color images in 10 classes, with 6,000
images in each class. The dataset is divided into 50,000 training images and
10,000 testing images. The classes are mutually exclusive and there is no
overlap between them.

(train_images, train_labels), (test_images, test_labels) =\

datasets.cifar10.load_data()
# Normalize pixel values to be between 0 and 1
train_images,test_images=train_images/255.,test_images/255.

15.7.3 Verification of the data

To verify the dataset, let us plot the first 25 images from the training set
and display the class name below each image.
class_names = ['airplane', 'automobile', 'bird', 'cat', 'deer',
'dog', 'frog', 'horse', 'ship', 'truck']

plt.figure(figsize=(10,10))

for i in range(36):
plt.subplot(6,6,i+1)
plt.xticks([])
plt.yticks([])
plt.grid(False)
plt.imshow(train_images[i], cmap=plt.cm.binary)
# The CIFAR labels happen to be arrays,
554 Machine Learning with Python: Theory and Applications

# which is why you need the extra index

plt.xlabel(class_names[train_labels[i][0]])

plt.show()

Figure 15.12: Sample images from the CIFAR10 dataset.

15.7.4 Creation of Conv2D layers

The code below defines a CNN using a common pattern: a stack of Conv2D
and MaxPooling2D layers. It takes in the image in the form of tensors of
shape (image height, image width, color channels). Color channels refer to
(R,G,B). This example configures a CNN to process inputs of shape (32,
32, 3), which is the format of CIFAR images. This is done by passing the
argument input shape to the first layer.
 Convolutional Neural Network (CNN): Classification and Object Detection 555

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu', \
input_shape=(32, 32, 3)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))

Let us display the architecture of our CNN model.

model.summary()

Model: "sequential"

Layer (type) Output Shape Param #

==============================================================
conv2d (Conv2D) (None, 30, 30, 32) 896

max pooling2d (MaxPooling2D) (None, 15, 15, 32) 0

conv2d 1 (Conv2D) (None, 13, 13, 64) 18496

max pooling2d 1 (MaxPooling2D (None, 6, 6, 64) 0

conv2d 2 (Conv2D) (None, 4, 4, 64) 36928

==============================================================
Total params: 56,320
Trainable params: 56,320
Non-trainable params: 0

It is shown that the output of every Conv2D and MaxPooling2D layer is

a 3D tensor with its own shape (height, width, channels). The dimensions
of width and height are often set to shrink and those of the channels are set
to grow as the net gets deeper in the network. Naturally, as the width and
height of the Conv2D layer shrink, more output channels can be added. The
number of output channels (e.g., 32 or 64 for this example) for each Conv2D
layer is controlled by the first argument.
556 Machine Learning with Python: Theory and Applications

15.7.5 Add Dense layers to the Conv2D layers

The output tensor from the last Conv2D layer of shape (4, 4, 64) is fed into
one or more dense layers to perform multi-classification. Because the dense
layers take 1D vectors as input, the 3D Conv2D output shall be flattened
to 1D. Since the CIFAR has 10 classes at the output, the final Dense layer
shall have 10 output neurons with a softmax activation.
model.add(layers.Flatten())
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(10))

The complete architecture of our CNN model is listed below:

model.summary()

Model: "sequential"

Layer (type) Output Shape Param #

==============================================================
conv2d (Conv2D) (None, 30, 30, 32) 896

max pooling2d (MaxPooling2D) (None, 15, 15, 32) 0

conv2d 1 (Conv2D) (None, 13, 13, 64) 18496

max pooling2d 1 (MaxPooling2D (None, 6, 6, 64) 0

conv2d 2 (Conv2D) (None, 4, 4, 64) 36928

flatten (Flatten) (None, 1024) 0

dense (Dense) (None, 64) 65600

dense 1 (Dense) (None, 10) 650

==============================================================
Total params: 122,570
Trainable params: 122,570
Non-trainable params: 0
 Convolutional Neural Network (CNN): Classification and Object Detection 557

As shown above, the 3D (4, 4, 64) outputs at conv2d 2 are flattened into
a 1D vector of shape (1024) before feeding to the first dense layer.

15.7.6 Compile and train the CNN model

model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=True), metrics=['accuracy'])

history = model.fit(train_images, train_labels, epochs=10,

validation_data=(test_images, test_labels))

Epoch 1/10
1563/1563 [================] - 45s 28ms/step - loss: 1.7221 -
accuracy: 0.3648 - val loss: 1.2347 - val accuracy: 0.5573
Epoch 2/10
1563/1563 [================] - 42s 27ms/step - loss: 1.1582 -
accuracy: 0.5869 - val loss: 1.0298 - val accuracy: 0.6363
.....
Epoch 9/10
1563/1563 [================] - 39s 25ms/step - loss: 0.5681 -
accuracy: 0.7993 - val loss: 0.8510 - val accuracy: 0.7178
Epoch 10/10
1563/1563 [================] - 39s 25ms/step - loss: 0.5337 -
accuracy: 0.8114 - val loss: 0.9123 - val accuracy: 0.6919

The training of our CNN model is done for 10 epochs with about 40 s
running time for each.

15.7.7 Evaluation of the trained CNN model

plt.plot(history.history['accuracy'], label='accuracy')
plt.plot(history.history['val_accuracy'],label='val_accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.ylim([0.5, 1])
plt.legend(loc='lower right')
test_loss,test_acc=model.evaluate(test_images,test_labels, verbose=2)

313/313 - 3s - loss: 0.9123 - accuracy: 0.6919

558 Machine Learning with Python: Theory and Applications

Figure 15.13: Accuracy evolution of the CNN model during training.

This simple CNN has achieved a test accuracy of about 70%, as shown
in Fig. 15.13. Not bad for a few lines of code! For another CNN model,
readers may take a look at an example using the Keras subclassing API and
a tf.GradientTape here.

15.8 Applications of YOLO for Object Detection

Object detection involves identifying the objects, including their presence,

identities, dimensions, and locations from a given image. We would need
efficient techniques, such as object recognition, classification, localization,
and dimension determination. For practical applications such as self-driven
cars, we would need to perform such an object detection in real time. YOLO
is one of the best choices.
Sophisticated source codes for versions of YOLO and pre-trained models
are available, many at the official DarkNet GitHub repository written in C,
with tutorials on the usage. There are also third-party implementations of
YOLO in Keras. A widely used project is the “keras-yolo3: Training and
Detecting Objects with YOLO3” by Huynh Ngoc Anh. The project is called
experiencor for short. The codes in the project are available under the MIT
open source license, which is permissive. It provides scripts to load, uses pre-
trained YOLO models, and transfers learning for developing new YOLOv3
models on new datasets.
In this section, we will introduce the basic procedure using a pre-trained
YOLOv3 model for object detection in a given image. The materials and
codes used in this section are based on a very clear and detailed online
 Convolutional Neural Network (CNN): Classification and Object Detection 559

Figure 15.14: YOLOv3 object detection: cars, persons, traffic light around a traffic light
junction.

article by Jason Brownlee, on the experiencor YOLO3 Project. The code is

a little lengthy and thus will not be listed here, but it should be quite easy
to follow after the fundamental understanding is established. The sources
are also listed in the additional reference section.
First, we take a photo with some interested objects to detect if it contains
persons, cars, and traffic lights. We use the trained YOLOv3 model to make
predictions on the objects in the photo. The results are shown in Fig. 15.14.
The YOLOv3 located 3 cars out of 4. The missed one has too much
blockage, and even a human may not be able to see it. All the three persons
are correctly located. It has located 4 traffic lights, 3 out of 4 are correct.
The small box detected as a traffic light is box container for park users. It
is not exactly a traffic light, but quite similar. Thus, this may be arguable.
Overall, the YOLOv3 did an excellent job in locating all these interest objects
from such a messy image. The time taken for the detection is about 20 s on
author’s laptop.
Using the same code, we detected a few more photos, and the results are
summarized below.
For the above case, YOLOv3 found a person on the truck with 95.86%
confidence. It also located the truck with 66.15% confidence (because it is
in fact a trailer) and all the four cars with confidences of 61.95%, 84.14%,
99.63%, and 98.79%. All these objects are located and boxed as shown in
Figs. 15.15 and 15.18.
560 Machine Learning with Python: Theory and Applications

Figure 15.15: YOLOv3 has successfully detected a man standing on a truck (on the right
side of the photo and quite difficult for humans to see), despite that part of the image
being rather dark.

Figure 15.16: Persons in winter clothes detected by the trained YOLOv3.

Additional references and further readings:

• Jason Brownlee, 2019, How to Perform Object Detection With YOLOv3

in Keras.
• experiencor/keras-yolo3, GitHub.
• You Only Look Once: Unified, Real-Time Object Detection, 2015.
• YOLO9000: Better, Faster, Stronger, 2016.
• YOLOv3: An Incremental Improvement, 2018.
• matplotlib.patches.Rectangle API.
• YOLO: Real-Time Object Detection, Homepage.
• Official DarkNet and YOLO Source Code, GitHub.
 Convolutional Neural Network (CNN): Classification and Object Detection 561

Figure 15.17: Four cars and one fire hydrant detected by the trained YOLOv3.

Figure 15.18: Two birds on marsh land detected by the trained YOLOv3.

• Official YOLO: Real Time Object Detection.

• Huynh Ngoc Anh, experiencor, Home Page.
• allanzelener/YAD2K, GitHub.
• qqwweee/keras-yolo3, GitHub.
• xiaochus/YOLOv3 GitHub.
• Blog of Rahmad Sadli.

Before ending this chapter, we may note that from MLP to conv2D, there
is quite a significant change. Such a change has led to huge advancements
in ML. This tells us that our mind needs to be wide open for advancements
in sciences and technology. One may ask now: Can we have conv1D? The
answer is yes. The interested reader may take a look at the online article by
562 Machine Learning with Python: Theory and Applications

Jason Brawnlee at https://machinelearningmastery.com/cnn-models-for-

human-activity-recognition-time-series-classification/. Further, what about
conv3D? Yes. You have guessed it.

References
[1] K. Fukushima, Neocognitron: A self-organizing neural network model for a mechanism
of pattern recognition unaffected by shift in position, Biological Cybernetics, 36, 193–
202, 2004.
[2] D. Ciresan, M. Ueli and J. Schmidhuber, Multi-column deep neural networks for image
classification, 2012 IEEE Conference on Computer Vision and Pattern Recognition,
3642–3649, 2012.
[3] I. Goodfellow, B. Yoshua and C. Aaron, Deep learning, Nature, 521, 436–444, 2015.
[4] M.V. Valueva, N.N. Nagornov, P.A. Lyakhov et al., Application of the residue
number system to reduce hardware costs of the convolutional neural network
implementation, Mathematics and Computers in Simulation, 177, 232–243, 2020.
https://www.sciencedirect.com/science/article/pii/S0378475420301580.
[5] Duan Shuyong, Ma Honglei, G.R. Liu et al., Development of an automatic lawnmower
with real-time computer vision for obstacle avoidance, International Journal of
Computational Methods, Accepted, 2021.
[6] Duan Shuyong, Lu Ningning, Lyu Zhongwei et al., An anchor box setting technique
based on differences between categories for object detection, International Journal of
Intelligent Robotics and Applications, 1–14, 2021.
[7] V. Dumoulin and F. Visin, A guide to convolution arithmetic for deep learning, arXiv,
1603.07285, 2018.
[8] Y. Lecun, L. Bottou, Y. Bengio et al., Gradient-based learning applied to document
recognition, Proceedings of the IEEE, 86(11), 2278–2324, 1998.
[9] A. Krizhevsky, I. Sutskever and G. E., Imagenet classification with deep convolutional
neural networks, Advances in Neural Information Processing Systems, 2012.
[10] S. Karen and Z. Andrew, Very deep convolutional networks for large-scale image
recognition, arXiv preprint arXiv:1409.1556, 2014. https://www.bibsonomy.org/bib
tex/2bc3ee27a1dd159f48b10ac3555879865/buch jon.
[11] He Kaiming, Zhang Xiangyu, Ren Shaoqing et al., Deep residual learning for image
recognition, CoRR, abs/1512.03385, 2015. http://arxiv.org/abs/1512.03385.
[12] Christian Szegedy, Wei Liu, Yangqing Jia et al., Going Deeper with Convolutions,
2014.
[13] Min Lin, Qiang Chen and Shuicheng Yan, Network In Network, 2014. https://arxiv.
org/abs/1312.4400.
[14] R. Joseph, D. Santosh, G. Ross et al., You only look once: Unified, real-time object
detection, CoRR, abs/1506.02640, (1), 2015. http://arxiv.org/abs/1506.02640.
 Chapter 16

Recurrent Neural Network (RNN)

and Sequence Feature Models

Recurrent neural networks (RNNs) are yet another quite special class of
artificial neural networks (NNs) for datasets with features of sequence or
temporal significance. An RNN typically has an internal state recorded
in a memory cell, so that it can process sequential inputs, such as
video and speech records. Affine transformations are used in an extended
affine space that includes memorized states as additional features. RNNs
have important applications in various areas including natural language
processing, speech recognition, translation, music composition, robot control,
and unconstrained handwriting recognition [1], to name just a few. A recent
review on RNNs can be found in [2].
Different types of RNNs have been developed [2]. This chapter focuses on
the so-called long short-term memory networks (LSTMs) presented originally
by Sepp Hochreiter and Jürgen Schmidhuber [3]. This is because (1) LSTM
is most widely used and studied; (2) when LSTM is well understood, it is
relatively easy to understand other RNNs; and (3) most publicly available
modules have LSTM classes built in for ready use.
A typical LSTM unit is composed of a cell, an update gate, a forget
gate, and an output gate. The cell remembers values over time sequence
corresponding to the sequential data inputs. These three gates regulate the
information exchange between the memory cell and newly inputted data. The
level of the regulation is controlled by nested neural network (NN) layers with
learning parameters. These parameters are trained via a supervised process
similar to the NN or MLP models discussed in previous chapters.
LSTMs were developed to deal with the vanishing gradient problem
encountered in training traditional NNs, because they allow information to
travel deep in the layers through interconnections built in the LSTM.

563
564 Machine Learning with Python: Theory and Applications

In writing this chapter, the following materials are referenced in addition

to those listed in the reference section: Keras documentation on LSTM,
Wikipedia page on LSTM, etc.

16.1 A Typical Structure of LSTMs

Figure 16.1 shows a schematic drawn of an unfolded LSTM in processing a

sequential dataset. Consider a dataset with p dimensional vectors varying in
time. Let
T : the number of the time sequences
p: the dimension of the feature vectors (16.1)
l: the dimension of LSTM output spaces
We denote,
Input data:
x = [x(1) , x(2) , . . . , x(t) , . . . , x(T ) ] ∈ XT ×p , x(t) ∈ Xp
Activation states:
(16.2)
a = [a(0) , a(1) , a(2) , . . . , a(t) , . . . , a(T ) ] ∈ X(T +1)×l , a(t) ∈ Xl
Memory cell states:
c = [c(0) , c(1) , c(2) , . . . , c(t) , . . . , c(T ) ] ∈ X(T +1)×l , c(t) ∈ Xl

In an LSTM, the initial vectors a(0) and c(0) are usually set to zero.
These activations can be outputted at each time sequence or only at the
final time T , depending on the type of labels used in the training for the
problem at hand.

Figure 16.1: Unfolded identical LSTM units for processing a sequential dataset in an
LSTM RNN. For a given sequential dataset x, the LSTM uses a memory cell to memorize
the state c at any time. The corresponding activation becomes a, which is updated and
can be outputted at each time sequence or only at the final time T .
 Recurrent Neural Network (RNN) and Sequence Feature Models 565

Figure 16.2: An LSTM unit. It has, recurrently, a memory cell c(t−1) that memorizes
the state and an activation state a(t−1) , both in Xl . The unit corresponds to data input
x(t) ∈ Xp at time t. The orange blocks are neural network layers, or an MLP, that perform
1+p+l
affine transformations in an extended affine space X producing a vector in Xl . Yellow
circles and ellipse denote element-wise operations.

These gates are equipped with independent learning parameters, and used
to control the information exchange between x, a and c.
Note that Fig. 16.1 is used to show the sequential process inside an LSTM.
In actual computation, only one LSTM unit is needed in a loop, by feeding
back the a(t) and c(t) to the next time sequence. A typical unit is shown in
Fig. 16.2, where all the detailed operations are shown.

16.2 Formulation of LSTMs

16.2.1 General formulation

Consider time t. At the start, the activation state is a(t−1) , and the memory
cell is c(t−1) . Let l be the dimension of the output of our LSTM. Data x(t)
is a vector of p components, and is now inputted. Date x(t) and previous
activation a(t−1) are concatenated to form an extended affine space together,
1, which shall have a dimension of 1 + p + l. We perform the following
operations using all the current learning parameters:
1+l+p
xa = [1, x(t) , a(t−1) ] ∈ X extended affine space
c̃ = tanh(xa Ŵc ) ∈ (−1, 1)l candidate memory
gu = σ(xa Ŵu ) ∈ (0, 1)l update gate
gf = σ(xa Ŵf ) ∈ (0, 1)l forget gate (16.3)
go = σ(xa Ŵo ) ∈ (0, 1)l output gate
c(t) = gu ∗ c̃ + gf ∗ c(t−1) ∈ (−1, 1)l new memory
a(t) = go ∗ tanh(c(t) ) ∈ (−1, 1)l new activation
566 Machine Learning with Python: Theory and Applications

Figure 16.3: LSTM layer. An LSTM can be simply regarded as a normal NN layer that
can be placed horizontally (left) or vertically (right). The input x has typically a shape of
(m, T, p) and the output a has a shape of (m, T, l), where m stands for mini-batch size. If
only the final activation output is needed, the shape of a(T ) has a shape of (m, l).

We highlight the input in blue, and output in red. These two are
the connection points between LSTM units. The learning parameters are
Ŵc , Ŵu , Ŵf , and Ŵo . If single NN layers are used in an LSTM, they
all are in W(1+l+p)×l , and thus the total number of training parameters is
P = 4 × (1 + l + p) × l. Equation (16.3) is practically a computational graph
of LSTM, which gives both forward and backward paths. Since all these
operations are chained and all the derivatives of these activation functions
are available, computing the gradients with respect to any of these training
samples can be done in the standard way using autograd, as discussed in
Chapter 8.

16.2.2 LSTM layer and standard neural layer

It is important to note that Eq. (16.3) and the figures may looks complicated,
but it can be simply regarded as a single NN layer. The input to LSTM is
only x(t) , and the output from it is just a(t) . All the others are the internal
variables. An LSTM layer can be shown in the Fig. 16.3. Therefore, an LSTM
layer is very easy to use, and can be used as a hidden layer in a neural network
model that has a proper loss function defined at the terminal output layer
of the entire model.

16.2.3 Reduced LSTM

Consider now a special case when the number of the time sequences T = 1.
In this case, a(t−1) = c(t−1) = 0. Our formulation is reduced to
 Recurrent Neural Network (RNN) and Sequence Feature Models 567

1+p
x = [1, x(t) ] ∈ X standard affine space
c̃ = tanh(x Ŵc ) ∈ (−1, 1)l candidate memory
gu = σ(x Ŵu ) ∈ (0, 1)l update gate
gf = σ(x Ŵf ) ∈ (0, 1)l forget gate (16.4)
go = σ(x Ŵo ) ∈ (0, 1)l output gate
c(t) = gu ∗ c̃ ∈ (−1, 1)l new memory
a(t) = go ∗ tanh(c(t) ) ∈ (−1, 1)l new activation

which can be simplified as

 
a(t) = σ(x Ŵo ) ∗ tanh σ(x Ŵu ) ∗ tanh(x Ŵc ) ∈ (−1, 1)l new activation
(16.5)

Note that ∗ stands for element-wise multiplication. The learning parame-

ters become Ŵc , Ŵu , and Ŵo . They are all now in W(1+p)×l . It is essentially
1+p
a mapping from X to (−1, 1)l ∈ Xl on the standard affine space. The
reduced LSTM is equipped with three times more learning parameters
1+p
compared to a standard NN layer that maps from X to (−1, 1)l ∈ Xl .
The total number of training parameters are P = 3 × (1 + p) × l.
The 1st term on the right-hand side of Eq. (16.5) is essentially a standard
NN layer with sigmoid activation, which is regulated by a tanh function
that, in turn, is a function of another NN regulated by tanh activation of
yet another NN. This configuration makes all these three groups of learning
parameters independent. Therefore, the reduced LSTM is an NN regulated
by the other two nested regulated NNs, resulting in three times more learning
parameters.

16.3 Peephole LSTM

There are many other versions of LSTM, and peephole LSTM [4] is one of
them. Peephole LSTM is said to allow the gates to access the constant error
carousel, using the cell state c(t−1) instead of activation a(t−1) . Note that
c(t−1) and a(t−1) have the same shape, as discussed earlier. The formulation
is thus quite similar and is given as follows:
1+p
x = [1, x(t) ] ∈ X standard affine space
1+l+p
xc = [x, c(t−1) ] ∈ X extended affine space
568 Machine Learning with Python: Theory and Applications

c̃ = tanh(x Ŵc ) ∈ (−1, 1)l candidate memory

gu = σ(xc Ŵu ) ∈ (0, 1)l update gate (16.6)
gf = σ(xc Ŵf ) ∈ (0, 1)l forget gate
go = σ(xc Ŵo ) ∈ (0, 1)l output gate
c(t) = gu ∗ c̃ + gf ∗ c(t−1) ∈ (−1, 1)l new memory
a(t) = go ∗ tanh(c(t) ) ∈ (−1, 1)l new activation

All the other operations of the peephole LSTM are similar to the LSTM.

16.4 Gated Recurrent Units (GRUs)

Gated recurrent units (GRUs) were introduced by Kyunghyun Cho et al. [5].
It may be viewed as a version of LSTM without a dedicated memory cell
and hence with less learning parameters than LSTM. The formulation is as
follows.
Consider time t. The activation state at its previous time is a(t−1) with
a(0) = 0. Let l be the dimension of the output of the GRU. Data x(t) is
a vector of p components, and is now inputted. Data x(t) and previous
activation a(t−1) are concatenated to form an extended affine space together,
1, which shall have a dimension of 1 + p + l. We perform the following
operations:

1+l+p
xa = [1, x(t) , a(t−1) ] ∈ X extended affine space
gu = σ(xa Ŵu ) ∈ (0, 1)l update gate
gr = σ(xa Ŵr ) ∈ (0, 1)l reset gate
1+l+p
(16.7)
xra = [1, x(t) , gr a(t−1) ] ∈ X reset extended affine space
ã = tanh(xra Ŵa ) ∈ (−1, 1)l candidate activation
a(t) = (1 − gu )a(t−1) + gu ã ∈ (−1, 1)l new output activation

The learning parameters in a GRU are Ŵu , Ŵr , and Ŵa , 3/4 of those in
an LSTM. Alternative versions of GRUs are also available. Interested readers
my visit the Wiki page on GRUs and then the links therein.
 Recurrent Neural Network (RNN) and Sequence Feature Models 569

16.5 Examples

16.5.1 A simple reduced LSTM with a standard NN layer

for regression

Let us first look at an example for the reduced LSTM that can be
easily handcrafted for comprehension. We will use the simple and familiar
regression problem examined step by step in Chapter 10, using a synthesis
dataset and the MxNet. The details on the problem setting will not be
repeated here. We will follow our xw formulation in the following code:
from __future__ import print_function
from mxnet import nd, autograd, gluon
import numpy as np
import mxnet as mx
import matplotlib.pyplot as plt
%matplotlib inline
np.set_printoptions(precision=4)

mx.random.seed(1) # fix seed for repeatable random numbers.

# Specify GPU or CPU for the computation, mxnet's syntax.

data_ctx, model_ctx = mx.cpu(), mx.cpu()
num_inputs = 2 # p, 2D space with two features: x1, x2
num_outputs = 1 # k, one prediction y
num_samples = 1000 # Sample number, m, use 10 for debugging

p1 = num_inputs + 1 # dimension of the affine space

wr=nd.array([[4.2],[2.0],[-3.4]]) # true w_=[b,w0,w1].T,1 neuron

print(f" True w:{wr},{wr.shape}")

def true_fn(X, w): # linear func, 3 learning parameters:w1,w2,b

return nd.dot(X,w) # X is the X_; 2*x1-3.4*x2 + 4.2

X=nd.random_normal(shape=(num_samples,p1),ctx=data_ctx)
# randomly generate samples for x1 and x2
noise=.1*nd.random_normal(shape=(num_samples,num_outputs),
ctx=data_ctx)
# 10% noise, zero mean, 1 variance
X[:,0] = 1.0 # / num_samples
print(f" X in affine space:{X[0:4]},{X.shape}")

y = true_fn(X, wr) + noise # Generate labels with noise.

print(f"y={y[0:4]}{y.shape}") # A peek at the dataset
570 Machine Learning with Python: Theory and Applications

True w:
[[ 4.2]
[ 2. ]
[-3.4]]
<NDArray 3x1 @cpu(0)>,(3, 1)
X in affine space:
[[ 1. -0.4902 -0.9502]
[ 1. -0.7298 -2.0401]
[ 1. 1.0408 -0.4526]
[ 1. -0.8367 -0.7883]]
<NDArray 4x3 @cpu(0)>,(1000, 3)
y=
[[6.5203]
[9.5923]
[7.7648]
[5.2164]]
<NDArray 4x1 @cpu(0)>(1000, 1)

# Plot the data points.

plt.scatter(X[:, 1].asnumpy(),y.asnumpy(),color='r') # 1st feature
plt.scatter(X[:, 2].asnumpy(),y.asnumpy(),color='b') # 2nd feature
plt.show()

Figure 16.4: Computer-generated data-points for study LSTM.

 Recurrent Neural Network (RNN) and Sequence Feature Models 571

# Create data iterators

batch_size=32
train_data=gluon.data.DataLoader(gluon.data.ArrayDataset(X,y),
batch_size=batch_size, shuffle=True)

# Initialize the training parameters, only W_ is needed!

def init_wb(p1,l,k): # p1: affine space dimension
'''p:model num_inputs; l:lstm outputs; k:model num_outputs'''
wo = nd.random_normal(shape=(p1,l),ctx=model_ctx)
wu = nd.random_normal(shape=(p1,l),ctx=model_ctx)
wc = nd.random_normal(shape=(p1,l),ctx=model_ctx)
w = nd.random_normal(shape=(l,k),ctx=model_ctx)
return w, wo, wu, wc

l = 50
w, wo, wu, wc = init_wb(p1,l,num_outputs) #p1 = p+1
params = [w, wo, wu, wc]

# Request for gradients

for param in params:
param.attach_grad()
#print(param.grad) # check initial gradients

# Define a reduced LSTM layer

def rd_lstm(X): #only X_ is needed!

'''Reduced LSTM, no T-loop'''
tc = nd.tanh(nd.dot(X, wc))
su = nd.sigmoid(nd.dot(X, wu))
rd_lstm = nd.sigmoid(nd.dot(X, wo))*nd.tanh(su*tc)
return rd_lstm

# Define a standard NN layer

def net(X): # X_ is needed
return nd.dot(X, w) # only w_ is needed!

# Define a L2 loss function for regression

def square_loss(yhat, y):
return nd.mean((yhat - y) ** 2)

# Optimizer using the stochastic gradient descent (SGD)

def SGD(params, lr): # lr: learning rate
572 Machine Learning with Python: Theory and Applications

for param in params:

param[:] = param - lr * param.grad

# Training and plot the training process and results.

epochs = 50 # times of getting batch-data from dataset

num_batches = num_samples/batch_size

# Re-initialize parameters. They may be trained earlier

w, wo, wu, wc = init_wb(p1,l,num_outputs) # p1 = p+1
params = [w, wo, wu, wc]
for param in params:
param.attach_grad()

# Function to plot the losses over time

def plot(losses, X, s_s = 100): #s_s: sample_size
xs = list(range(len(losses)))
f, (fg1, fg2) = plt.subplots(1, 2)
fg1.set_title('Loss during training')
fg1.plot(xs, losses, '-r')
fg2.set_title('Estimated vs real function')
fg2.plot(X[:s_s,2].asnumpy(),net(rd_lstm(X[:s_s,:])).asnumpy(),
'or',label='Estimated')
fg2.plot(X[:s_s, 2].asnumpy(), true_fn(X[:s_s,:],wr).asnumpy(),
'*g',label='Real')
fg2.legend()
plt.show()

learning_rate = .1
losses = []
plot(losses, X)

for e in range(epochs):
cumulative_loss = 0
for i, (data, label) in enumerate(train_data):
data = data.as_in_context(model_ctx)
label = label.as_in_context(model_ctx).reshape((-1,1))

with autograd.record(): #Computational graph

output = net(rd_lstm(data))
loss = square_loss(output, label)

loss.backward()
SGD(params, learning_rate)
cumulative_loss += loss.asscalar()
 Recurrent Neural Network (RNN) and Sequence Feature Models 573

if e%10 ==0:
print("Epoch %s, batch %s. Mean loss: %s" %
(e, i, cumulative_loss/num_batches))

losses.append(cumulative_loss/num_batches)

plot(losses, X)

Figure 16.5: Prediction using LSTM before training.

Epoch 0, batch 31. Mean loss: 3.1009296717643737

Epoch 10, batch 31. Mean loss: 0.03779914233088493
Epoch 20, batch 31. Mean loss: 0.020572829380631447
Epoch 30, batch 31. Mean loss: 0.01705670604109764
Epoch 40, batch 31. Mean loss: 0.01468234208971262

Figure 16.6: Prediction using the trained LSTM.

574 Machine Learning with Python: Theory and Applications

This example demonstrates that a reduced LSTM together with a

standard NN layer can perform a regression task. The use of LSTM for this
simple example may not be helpful. In fact, it makes the model unnecessarily
more complex, and it is overkill. However, this example does show clearly
that an LSTM can have multiple learn parameters built into it in a nested
fashion, which may have more capability for complicated problems.
We can easily create a loop in our simple code and turn our reduced
LSTM to a full LSTM. However, is would be much easier to just make use
of the well-developed LSTM classes available in a publicly available module.

16.5.2 LSTM class in tensorflow.keras

Let us this time use the LSTM class in tensorflow.keras. We first examine
the dimensionality of the input and outputs to LSTM using a randomly
generated synthetic dataset.

import numpy as np
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense

np.set_printoptions(precision=3)
tf.keras.backend.clear_session()

m, T, p = 32, 100, 28 # T: timesteps; p: features; m: mimi-batch

l = 4 # l: dimension of the LSTM output

tf.random.set_seed(8)
inputs = tf.random.normal([m, T, p])
#inputs: A 3D tensor with shape [m, T, p].

lstm1 = tf.keras.layers.LSTM(l) # create a LSTM instance

output0 = lstm1(inputs)

print(inputs[-1, T-1,0:p:10],'\n',inputs.shape)
print(output0[-1,:],'\n',output0.shape) #(32, 4)

tf.Tensor([ 0.521 -1.34 1.215], shape=(3,), dtype=float32)

(32, 100, 28)
tf.Tensor([ 0.113 0.059 -0.321 0.248], shape=(4,), dtype=float32)
(32, 4)
 Recurrent Neural Network (RNN) and Sequence Feature Models 575

lstm2=tf.keras.layers.LSTM(l,return_sequences=True,return_state=True)

whole_seq_output, final_memory_state, final_carry_state=lstm2(inputs)

print(f"whole_seq_output.shape: {whole_seq_output.shape}") #(32,10,4)

print(f"final_memory_state.shape:{final_memory_state.shape}") #(32,4)
print(f"final_carry_state.shape:{final_carry_state.shape}") #(32,4)
print(f"final_memory_states:{final_memory_state[-1,:]}")
print(f"final_carry_stfinal_memory_stateate",
f"{final_carry_state[-1,:]}")

whole seq output.shape: (32, 100, 4)

final memory state.shape:(32, 4)
final carry state.shape:(32, 4)
final memory states:[ 0.079 0.01 -0.295 0.048]
final carry stfinal memory stateate [ 0.17 0.018 -0.863 0.099]

16.5.3 Using LSTM for handwritten digit recognition

We now use LSTM for handwritten digit recognition. We first train an LSTM
model using the training dataset of MNIST and perform the testing using
the test dataset to examine the accuracy of the recognition of the trained
LSTM model.
We used the MNIST dataset four times. It contains 28x28 images of digits
with labels. It is a static dataset. However, it can be treated as a dynamic
one. We first treat the image pixels in a row of the image as a vector of
features, and thus p = 28. Each row is then treated and a time sequence,
which gives T = 28. We simply use the default mini-batch size for our LSTM
model, which is m = 32.

mnist=tf.keras.datasets.mnist # 28×28 images with labels

(x_train,y_train),(x_test,y_test)=mnist.load_data()

x_train, x_test = x_train/255., x_test/255. #normalization

print(x_train.shape, x_train.shape[1:])
T = x_train.shape[1:][0]
p = x_train.shape[1:][1]
k = len(np.unique(y_train)) # k-classes of MNIST dataset
l = int(x_train[0].shape[0]*x_train[0].shape[1]/10)
# use 1/10th the total image features as the LSTM output, l
p1 = 1 + p + l # dimension of the extended affine space
print(f"Number of classes k ={k}")
576 Machine Learning with Python: Theory and Applications

print(f"Number of features p ={p}")

print(f"Dimension of the affine space ={p1}")
print(f"Number of outputs from the LSTM layer ={l}")
print(f"Training Parameters in LSTM ={4*p1*l}")

model = Sequential()
# LSTM layer to process sequential data [m, T, k]
# It produces an output [m, l]. Use default activation:tanh
# Outputs only the final activation a^T.
model.add(LSTM(l, input_shape=(x_train.shape[1:])))
# Add a dense layer for k-classification, using LSTM outputs.
model.add(Dense(k, activation='softmax')) # a^T is feed here
optimizer=tf.keras.optimizers.Adam(learning_rate=.01,decay=1e-6)

# Compile model: 1-LSTM, 1-Dense, use Adam

model.compile(loss='sparse_categorical_crossentropy',
optimizer=optimizer, metrics=['accuracy'])

model.summary()

(60000, 28, 28) (28, 28)

Number of classes k =10
Number of features p =28
Dimension of the affine space =107
Number of outputs from the LSTM layer =78
Training Parameters in LSTM =33384
Model: "sequential"

Layer (type) Output Shape Param #

===============================================================
lstm 2 (LSTM) (None, 78) 33384

dense (Dense) (None, 10) 790

===============================================================
Total params: 34,174
Trainable params: 34,174
Non-trainable params: 0

Our model has 33,384 training parameters for the LSTM layer, same as
our calculation. The whole model consists of one LSTM layer and one dense
NN layer. We shall now train the model.
 Recurrent Neural Network (RNN) and Sequence Feature Models 577

model.fit(x_train,y_train,epochs=3,validation_data=(x_test,y_test));
# Using the default m_size 32

Epoch 1/3
1875/1875 [==============================] - 32s 16ms/step - loss:
0.2583 - accuracy: 0.9175 - val loss: 0.1020 - val accuracy: 0.9703
Epoch 2/3
1875/1875 [==============================] - 29s 16ms/step - loss:
0.1091 - accuracy: 0.9683 - val loss: 0.0897 - val accuracy: 0.9716
Epoch 3/3
1875/1875 [==============================] - 30s 16ms/step - loss:
0.0882 - accuracy: 0.9731 - val loss: 0.0792 - val accuracy: 0.9772

This model achieved an accuracy of ∼98%, which is a quite good

performance. One can easily add in additional LSTM layers and/or dense
layers to form a deep net. The following code builds an LSTM-MLP, just for
demonstration purposes:

Lmlp = Sequential()
l = int(l/2) #reduce LSTM output, because using more layers
Lmlp.add(LSTM(l, input_shape=(x_train.shape[1:]),
return_sequences=True)) # True: allows feed to next layer
Lmlp.add(LSTM(l, input_shape=(x_train.shape[1:])))
Lmlp.add(Dense(2*k, activation='relu')) # a^T is feed here
Lmlp.add(Dense(k, activation='softmax')) # for classification
optimizer=tf.keras.optimizers.Adam(learning_rate=.001,decay=1e-6)

# Compile our Lmlp: 2-LSTM layers, 2-Dense layers, use Adam

Lmlp.compile(loss='sparse_categorical_crossentropy',
optimizer=optimizer, metrics=['accuracy'])

Lmlp.summary()

Model: "sequential 1"

Layer (type) Output Shape Param #

===============================================================
lstm 3 (LSTM) (None, 28, 39) 10608

lstm 4 (LSTM) (None, 39) 12324

578 Machine Learning with Python: Theory and Applications

dense 1 (Dense) (None, 20) 800

dense 2 (Dense) (None, 10) 210

===============================================================
Total params: 23,942
Trainable params: 23,942
Non-trainable params: 0

Lmlp.fit(x_train,y_train,epochs=3,validation_data=(x_test,y_test));

Epoch 1/3
1875/1875 [===================] - 39s 19ms/step - loss:
0.4861 - accuracy: 0.8447 - val loss: 0.1857 - val accuracy: 0.9420
Epoch 2/3
1875/1875 [===================] - 36s 19ms/step - loss:
0.1394 - accuracy: 0.9597 - val loss: 0.1105 - val accuracy: 0.9674
Epoch 3/3
1875/1875 [===================] - 36s 19ms/step - loss:
0.0965 - accuracy: 0.9711 - val loss: 0.0818 - val accuracy: 0.9755

The “deep” LSTM-MLP achieved a similar level of accuracy with about

2/3 training parameters.

16.5.4 Using LSTM for predicting dynamics of moving vectors

This example builds an LSTM for predicting the movements in time of

vectors in 2D space.

import numpy as np
import keras
import tensorflow as tf
from keras.models import Sequential
from keras.layers.core import Dense, Activation
from keras.layers.recurrent import LSTM
import matplotlib.pylab as plt
import sys
%matplotlib inline
 Recurrent Neural Network (RNN) and Sequence Feature Models 579

print(f"keras:{keras.__version__}, tensorflow: {tf.__version__}")

keras:2.6.0, tensorflow: 2.6.2

np.set_printoptions(precision=5)
tf.keras.backend.clear_session()

# Generate a raw dataset of moving 2D vectors

x1=np.array(np.sin( 2*np.pi*np.arange(0,10,.001))) #component 1
x2=np.array(np.cos(2.3*np.pi*np.arange(0,10,.001))) #component 1

noise_level = 0.1 # 10% Gaussian

noise = np.random.normal(0, 1, len(x1)) * noise_level
x1 = x1 + noise
noise = np.random.normal(0, 1, len(x1)) * noise_level
x2 = x2 + noise
print(x1[:5],x1.shape, x2.shape)

X = np.vstack((x1,x2)).T # Form the dataset

print(X[:5],X.shape,len(X))

[ 0.00639 -0.04602 0.06577 0.07202 0.04981] (10000,) (10000,)

[[ 0.00639 0.8971 ]
[-0.04602 0.9104 ]
[ 0.06577 1.22828]
[ 0.07202 0.982 ]
[ 0.04981 1.01548]] (10000, 2) 10000

# Set configurational parameters for the model

T, p, L = 100, 2, 20 # T: timesteps; p: features
#L: dimension of the LSTM output
out_d = 2 # output dimension for the whole model, or k
p1 = 1 + p + L # dimension of the extended affine space
print(f"Final output dimension ={out_d}")
print(f"Number of time sequences T={T}")
print(f"Number of features p ={p}")
print(f"Dimension of the extended affine space ={p1}")
print(f"Number of outputs from the LSTM layer ={L}")
print(f"Training Parameters in LSTM ={4*p1*L}")
580 Machine Learning with Python: Theory and Applications

Final output dimension =2

Number of time sequences T=100
Number of features p =2
Dimension of the extended affine space =23
Number of outputs from the LSTM layer =20
Training Parameters in LSTM =1840

# Take a look a the raw data

plt.rcParams["figure.figsize"] = (4, 2.5)
plt.figure(dpi=300)
plt.plot(X[:1000:10][:,0],"-")
plt.plot(X[:1000:10][:,1],"--")
plt.xlabel('Time t')
plt.ylabel('Coordinates of 2D vectors')
plt.legend(["x1", "x2"]);

Figure 16.7: Sample coordinates of the trajectory of a moving vector with noise.

def createDataset(X, T = 100):

'''Create dataset with proper shape for keras LSTM class
input: X[*,p]; output: Xt[*,T,p]'''
Xt, yt = [], []
for i in range(len(X)-T):
Xt.append(X[i:i+T])
yt.append(X[i+T])
Xt = np.array(Xt)
yt = np.array(yt)
return Xt, yt
 Recurrent Neural Network (RNN) and Sequence Feature Models 581

def sequentialSplit(X, y, train_rate=0.8):

'''Sequentially split data, without random shuffling.
1st part for training, the rest for test'''
n_train = round(len(X) * train_rate)
X_train, y_train = X[0:n_train], y[0:n_train]
X_test, y_test = X[n_train:], y[n_train:]
return X_train, y_train, X_test, y_test

def randomSplit(Xt, yt): # not used in this example

''' This code may be used for reshuffled data batches '''
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(
Xt, yt, train_size=.8, random_state=8)
return X_train, y_train, X_test, y_test

Xt, yt = createDataset(X, T = T )
print(Xt[1][0],Xt.shape,Xt.shape[1])
print(yt[1],yt.shape)

[-0.04602 0.9104 ] (9900, 100, 2) 100

[0.61192 0.61808] (9900, 2)

We split the dataset sequentially, so that we can train the model using
the first part of the data, and then predict the future movements of the 2D
vectors. One may shuffle the dataset and then split the dataset. In this case,
we train the model for predicting the hidden features of the dataset.

X_train, y_train, X_test, y_test = sequentialSplit(Xt, yt)

#X_train, y_train, X_test, y_test = randomSplit(Xt, yt)

print(X_train[1][0],X_train.shape)
print(X_test[1][0],X_test.shape)

[-0.04602 0.9104 ] (7920, 100, 2)

[-0.62915 0.92939] (1980, 100, 2)
582 Machine Learning with Python: Theory and Applications

# Build the LSTM-NN model

model = Sequential()
model.add(LSTM(L, input_shape=(T, p)))
model.add(Dense(out_d, input_dim=L))
model.add(Activation("linear"))
model.compile(loss="mean_squared_error", optimizer="rmsprop")

model.summary()

Model: "sequential"

Layer (type) Output Shape Param #

===============================================================
lstm (LSTM) (None, 20) 1840

dense (Dense) (None, 2) 42

activation (Activation) (None, 2) 0

===============================================================
Total params: 1,882
Trainable params: 1,882
Non-trainable params: 0

model.fit(X_train,y_train,batch_size=128,epochs=5,validation_split=.05);

Epoch 1/5
59/59 [==============================] - 8s 78ms/step - loss:
0.1552 - val loss: 0.0290
Epoch 2/5
59/59 [==============================] - 4s 73ms/step - loss:
0.0151 - val loss: 0.0154
Epoch 3/5
59/59 [==============================] - 4s 73ms/step - loss:
0.0124 - val loss: 0.0126
Epoch 4/5
59/59 [==============================] - 4s 73ms/step - loss:
0.0121 - val loss: 0.0128
 Recurrent Neural Network (RNN) and Sequence Feature Models 583

Epoch 5/5
59/59 [==============================] - 4s 73ms/step - loss:
0.0119 - val loss: 0.0125

# Accuracy assessment
from numpy.linalg import norm
predicted = model.predict(X_test)
rmse = np.sqrt(((predicted - y_test) ** 2).mean(axis=0))
print(f"Root Mean Square Error = {rmse}")
print(f"Relative Root Mean Square Error= {rmse/norm(y_test)}")

Root Mean Square Error = [0.10676 0.11551]

Relative Root Mean Square Error= [0.00238 0.00257]

# Plot the prediction results

nt, ns = len(y_test), 50
plt.rcParams["figure.figsize"] = (4, 2.5)
plt.figure(dpi=300)
plt.plot(predicted[:nt:ns][:,0],"--")
plt.plot(y_test[:nt:ns][:,0],":")
plt.legend(["Predict", "Test"])
plt.xlabel('Time t')
plt.ylabel('Coordinates of vectors, x1');

Figure 16.8: LSTM predicted coordinate x1 of the trajectory of a moving vector, in

comparison with the testing data.

plt.rcParams["figure.figsize"] = (4, 2.5)

plt.figure(dpi=300)
plt.plot(predicted[:nt:ns][:,1],"--")
584 Machine Learning with Python: Theory and Applications

plt.plot(y_test[:nt:ns][:,1],":")
plt.legend(["Predict", "Test"])
plt.xlabel('Time t')
plt.ylabel('Coordinates of vectors, x2');

Figure 16.9: LSTM predicted coordinate x2 of the trajectory of a moving vector, in

comparison with the testing data.

It is seen in Fig. 16.9 that our LSTM model gives a reasonable prediction
based on this dataset. It captures the major waving features of the moving
vectors.

16.6 Examples of LSTM for Speech Recognition

LSTM is one of the most powerful tools for speech recognition. There are
many open-source examples. Interested readers may take a look at the mxnet
on Speech LSTM. The source code is also available at their GitHub site.

References

[1] G. Alex, L. Marcus, F. Santiago et al., “A Novel Connectionist System for Uncon-
strained Handwriting Recognition”, IEEE Transactions on Pattern Analysis and
Machine Intelligence, 31, pp. 855–868, 2009.
[2] Yu Yong, Si Xiaosheng, Hu Changhua et al., “A Review of Recurrent Neural Networks:
LSTM Cells and Network Architectures”, Neural Computation, 31(7), 1235–1270,
2019.
[3] H. Sepp and S. Jürgen, “Long Short-Term Memory”, Neural Computation, 9(8), 1735–
1780, 1997. https://doi.org/10.1162/neco.1997.9.8.1735.
[4] F.A. Gers and E. Schmidhuber, “LSTM recurrent networks learn simple context-
free and context-sensitive languages”, IEEE Transactions on Neural Networks, 12(6),
1333–1340, 2001.
[5] Cho KyungHyun, On the Properties of Neural Machine Translation: Encoder-Decoder
Approaches, CoRR, abs/1409.1259, 2014. http://arxiv.org/abs/1409.1259.
 Chapter 17

Unsupervised Learning Techniques

17.1 Background

The world now has a massive amount of data, and it is growing. How to
make use of these data becomes very important. There are mainly two ways
to make the data useful. One is to have the data examined by experts in
the related areas, and have it labeled, and then the data can be used to
train machine learning models. Labeling data, however, is a very expensive
and time-consuming task. The other way is to develop machine learning
algorithms that read the (unlabeled) raw data, and try to extract some
latent features or characteristics from the data. This is the unsupervised
learning that we are discussing in this chapter.
There are many applications for unsupervised learning, including data
compression, de-noising, clustering, recommendation, abnormality detection,
classification, just to name a few.
Unsupervised machine learning methods include principle components
analysis (PCA) for principal value extraction from datasets, the family of
K-means clustering methods, mean-shift methods, and the family of Autoen-
coders. We have already discussed PCA in Chapter 3 and hence will simply
use it here for clustering algorithms. In this chapter, we shall discuss in detail
K-means clustering, mean-shift clustering, and Autoencoders. We shall focus
more on the fundamentals of these methods to help readers understand other
related methods and techniques that are not covered in this book.

17.2 K-means for Clustering

K-means clustering is an unsupervised machine learning algorithm using

unlabeled datasets. It is originally from signal processing for partitioning a

585
586 Machine Learning with Python: Theory and Applications

cloud of data-points into k clusters. In each cluster, data-points have the

nearest mean with respect to its cluster centers or centroid. K-means clus-
tering minimizes within-cluster variances, that is, the squared (Euclidean)
distances (hence, it becomes a least squares minimization problem, which is
quadratic). Such an algorithm essentially partitions these data-points into
Voronoi cells generated using the k centroids.
The clustering problem is in general NP-hard, meaning that it is
computationally intractable in theory. However, efficient heuristic algorithms
have been developed to effectively find local optimal solutions. The K-means
is one such algorithm that clusters given unlabeled data-points into pre-
specified k clusters. It is loosely related to the k-nearest neighbor classifier
and it has a number of variations. The mean-shift clustering algorithm is one
such variation, in which the number of clusters need not be pre-specified.
More details on K-means and its relation to other algorithms can be found
at the Wikipedia page (https://en.wikipedia.org/wiki/K-means clustering)
and the widely used Sklearn sites.
The typical K-means algorithm uses an iterative updating process. It
is referred to as Lloyd’s algorithm. Here, we discuss in detail this basic
algorithm, hoping this can be of help in understanding and using other
related and more advanced algorithms. The K-means algorithm starts with
an initialization of k means.

17.2.1 Initialization of means

Consider a dataset with a cloud of n data-points in the feature space Xp
with p features. Our objective is to group these data-points into k clusters
(sub-clouds), by extracting k geometrical distributive characteristics of these
data-points. A typical K-means algorithm starts with generating k initial
means randomly. A discussion on random initiation methods is given in a
paper by Hamerly and Elkan [1]. One of the methods is to randomly choose
k points from the given dataset and use these as the initial means, which
of course can be far away from the true ones. Another method randomly
assigns clusters and then computes the initial means for these clusters. In this
case, this initial clustering will also be far away from the true clustering. In
either case, the initial start has some randomness involved. Hence, it is often
necessary to run the K-means for a number of times each with a random set
of initial means and then choose the best one based on some criterion.
There are a number of specially designed strategies for K-means initia-
tion, including centroids of random subsamples, group representative points,
random farthest points, or k-means++. Interested readers may refer to this
 Unsupervised Learning Techniques 587

online discussion page (https://stats.stackexchange.com/questions/317493/

methods-ofinitializing-k-means-clustering) for more details. We will use
k-means++ in our example problems later in benchmark studies.
The PCA (discussed in Chapter 3) can be a good choice for generating the
initial k means. One can conduct a PCA to obtain k principal components
using the given dataset, and then use those principal components as the
initial k means. In this case, the initiation becomes somewhat deterministic,
and only one K-means run is conducted. It is believed to improve the
accuracy of the clustering results because the K-means is known to be
sensitive to scaling, and PCA can offer more stable scaling and also noise
reduction features. PCA can also help in convergence behavior because of
possible better initial guesses. Moreover, PCA is used for visualizing the
results of K-means clustering. As discussed in Chapter 3, PCA can be
expensive. Fortunately, in K-means applications, k is usually not a big
number (compared to the number of samples in big datasets), and hence the
PCA cost is usually not a big issue for many problems. We shall examine
this using example problems later. Let us now look at the basic procedure,
formulation, and algorithm for typical K-means clustering. It is iterative,
and each iteration has mainly two steps.

17.2.2 Assignment of data-points to clusters

(0) (0) (0)
At t = 0, we have a set of initial k means m1 , m2 , . . . , mk corresponding
to these initial k clusters for a given cloud of data-points in Xp . These means
are also referred to as centroids or centers of the clusters. Note that in K-
(t)
means clustering, we pre-specify the number of clusters k. Also, mi is in
general a vector in Xp .
At step t, we compute squared Euclidean distances between a data-point
xq to these means, and then assign this data-point to the cluster with the
“nearest mean”, meaning that the Euclidean distance between the data-point
and the cluster mean is the shortest.

(t) (t) (t)

Si = {xq : xq − mi 2 ≤ xq − mj 2 ∀j, 1 ≤ j ≤ k}, i = 1, 2, . . . , k
(17.1)

If there is more than one nearest means, we choose one of those, so that
(t)
data-point xq is assigned only to one set Si . This process is done for all the
data-points in the dataset, which leads to a partitioning of all these data-
points by the edges of the Voronoi diagram generated by these means (this
588 Machine Learning with Python: Theory and Applications

will be shown more clearly in the case study examples later). Each Voronoi
(t)
cell hosts a cluster with data-points in set Si .

17.2.3 Update of means

When partitioning is done, we re-compute the mean for each of the k

clusters using these data-points assigned to the cluster. For the ith cluster,
we have

(t+1) 1 
mi = xj , i = 1, 2, . . . , k (17.2)
ni (t)
xj ∈Si

(t)
where ni is the number of data-points in set Si .
This leads to an updated set of k means for the t + 1 step, which is then
used for a new round of iteration: re-assignment of data-points to a new set
of k clusters using Eq. (17.1), and then re-update the k means for the t + 2
step. This is done until the process converges, meaning that these k means
do not change significantly based on some criteria.
Note that this algorithm does not guarantee finding the global optimum
[2]. In practice, however, the K-means algorithm is fast, and it may be one
of the fastest clustering algorithms available for local minima. Because the
initiation of the k means is of random nature, one may need to perform the
clustering several times and choose the best result.
Lloyd’s algorithm discussed above has an average complexity of O(kmT ),
where m is the number of data-points and T is the number of iterations.
Even faster variations of Lloyd’s algorithm have been developed, one of
which is Elkan’s algorithm, which uses the triangular inequality property
to significantly improve the efficiency. Elkan’s algorithm is quite frequently
used, and we will demonstrate Elkan’s algorithm in the case study examples.
Note also that one may use a different distance norm measure other than
the squared Euclidean distance for these K-means type of algorithms. This
may change the convergence behavior. More details on this can be found
at the Wikipedia page (https://en.wikipedia.org/wiki/K-means clustering)
and the links therein.
Figure 17.1 is a picture taken from a nicely made animation by
Chire with CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0)
via the Wikimedia Commons (https://commons.wikimedia.org/wiki/File:K-
means convergence.gif). This animation shows clearly how the iteration is
performed in a typical K-means algorithm.
 Unsupervised Learning Techniques 589

Figure 17.1: Convergence process of a typical K-means algorithm. The data-points are
given and fixed in 2D feature space, which are to be grouped in k = 3 clusters: blue, yellow,
and red ones.

To examine the process more closely, we plot snapshots at four iterations:

0th (initial), 2nd, 5th, and 14th, as shown in the Fig. 17.2.
As shown in Fig. 17.1, the data-points are given and fixed in 2D space,
which are to be grouped in k = 3 clusters. At the beginning, a set of 3
means is generated (randomly or otherwise) based on these data-points.
These three means (cluster centers) are marked with a blue circle, yellow x,
and red cross. Based on these 3 means, a Voronoi diagram is formed, via
the “nearest mean” criterion by computing the squared Euclidean distances
of the data-points to these 3 means. The edges (these 3 black lines) of the
Voronoi diagram divide the data-points into 3 clusters: the blue, yellow, and
red ones. At the 0th iteration, these 3 means are quite randomly placed. The
blue circle is very close to the yellow x, and both are far apart from the red
cross. Next, 3 means of the coordinates of the data-points for these 3 clusters
are computed, using Eq. (17.2). This results in 3 updated means. At the 2nd
iteration, these 3 means are more reasonably separated and the data-points
are also better clustered. With the updated 3 means, the data-points are
re-clustered via the nearest mean criterion. This process iterates until the
update becomes sufficiently small. The clustering process converges. At the
590 Machine Learning with Python: Theory and Applications

Figure 17.2: Examination of the converging process of a typical K-means algorithm. The
data-points are given in 2D feature space, and are to be grouped into k = 3 clusters: blue,
yellow, and red ones.

5th iteration, the clustering is already quite good and not too far from that
achieved at the final 14th iteration.
Note that when the number of the clusters k increases, the Voronoi
diagram will become more complicated, as will be shown in one of the
examples later for a 10-digit clustering problem.

17.2.4 Example 1: Case studies on comparison of initiation

methods for K-means clustering

In this first example, we compare the various initialization methods for

K-means clustering in terms of computation time and quality of the
clustering results, using various evaluation metrics. We will use randomly
generated synthetic data for this study, so that we can design four different
cases for this study. We will write a simple code for this study using routines
available at Scikit-learn [3].
 Unsupervised Learning Techniques 591

17.2.4.1 Define a function for benchmarking study

We will use the following function to evaluate the performance of the method.
It uses the SKlearn class make pipeline to compute the evaluation metrics.
This function will be called when a K-means run is completed using an
initiation means.

# From Sklearn online examples

# import all necessary modules
import numpy as np
import matplotlib.pyplot as plt
from sklearn.cluster import MeanShift, estimate_bandwidth
from sklearn.datasets import make_blobs
from sklearn.datasets import load_digits
from sklearn import metrics
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from time import time
%matplotlib inline

def bench_k_means(kmeans, name, data, labels):

"""To evaluate the KMeans initialization methods.

Parameters
----------
kmeans : KMeans instance
class:`~sklearn.cluster.KMeans` instance with the
initialization already set.
name : str
Name given to the strategy. It will be used to show
the results in a table.
data : ndarray of shape (n_samples, n_features)
The data for clustering.
labels : ndarray of shape (n_samples,)
The labels used to compute the clustering metrics
which require some supervision.
"""
t0 = time()
(n_samples, n_features), n_k = data.shape,
np.unique(labels).size
estimator = make_pipeline(StandardScaler(), kmeans).fit(data)
fit_time = time() - t0
results = [name, fit_time, estimator[-1].inertia_]
592 Machine Learning with Python: Theory and Applications

# Define the metrics which require only the true labels

# and estimated labels
clustering_metrics = [metrics.homogeneity_score,
metrics.completeness_score,
metrics.v_measure_score,
metrics.adjusted_rand_score,
metrics.adjusted_mutual_info_score,]

results+=[m(labels,estimator[-1].labels_) for m in
clustering_metrics]

# The silhouette score requires the full dataset

results+=[metrics.silhouette_score(data, estimator[-1].labels_,
metric="euclidean", sample_size=n_samples,)]

# Show the results

formatter_result = ("{:9s}\t{:.3f}s\t{:.0f}\t{:.3f}\t{:.3f}"
"\t{:.3f}\t{:.3f}\t{:.3f}\t{:.3f}")
print(formatter_result.format(*results))

# function to print out key information about the dataset

def print_data_info(data, labels,n_features,n_k,n_samples,n_comp):

print('data=',data[0:1,0:10], '\n data.shape=', data.shape)

print(f"No of clusters k:{n_k}; No of samples:{n_samples};

"+ f"No of features p:{n_features};
No of PCA comps:{n_comp}")

print('data.shape:',data.shape)
print('labels:',labels,' Shape:',labels.shape)

#____________________________________________________________

# Function to print out evaluation metrics

def print_evaluation_metrics():
print('inertia: sum of squared distances of samples to their
centroid')
print('homo: homogeneity [0,1] of the clusters with their own
data-points')
print('compl: completeness [0,1], how well the same
class points fit the clusters')
print('v_meas: v_measure score [0,1], consistence btw two
clustering results')
 Unsupervised Learning Techniques 593

print('ARI: adjusted rand index [-1,1], distances btw

different sample splits')
print('AMI: adjusted mutual info index [0,1], the mutual
infor. for two splits')
print('silht: silhouette [-1,1] mean-distance ratio of
in- & btw-clusters points')

#____________________________________________________________
# Define a function to plot the clusters
# The data needs to be reduced to 2D, if it not already 2D
# kmeans needs already been fitted

def plot_clusters(reduced_data,kmeans,fig_title):

# Mesh step size. Decrease to increase the quality of VQ.

h = .02 # points in [x_min, x_max]x[y_min, y_max].

plt.figure(figsize=(4.0, 4.0),dpi=100)
# Plot the decision boundary. Assign a color to each
x_min, x_max = reduced_data[:, 0].min() - 1,
reduced_data[:, 0].max() + 1
y_min, y_max = reduced_data[:, 1].min() - 1,
reduced_data[:, 1].max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
np.arange(y_min, y_max, h))

# Labels for each point in mesh. Use last trained model.

Z = kmeans.predict(np.c_[xx.ravel(), yy.ravel()])

# Put the result into a color plot

Z = Z.reshape(xx.shape)
plt.figure(1)
plt.clf()
plt.imshow(Z, interpolation="nearest",
extent=(xx.min(), xx.max(), yy.min(), yy.max()),
cmap=plt.cm.Paired, aspect="auto", origin="lower")

plt.plot(reduced_data[:, 0], reduced_data[:, 1], 'k.',

markersize=2)

# Plot the centroids as a white X

centroids = kmeans.cluster_centers_
plt.scatter(centroids[:, 0], centroids[:, 1], marker="x",
s=169, linewidths=3, color="w", zorder=10)

plt.title(fig_title)
594 Machine Learning with Python: Theory and Applications

#xyd = 3.5
xyd = min(abs(x_min), x_max, abs(y_min), y_max)
x_min, x_max, y_min, y_max = -xyd, xyd, -xyd, xyd
plt.xlim(x_min, x_max)
plt.ylim(y_min, y_max)

plt.xticks(())
plt.yticks(())
plt.show()

#print('centroids=',centroids)

17.2.4.2 Generation of synthetic data-points

We now generate a set of synthetic data-points using the SKlearn make blobs
class in a random manner. In this example, we designed a total of four study
cases, which are detailed in the cell below. Readers may simply assign a
number (from 1 to 4) to a variable “Case”, and the date points for this case
will then be generated.

# Choose a case to generate a set of synthetic data-points.

# Number of features (n_features) and number of clusters (k)
# will be computed after the data-points are generated

Case=3 # 1 or 2 or 3 or 4 # 4 cases designed for this example

if Case == 1:
#_______________________________________________________
# Case 1: synthetic sample with 2 groups of data-points
# in 2D space, n_features=2, n_k=2
centers = [[1, 1], [-1, -1]] # centers for 2D points
#_______________________________________________________
elif Case == 2:
# Case 2: synthetic sample data that has 3 groups of
# data-points in 3D space, n_features=3, n_k=3
centers = [[1, 1, 1], [-1,-1,-1], [1,-1,0.5]] # for 3D points
#_______________________________________________________
elif Case == 3:
# Case 3: synthetic sample data that has 3 groups of
# data-points in 2D space, n_features=2, n_k=3
centers = [[1, 1], [-1,-1], [1,-1]] # set 3 centers 2D points
#_______________________________________________________
 Unsupervised Learning Techniques 595

elif Case == 4:
# Case 4: synthetic sample data that has 2 groups of
# data-points in 6D space, n_features=6, n_k=5
centers = [[ 1, 0, 1, 1, 1, 0], [-1,-1,-1,-1, 0,-1],
[1,-1,-1, 0,-1,-1], [ 1, 1, 0,-1,-1,-1],
[ 0, 1, 1, 0,-1, -1]]
# set 5 centers for 5 group of data-points in 6D
#________________________________________________________
else:
print('There are a total of 4 cases for this example!!!')

print('Cluster centers set for the data-points =', centers)

# Generation of data-points
n_s, c_std = 1888, 0.4888 # n_s: number of samples;
#c_std: standard deviation for random data-points generation
data,labels=make_blobs(n_samples=n_s,centers=centers,
cluster_std=c_std) # number of samples

(n_samples, n_features), n_k = data.shape, np.unique(labels).size

n_comp = min(n_features, n_k) # number of prin. cmpnts when

# PCA is used to create initial means for K-means

# print out key data information

print('This example has a total of 4 cases. This is Case-'
+str(Case)+'; Data:'+str(n_features)+'D')

print_data_info(data, labels,n_features,n_k,n_samples,n_comp)

Cluster centers set for the data-points =

[[1, 1], [-1, -1], [1, -1]]
This example has a total of 4 cases. This is Case-3; Data:2D
data= [[ 0.60389228 -1.62706035]]
data.shape= (1888, 2)
No of clusters k:3; No of samples:1888; No of features p:2;
No of PCA comps:2
data.shape: (1888, 2)
labels: [2 0 2 ... 1 0 2] Shape: (1888,)
596 Machine Learning with Python: Theory and Applications

17.2.4.3 Examination of different initiation methods

We conduct a comparison study on different initialization methods for

K-means clustering. This includes three initiation methods:

• Initialization using the kmeans++ method. This is a random-farthest-point

for initial mean generation. We will run the initiation and clustering for
n i times, and choose the best results based on the inertia value.
• Initialization using “random” option. This is also a pure random method.
We will run the initiation and clustering for n i times, and choose the best
results based on the inertia value.
• Initialization using PCA. The principal components of the PCA will be
used as the initial means. This is a deterministic method, and a single run
shall suffice.

# Comparison study runs

from sklearn.cluster import KMeans
from sklearn.decomposition import PCA

print('This example has a total of 4 cases. This is Case-'+

str(Case)+'; Data:'+str(n_features)+'D')
print_data_info(data, labels,n_features,n_k,n_samples,n_comp)
print(78 * '_')
print('init\t\ttime\tinertia\thomo\tcompl\tv-meas\tARI\
tAMI \tsilht')
n_i = 5 # The times the k-means algorithm will run with
# different centroid seeds. The final results will be the
# best output of n_i consecutive runs in terms of inertia.

if Case != 3: # (2D, n_features=2, n_k=3, PCA not applicable

pca = PCA(n_components=n_comp).fit(data)
kmeans = KMeans(init=pca.components_, n_clusters=n_k, n_init=1)
bench_k_means(kmeans=kmeans,name="PCA-init",data=data,
labels=labels)
kmeans = KMeans(init="random",n_clusters=n_k,n_init=n_i,
random_state=0)
# by default it uses the Elkan algorithm
bench_k_means(kmeans=kmeans, name="random", data=data,
labels=labels)

kmeans = KMeans(init="k-means++", n_clusters=n_k, n_init=n_i,

random_state=0)
 Unsupervised Learning Techniques 597

bench_k_means(kmeans=kmeans, name="k-means++",data=data,
labels=labels)
r_data = PCA(n_components=n_comp).fit_transform(data)
bench_k_means(kmeans=kmeans,name="k-means++pca",data=r_data,
labels=labels)

print(78 * '_')
print_evaluation_metrics()

This example has a total of 4 cases. This is Case-3; Data:2D

data= [[ 0.60389228 -1.62706035]]
data.shape= (1888, 2)
No of clusters k:3; No of samples:1888; No of features p:2;
No of PCA comps:2
data.shape: (1888, 2)
labels: [2 0 2 ... 1 0 2] Shape: (1888,)

init time inertia homo compl v-meas ARI AMI silht

random 0.018s 763 0.890 0.890 0.890 0.929 0.890 0.561
k-means++ 0.015s 763 0.890 0.890 0.890 0.929 0.890 0.561
k-means++pca 0.013s 892 0.830 0.830 0.830 0.878 0.830 0.555

inertia: sum of squared distances of samples to their centroid

homo: homogeneity [0,1] of the clusters with their own data-points
compl: completeness [0,1], how well the same class points fit the clusters
v meas: v measure score [0,1], consistence btw two clustering results
ARI: adjusted rand index [-1,1], distances btw different sample splits
AMI: adjusted mutual info index [0,1], the mutual infor. for two splits
silht: silhouette [-1,1] mean-distance ratio of in- & btw-clusters points

Set Case = 1 in the codes given above; the clustering has been completed
on the author’s laptop and the results are summarized in the same output
table above. We may observe the following:

• First is the computation time for clustering. They are all very fast in
the order of 20 ms. The K-means with PCA initiated means (the PCA-
init method) takes about 7 ms, and it runs fastest (without counting the
time for the PCA). This is because it runs only once. The time for the
other three methods is roughly the same, and slower than the PCA-init,
because they are set to run for 5 times. Because our toy problem with
synthetic data is very small in scale, the measure of computation time is
not accurate. We cannot make much point.
• Next is the quality of clustering. These first three methods listed in the
output table give the same quality results in all the measuring metrics.
Because this toy problem is too simple, there may just be one global
598 Machine Learning with Python: Theory and Applications

optimal solution, and any of these methods give the same answer. However,
the k-means++pca gives a result that is different from that of all the other
three methods, especially on the inertia measure. This is because the K-
means is done using the PCA processed data. For Case-1, the PCA did not
reduce the data dimension, but projected the dataset to its two principle
axes. The dataset in the principle component space become flatter (see
Chapter 3). Thus, the measured inertia becomes much larger. The scores
measured in other metrics have some change but not very much.
• Using the same code, we set the n init to 1 (instead for 5) and run. We
obtained the same results. This also supports the argument that there is
likely just one optimal solution to this toy problem, because randomly
selected initial means do not give a different solution for all the methods
used. Readers can easily confirm this using the code given.
Readers may set Case to other numbers (2 or 3 or 4). Similar observations
can be made for all these toy cases.

17.2.4.4 Visualize the clustering results

Our toy example problem has two cases (Case-2 and Case-4) with higher-
dimensional data in the feature space (p = 3 and p = 6). K-means can have
all these datasets clustered in these high dimensions. However, it is difficult
to visualize these clusters in high dimension. Therefore, we will use PCA to
project the datasets into a 2D principal component space, conduct K-means
clustering, and then plot the data-points and their clusters on a 2D plane.
This is done using the following codes and the plot clusters() defined earlier:

import matplotlib.pyplot as plt

if Case == 1 or Case == 3: # Case 1 or 3: data in 2D space,

n_k =2 or 3
#________________________________________________________
pca_compf = 0 # no PCA used for plotting the data
reduced_data = data # no dimension reduction
#________________________________________________________
elif Case == 2 or Case == 4: # Case 2: 3 group data in 3D space,
# n_features=3, n_k =3
# Case 4: 5 group data in 6D space,
# n_features=6, n_k=5
pca_compf = 2 # dimension reduction to show in 2D
reduced_data = PCA(n_components=pca_compf).fit_transform(data)
#________________________________________________________
 Unsupervised Learning Techniques 599

else:
print('There are a total of 4 cases for this example!!!')

print_data_info(data, labels,n_features,n_k,n_samples,n_comp)
kmeans = KMeans(init="k-means++", n_clusters=n_k, n_init=n_i)
kmeans.fit(reduced_data)
fig_title = "Case-"+str(Case)
plot_clusters(reduced_data,kmeans,fig_title)
print('Cluster centers set for the data-points =', centers)
data= [[ 0.60389228 -1.62706035]]
data.shape= (1888, 2)
No of clusters k:3; No of samples:1888; No of features p:2;
No of PCA comps:2
data.shape: (1888, 2)
labels: [2 0 2 ... 1 0 2] Shape: (1888,)

Figure 17.3: Clustered data-points for Case-3 using the K-means.

Cluster centers set for the data-points = [[1,1], [-1,-1], [1,-1]]

Figure 17.4 plots the clusters and the data-points for all these four cases.
We note the following:
• All these cases are well clustered using K-means algorithm, with the
converged centroids (marked with white X) each for a cluster.
600 Machine Learning with Python: Theory and Applications

Figure 17.4: K-means clusters for synthetic data-points generated randomly for four
different cases of datasets. The original datasets for Case-2 and Case-4 are in higher-
dimensional feature spaces. For these datasets, its feature space is reduced to 2D and then
the K-means clustering is performed.

• These clusters are hosted in the colored Voronoi cells (one cluster per
color), and the boundaries between these Voronoi cells divide these
clusters.
• The line that connects these centroids (for each of the cases) is perpendic-
ular (orthogonal) to the Voronoi cell boundary (or its extended straight
line). This orthogonality is proof that these cells are essentially the Voronoi
cells by definition. It is produced by minimizing the squared Euclidean
distances of a point in cluster to its centroid in the K-means clustering
algorithm.

We are now ready to look at a more complicated real-life example.

 Unsupervised Learning Techniques 601

17.2.5 Example 2: K-means clustering on the handwritten

digit dataset

We now apply the K-means clustering algorithm to a set of images of hand-

written digits. This dataset is the test set of the UCI ML hand-written
digits (https://archive.ics.uci.edu/ml/datasets/Optical+Recognition+of+
Handwritten+Digits). The functions defined earlier for our toy dataset
example can be used as is for this study. We can also use the same metric
for evaluation of the efficiency and quality of the clustering.

17.2.5.1 Load handwritten digit dataset

We first load the dataset using the SKlearn load digits() and perform the
clustering using the same SKlearn’s K-means. This dataset contains images
of 10 handwritten digits from 0 to 9. We would like each cluster to contain
the images of the same digits. We thus have k = 10 clusters for the problem.
Each of these images has 8 × 8 pixel, meaning that the number of features
is 64.
# Import necessary module for this task
import numpy as np
from sklearn.datasets import load_digits
import matplotlib.pyplot as plt
%matplotlib inline
# load the digits dataset
data, labels = load_digits(return_X_y=True)
(n_samples,n_features),n_digits=data.shape,np.unique(labels).size
n_k = n_digits
n_comp = 2
print_data_info(data, labels,n_features,n_k,n_samples,n_comp)
print('Take a look at the image for digit 9:')
plt.imshow(data[9].reshape(8,8))

data= [[ 0. 0. 5. 13. 9. 1. 0. 0. 0. 0.]]

data.shape= (1797, 64)
No of clusters k:10; No of samples:1797; No of features p:64;
No of PCA comps:2
data.shape: (1797, 64)
labels: [0 1 2 ... 8 9 8] Shape: (1797,)
Take a look at the image for digit 9:

<matplotlib.image.AxesImage at 0x1fdcd0ca1d0>
602 Machine Learning with Python: Theory and Applications

Figure 17.5: Sample image of a handwritten digit in the dataset of UCI ML hand-written
digits.

As shown in Fig. 17.5, in this dataset, each of the images has 64 features
that are the image values at the 8 × 8 pixels.

17.2.5.2 Examination of different initiation methods

We conduct a comparison study on different initialization methods for K-
means clustering. We compare the same three initiation methods as in the
previous example and in a similar manner.

from sklearn.cluster import KMeans

from sklearn.decomposition import PCA

print('data.shape:',data.shape)
print(f"Number of digits (k):{n_digits}; Number of samples (n):\
{n_samples};Number of features (p) {n_features}")
print(78 * '_')
print('init\t\ttime\tinertia\thomo\tcompl\tv-meas\tARI\
tAMI\tsilht')
n_i = 5
pca = PCA(n_components=n_digits).fit(data)
kmeans = KMeans(init=pca.components_, n_clusters=n_digits,
n_init=1)
bench_k_means(kmeans=kmeans, name="PCA-init", data=data,
labels=labels) # by default it uses the Elkan algorithm
kmeans = KMeans(init="random",n_clusters=n_digits,n_init=n_i,
random_state=0)
bench_k_means(kmeans=kmeans, name="random", data=data,
labels=labels)
 Unsupervised Learning Techniques 603

kmeans = KMeans(init="k-means++", n_clusters=n_digits, n_init=n_i,

random_state=0)
bench_k_means(kmeans=kmeans, name="k-means++", data=data,
labels=labels)

n_compr = 10 # n_digits
r_data = PCA(n_components=n_compr).fit_transform(data)
bench_k_means(kmeans=kmeans,name="k-means++pca",data=r_data,
labels=labels)

print(78 * '_')
print_evaluation_metrics()

data.shape: (1797, 64)

Number of digits (k):10; Number of samples (n):1797;
Number of features (p) 64

init time inertia homo compl v-meas ARI AMI silht

PCA-init 0.065s 72686 0.636 0.658 0.647 0.521 0.643 0.146
random 0.089s 69952 0.545 0.616 0.578 0.415 0.574 0.137
k-means++ 0.143s 69485 0.613 0.660 0.636 0.482 0.632 0.162
k-means++pca 0.072s 9169 0.722 0.727 0.724 0.650 0.721 0.246

inertia: sum of squared distances of samples to their centroid

homo: homogeneity [0,1] of the clusters with their own data-points
compl: completeness [0,1], how well the same class points fit the clusters
v meas: v measure score [0,1], consistence btw two clustering results
ARI: adjusted rand index [-1,1], distances btw different sample splits
AMI: adjusted mutual info index [0,1], the mutual infor. for two splits
silht: silhouette [-1,1] mean-distance ratio of in- & btw-clusters points

The results in the above output table tell us the following:

• First is the computation time taken for the clustering. They are all very
fast for this handwritten digit dataset. The K-means with PCA initiated
means (the PCA-init method) takes only ∼20 ms. It runs fastest (without
counting the time for the PCA). This is because it runs only once. The
time for the other three methods is roughly the same, and slower than
the PCA-init, because they were set to run for 5 times. The k-means++
method took about 5 times more time. Note that the runtime can also
change depending on whether or not the computer is on other tasks at the
same time.
• The inertia value is the sum of squared distances of samples to their
centroid. The random and k-means++ methods give quite close inertia
values. The PCA-init gives a higher value and the k-means++pca gives
a much smaller value. This is because the use of PCA has reduced the
604 Machine Learning with Python: Theory and Applications

feature space dimension from 64 to 10, before the k-means++ clustering,

leading to significant reduction in inertial value.
• The other five evaluation metrics, homo, compl, v-meas, ARI, and AMI,
are objective measures on the quality of the clustered results, because
these measures all use the true labels of the original data as the basis for
comparison, although the measurement methods are different. Based on
the output table above, PCA-init is of least quality, partially because
it runs only once. The random method achieved better scores for all
these measures compared to PCA-init, partially because it ran five times.
Readers may change n init number to 1 and rerun the code, and we
shall still find that the random method performs better for this example.
The k-means++ method is slightly better than the random method. The
k-means++pca gives the best results, that is, quite significantly better
than the k-means++ on all these measures. This is largely due to
the dimension reduction to the original dataset before the k-means++
clustering.
• The silhouette score (shown in the last column in the output table) is
computed without the use of the true labels. It measures the ratio of the
mean distances of the data-points within the clusters and that between
the clusters, based on the clustered results. This measure is less objective,
because the true labels are not used as the basis for comparison. However,
this measure is useful because we do not usually have true labels in
unsupervised learning. In this silhouette measure, the PCA-init gives least
score. Both random and k-means++ methods give a higher score. The best
performer is still the k-means++pca method. Note that the silhouette
measure is also useful in K-means methods that do not pre-specify the
number of the clusters k. It can be used to determine the best k number
for unlabeled datasets.

Readers can easily try out different setting and the run this example using
the code given.
Let us now plot the clustered data-points in 2D plane with the help of
(again) the PCA.

17.2.5.3 Visualize the results for handwritten digit clustering using PCA
The dataset for the handwritten digit is 64-dimensional in the feature
space. It is not possible for us to visualize the clustered results is such a
high dimension. We thus use the PCA to reduce the feature space to a
two-dimensional PCA principal components space, and then plot the data
together with the clusters.
 Unsupervised Learning Techniques 605

import matplotlib.pyplot as plt

pca_components = 2 # we can only view this well in 2D space
reduced_data = PCA(n_components=pca_components).
fit_transform(r_data)
kmeans = KMeans(init="k-means++", n_clusters=n_digits, n_init=n_i)
kmeans.fit(reduced_data)
fig_title = str(n_digits)+" handwritten digits"
plot_clusters(reduced_data,kmeans,fig_title)

Figure 17.6: Results of the 10-means clustering using the dataset of UCI ML hand-written
digits.

Notice from Fig. 17.6 that we have a more complicated Voronoi diagram
for the 10-means clustering problem.

17.3 Mean-Shift for Clustering Without Pre-Specifying k

In using a K-means type of algorithm, we usually need to pre-specify the

number of clusters k. This implies that we do know some information about
the dataset. In many unsupervised learning applications, we do not even
know k, and we need to cluster the data-points in a dataset. This section
discusses one such method, called mean-shift method, for such a task. Mean
shift clustering aims to discover “blobs” in a “cloud” of data-points. It is a
centroid-based algorithm, which works by updating candidates for centroids
to be the mean of the data-points within a given region. These candidates
are then filtered in a post-processing stage to eliminate near duplicates to
form the final set of centroids. More details can be found in the Sklearn
documentation and Ref. [4].
606 Machine Learning with Python: Theory and Applications

Let us look at a demonstration of the mean-shift clustering algorithm

that uses a flat kernel. This demonstration is available at Scikit-learn [3].

%matplotlib inline
print(__doc__)
import numpy as np
from sklearn.cluster import MeanShift, estimate_bandwidth
from sklearn.datasets import make_blobs

# Generate synthetic sample data that has 3 groups of data-points

n_s, c_std = 1888, 0.4888
centers = [[1, 1], [-1, -1], [1, -1]] # set 3 centers of these
# data-points

X, _ = make_blobs(n_samples=n_s, centers=centers,
cluster_std=c_std)
print('X=',X[0:2:], ' X.shape=', X.shape)
# This dataset is the same as Case-3 in Example 1

# Compute clustering with MeanShift

# First, estimate the bandwidth, it is needed in the MeanShift
# algorithm
bandwidth = estimate_bandwidth(X, quantile=0.2, n_samples=500)
print('Bandwidth=', bandwidth)

# MeanShift clustering
ms = MeanShift(bandwidth=bandwidth, bin_seeding=True)
ms.fit(X)
labels = ms.labels_
cluster_centers = ms.cluster_centers_
# print('Predicted cluster centers=', cluster_centers)

labels_unique = np.unique(labels)
# Find the unique elements in an array
n_clusters_ = len(labels_unique)

print("Number of estimated clusters : %d" % n_clusters_)

# Plot the clustering result
import matplotlib.pyplot as plt
from itertools import cycle

fig_title ='MeanShift Clusters'

plot_clusters(X,ms,fig_title)
print('Cluster centers set for generating the dataset =', centers)
centroids_kmeans = [[ 0.98774496, 0.99979977],
[-0.96779894 -0.9892344 ], [ 1.02407913, -0.98301746]]
 Unsupervised Learning Techniques 607

print('Cluster centers found by the kmeans++ was =',

centroids_kmeans)

Automatically created module for IPython interactive environment

X= [[ 0.55586883 -0.56830437]
[ 1.02971177 -1.1907716 ]] X.shape= (1888, 2)
Bandwidth= 0.8892415258145057
Number of estimated clusters : 3

Figure 17.7: An example of MeanShift clustering resulting in three clusters.

Cluster centers set for generating the dataset = [[1, 1],

[-1, -1], [1, -1]]
Cluster centers found by the kmeans++ was = [[0.98774496,
0.99979977], [-1.95703334], [1.02407913, -0.98301746]]

# Compute clustering with MeanShift

# Estimate the bandwidth, needed in the MeanShift algorithm
# The complexity of estimate_bandwidth function is at
# least quadratic in n_samples. For large datasets, use a
# small value.
bandwidth = estimate_bandwidth(reduced_data, quantile=0.0519,
n_samples=450)
# Note: estimate_bandwidth can be sensitive to the
# parameters: quantile and n_samples.
print('Bandwidth=', bandwidth)
608 Machine Learning with Python: Theory and Applications

# MeanShift clustering
ms = MeanShift(bandwidth=bandwidth, bin_seeding=True)
ms.fit(reduced_data)
labels = ms.labels_
cluster_centers = ms.cluster_centers_
# print('Predicted cluster centers=', cluster_centers)

labels_unique = np.unique(labels)
# Find the unique elements in an array
n_clusters_ = len(labels_unique)
print("Number of estimated clusters : %d" % n_clusters_)
fig_title ='MeanShift for Digits'
plot_clusters(reduced_data,ms,fig_title)
print('Centroids by MeanShift is: Centroids by
kmeans++ was:\n')
ms_cntr = np.sort(ms.cluster_centers_,axis=0)
# sort for easy comparison
km_cntr = np.sort(kmeans.cluster_centers_,axis=0)
for i in range(n_clusters_):
print (f'{ms_cntr[i,0]:12.4f}{ms_cntr[i,1]:9.4f}'
f'{km_cntr[i,0]:19.4f}{km_cntr[i,1]:9.4f}')

Bandwidth= 5.722875000219844
Number of estimated clusters : 10

Figure 17.8: Results of a MeanShift clustering using the dataset of UCI ML hand-written
digits.
 Unsupervised Learning Techniques 609

Centroids by MeanShift is: Centroids by kmeans++ was:

-21.2030 -19.2417 -22.0921 -19.5846

-15.6402 -16.6971 -14.1853 -16.6444
-11.0635 -11.4488 -12.8868 -8.2211
-5.0131 -6.4602 -3.7937 -6.8557
-1.9861 -1.9128 -3.6150 -0.3548
1.1829 1.0722 1.4639 0.3682
1.5387 4.0261 4.6894 5.7448
8.0474 10.8061 9.2011 11.3140
15.6139 14.2532 15.9470 12.9644
25.0669 21.7812 22.6392 21.6084

We note the following points from this study:

• This handwritten digit dataset has 64 features. To simply this problem,

we used PCA to reduce the dimension number to 10. This also allows us to
compare the MeanShift results with the K-means results that we obtained
in the previous example.
• It is found that the clustered results obtained using the MeanShift are
similar but quite different from these obtained using K-means, as shown
in results for the centroids listed in the above table.
• Note also that the estimate bandwidth can be quite sensitive to the
parameters of quantile and n samples. The setting with quantile=0.0519
and n samples=450 produces 10 clusters for the handwritten digit dataset.
Changing these to other numbers can produce totally different results first
for the bandwidth and then for the final clustering results. For example,
when quantile is changed to 0.04, we obtained 16 clusters. Therefore, care
is needed when running this kind of prediction.
• This example has also demonstrated that constructing unsupervised
machine learning models is in general a quite difficult task. Getting some
kind of information on unlabeled datasets can be helpful. More concrete
research is needed in this area, and hence it is one of the hottest topics of
study.

17.4 Autoencoders

Compressing of data is widely used, because storing and transferring bigdata

are expensive and can be very time consuming, despite advancements in
computer telecommunication technology. Techniques for compressing data
610 Machine Learning with Python: Theory and Applications

without much loss in features are of importance. A number of compression

techniques for different forms of data have already been developed. The
principle component analysis (PCA) technique is an excellent tool for
compressing data with controllable loss of accuracy. We have seen some
examples in Chapter 3. The PCA is a very general and universally useful tool
for compressing data with features that are mostly linear. Neural network
based Autoencoders [5–9] can also be useful for a variety of types of data,
including huge datasets with linear or nonlinear features.

17.4.1 Basic structure of autoencoders

An autodecoder is a technique to encode data automatically using an
artificial neural network that is trained without labeled data. It has
essentially three major parts, an encoder, bottleneck, and decoder, as
shown in Fig. 17.9.
The encoder can have one or more layers with training parameters of
weights and biases. It takes the original input dataset, encodes it with these
parameters, and then feeds it to the bottleneck neurons. The compressed

Figure 17.9: A neural network for autodecoder. It consists of two sub-nets, one for the
encoder and one for the decoder, which are bridged by the bottleneck layer z. An input
data x into the encoder, with the current training parameters, produces latent features z.
The same z is then feed into the decoder and produces an output x̃. The input x itself is
used as the label at the final output layer, and the residual r = x̃ − x is minimized to force
the output x̃ close to the input x. Once trained properly with a dataset, the autoencoder
produces the latent features at the bottleneck, that are the extracted/hidden features in
the dataset.
 Unsupervised Learning Techniques 611

data is then feeds to the decoder that may have one or more layers and
their own training parameters. The decoder decodes the compressed data to
reconstruct the data.
The network is trained with a simple criterion that the decoded (recon-
structed) data at the output layer must be as close as possible to the
original input data. This is why the training does not require the input
data to be labeled. Cost functions defined in the previous chapter can then
be used to satisfy this criterion for different types of data. The training
process is essentially the same as training any other neural network, via a
back-propagation process. Once the training is completed, all the training
parameters in these three parts are set, and the autodecoder is capable of
reconstructing the data with the compressed data when it is needed. This is
the basic strategy for an autoencoder.
Autoencoders can be built with various types of layers including dense
layers, convolution layers, their combinations, and other forms of layers.

17.4.2 Example 1: Image compression and denoising

One can easily build an Autoencoder using an existing ML package such as

Keras, by simply stacking up layers of neurons in a configuration similar to
Fig 17.9. Reader may take a look at an example and the code provided
by Ali Abdelaal in his online article “Autoencoders for Image Recon-
struction in Python and Keras” (https://stackabuse.com/autoencoders-for-
image-reconstruction-in-python-and-keras/). It uses the LFW (http://vis-
www.cs.umass.edu/lfw/) dataset and shows a detailed procedure with code
for image compression and denoising.

17.4.3 Example 2: Image segmentation

Image segmentation can be useful in autonomous vehicles, where one may

need to segment objects such as persons, animals, bicycles, and cars.
Techniques of CNNs and Autoencoders are useful to creating effective models
for image segmentations. The following are some examples of image segmen-
tation using pixel libraries (https://github.com/ayoolaolafenwa/PixelLib):

import matplotlib.pyplot as plt

import numpy as np
from PIL import Image
%matplotlib inline
612 Machine Learning with Python: Theory and Applications

import pixellib
from pixellib.semantic import semantic_segmentation
segment_image = semantic_segmentation()
segment_image.load_pascalvoc_model
('deeplabv3_xception_tf_dim_ordering_ tf_kernels.h5')
segment_image.segmentAsPascalvoc('../images/inkedcars15.jpg',
output_image_name = '../images/cars15-s.jpg')
img = Image.open('../images/inkedcars15.jpg')
plt.figure(figsize=(10, 15))
plt.imshow(img);

Figure 17.10: Original image for segmentation.

img = Image.open('../images/cars15-s.jpg')
plt.figure(figsize=(10, 15))
plt.imshow(img);

Figure 17.11: Resulting image after segmentation. The object type of the image is color
coded.
 Unsupervised Learning Techniques 613

The color code is used to distinguish the type of objects. The details of the
color codes can be found at pixel libraries site. One can also overlay the
segmented objects with the original image

segment_image.segmentAsPascalvoc('../images/inkedcars15.jpg',
output_image_name = '../images/cars15-s.jpg', overlay = True)
img = Image.open('../images/cars15-s.jpg')
plt.figure(figsize=(10, 15))
plt.imshow(img);

Figure 17.12: Overlaid image of the original and the segmented images.

import pixellib
from pixellib.semantic import semantic_segmentation
import time
from pixellib.instance import instance_segmentation
segment_image = instance_segmentation()
segment_image.load_model("mask_rcnn_coco.h5")
start = time.time()
segment_image.segmentImage("../images/inkedcars-blocked.jpg",
output_image_name = "../images/cars-blocked-s.jpg")
end = time.time()
print(f"Inference Time: {end-start:.2f}seconds")
img = Image.open("../images/inkedcars-blocked.jpg")
plt.figure(figsize=(10, 15));
plt.imshow(img);
Processed image saved successfully in your current working
directory.
Inference Time: 19.08seconds
614 Machine Learning with Python: Theory and Applications

Figure 17.13: The original image for segmentation.

img = Image.open("../images/cars-blocked-s.jpg")
plt.figure(figsize=(10, 15))
plt.imshow(img)

<matplotlib.image.AxesImage at 0x1fa39721108>

Figure 17.14: Overlaid image of the original and segmented images. The object type is
color coded.

segment_image.segmentImage("../images/inkedcars-blocked.jpg",
output_image_name = "../images/cars-blocked-sf.jpg",
show_bboxes = True)
img = Image.open("../images/cars-blocked-sf.jpg")
plt.figure(figsize=(10, 15))
plt.imshow(img);
 Unsupervised Learning Techniques 615

Figure 17.15: Overlaid image of the original and segmented images. The object type is
color coded. Objects are boxed.

Examples of Autoencoders using the MNIST dataset and the codes by the
Keras Team are available at The Keras Blog (https://blog.keras.io/building-
autoencoders-in-keras.html). Variational Autoencoders (VAE) can also be
used for segmentations. The theoretical discussion about VAE will be given in
the last section of this chapter. Autoencoders have many other applications,
including Text Embeddings (http://yaronvazana.com/2019/09/28/training-
an-autoencoder-to-generate-text-embeddings/), Anomaly detection, and
Popularity prediction, just to name a few.

17.5 Autoencoder vs. PCA

The above examples demonstrate how an Autodecoder is built, trained, and

used. To build a high-quality and well-posed Autoencoder, one would need
to have a good understanding of its mathematical foundation. Here, we try
to establish a better mathematical understanding of an Autoencoder, and
examine the similarity and difference between a single-layer Autoencoder and
the PCA that we have studied in Chapter 3. This study follows an excel-
lent article by Ranjan (https://towardsdatascience.com/build-the-right-
Autoencodertune-and-optimize-using-pca-principles-part-i-1f01f821999b).
Readers may refer to the original article for more details. Figure 17.16 is from
this article and illustrates quite clearly the mathematical operations that are
used in both an Autoencoder and a PCA. Note that the formulation is a little
different from that in our book.
Figure 17.16 shows a single-layer linear Autoencoder together with
the mathematical operations involved. The bottom of the figure shows
616 Machine Learning with Python: Theory and Applications

Figure 17.16: Relationship between an Autoencoder and PCA [10] (with permission).

the similarity of the encoding process against the PC transformation.

PC transformation projects the original input data on their principal
components to yield orthogonal features, called principal scores or latents.
The decoding process is also similar to the PCA reconstructing process using
their principal scores. These weights in both Autoencoder and PCA models
are determined by minimizing the errors between the p features. Both models
use the same input data:
• The Autoencoder has p neurons in its input layer.
• The PCA uses the same p input samples.
For the encoder in an Autoencoder and the eigenvalue decomposition
in PCA, the outcome of the mathematical operations is equivalent. In
particular, the following points may be noted:
• The code-size of the encoder layer is k in the Autoencoder.
• In PCA, k is the number of the selected principal components (PCs, or
latents). In both models, we shall have k < p, so as to achieve dimension
reduction.
• When k ≥ p, it can lead to over-representative models, and the
reconstruction error can be made zero, if the training or reconstruction
is successful. Because of the existence of multiple solutions, the solution
obtained using one of these two models may be different from the other.
• The wij in the colored cells in the encoder in the foregoing figure is for the
ith neuron in the bottleneck corresponding to the jth feature (parameter)
 Unsupervised Learning Techniques 617

of the input data, where i = 1, 2, . . . , k and j = 1, 2, . . . , p. The Wk×p is

equivalent to the matrix of the first k eigenvectors generated in a PCA
algorithm.
• If the linear activation function is used in the bottleneck neurons, the
output of the bottleneck of an Autoencoder becomes zk×1 = Wk×p xp×1 ,
which is equivalent to the k principal components (or scores) in a PCA.

For the decoder in an Autoencoder and for the reconstruction process in

a PCA, the outcome of the mathematical operations is also equivalent. In
particular, the following points may be noted:

• The number of neurons in the output layer of the decoder in the

Autoencoder and the number of outputs in PCA reconstruction are the
same, and all are equal to p, that is, the number of the features in the input
data.
• The output (reconstructed) data from the decoder (after taking the out-
puts from the bottleneck neurons) become x̂p×1 = Wp×k  zk×1 , equivalent
to the p outputs in PCA, which are the principle scores projected back
to the original space (from the principal component space), using the
transpose of the matrix of the first k eigenvectors.

The multi-colors in the output neurons in the decoder indicate that the
weights related to the bottleneck neurons may all contribute to a neuron of
the output in the Decoder.
The above comparison uses words like “similar” and “equivalent”. If one
uses linear activation functions, it is possible to train an Autoencoder to
be the same as the linear PCA, by properly enforcing a set of constraints
during the training including tying the weights in the encoder and decoder,
and orthogonal conditions on the weights. For more details, one may refer to
the article by Ranjan. Constraint Autoencoder [9] is used for feature-space
dimension-reduction for inverse analysis using Tubenets [11–13].

17.6 Variational Autoencoder (VAE)

Variational autoencoders are a recent interesting version of Autoencoder.

Instead of training to encode usual discrete features from a dataset, a VAE
learns the parameters of a probability distribution. When samples from
the learned distribution, one can reconstruct, and also generate new data
samples. Thus, VAE is thus said a “generative model”. It has an architecture
shown in Fig. (17.17), which is quite similar to a standard autoencoder.
618 Machine Learning with Python: Theory and Applications

Figure 17.17: A variational autoencoders (VAE). It consists of two sub-nets, one for the
encoder and one for the decoder, which are bridged by the bottleneck layer z that consists
of three components: vector of means μ, covariance matrix Σ, and vector  drawn from the
standard normal distribution. An input data x into the encoder, with the current training
parameters, produces μ and Σ that is used together with  to form latent features z that is
a normal distribution. The same z is then feed into the decoder and produces an output x̃.
The input x itself is used as the label at the final output layer, and the residual r = x − x̃
is minimized to force the output x̃ close to the input x, while the KL-divergence of the
standard normal distribution from the distribution of z is also minimized. Once trained
properly with a dataset, the autoencoder produces the latent distributive features at the
bottleneck, that can be used to reconstruct data-points in the input dataset, or to generate
new data-points that may not in the dataset. [Image modified based on that of EugenioTL
from Wikimedia Commons, under the CC BY-SA 4.0 license].

17.6.1.1 Key ideas in VAE

The key ideas used in the VAE is as follows.
• First, a VAE encoder converts the input samples x into in a latent space
of distribution variables controlled by two sets of latent features: means
of distribution μ and covariance matrix Σ, at the bottleneck.
• Next, we sample randomly points z from the normal distribution N (μ, Σ),

z = μ + Σ1/2 

where  is randomly generated vector from a standard normal distribution

with zero means and identify standard variations N (0, I). This is because
for normal distributions, we have

N (μ, Σ) = μ + Σ1/2 N (0, I) (17.3)

• Finally, a VAE decoder maps these sampled latent space points to

reconstruct the input data or generate new data.
• The parameters of a VAE model are trained using the following loss
function.
   
Lvae W, W̃; x, N (0, I) = L(x̃, x) + DKL qW (z|x)pW̃ (x̃|z) (17.4)
 Unsupervised Learning Techniques 619

where W is the training parameters used in the encoder, and W̃ is that

used in the decoder. This loss function consists of two parts: one is the
reconstruction loss forcing the reconstructed samples to match the inputs. A
loss function defined in Chapters 10, 11 and 12 may be used. The other is a
KL-divergence defined in Chapter 4, which measures the differences between
the latent distribution qW (z|x) and the prior distribution pW̃ (x̃|z) acting as
a regularization term. The KL-divergence needs to minimized so that the
distributions is forced to match the prior distribution.
Note that the KL-divergence for any two normal distributions has a
simple closed-form in terms of only their means and variances, which is
critical for the practical implementation of VAEs. To show this, we use the
KL-divergence of normal distributions in continuous form.

17.6.1.2 KL-divergence for two single-variable normal distributions

Consider two normal distributions p(x) and q(x) of a single-variable x ∈ R1 .
Their probability density functions are given by
 2
x−μp
1 − √
p(x) = N (μp , σp2 ) = √ e 2σp
;
2πσp
 2 (17.5)
x−μq
1 − √
q(x) = N (μq , σq2 ) = √ e 2σq

2πσq

where μ and σ are, respectively, the mean and the standard deviation of the
distribution specified by the subscript. Note in this case, both p(x) and p(x)
are scalar functions of x, and their distributions are full controlled by their
means and standard deviations.
The KL-divergence from q(x) to p(x) can be written as
 
DKL (qp) = q(x) log(q(x)) − log(p(x)) dx

 


1 x − μq 2 1 x − μp 2
= q(x) − log(σq ) − + log(σp ) + dx
2 σq 2 σp
 



σp 1 x − μp 2 1 x − μq 2
= q(x) log + − dx
σq 2 σp 2 σq




 
σp 1 x − μp 2 x − μq 2
= Eq log + −
σq 2 σp σq
620 Machine Learning with Python: Theory and Applications

 
σp
1 2 1 2
= log + 2 Eq (x − μp ) − Eq (x − μq )
σq
2σp 2σq2

 
σp 1 2 1
= log + 2 Eq (x − μp ) −
σq 2σp 2


σp σq2 + (μq − μp )2 1
= log + −
σq 2σp2 2


1 σp σq2 (μq − μp )2
= 2 log + 2+ −1 (17.6)
2 σq σp σp2

In the last step we used the following simple manipulation:

(x − μp )2 = (x − μq + μq − μp )2 = (x − μq )2 + 2(x − μq )(μq − μp ) + (μq − μp )2 ,

(17.7)
Its expectation over q(x) thus becomes

Eq [(x − μp )2 ] = Eq [(x − μq )2 ] + Eq [2(x − μq )(μq − μp )] + Eq [(μq − μp )2 ]

= σq2 + 0 + (μq − μp )2 (17.8)

The 2nd term in the right-hand-side of Eq. (17.8) vanishes, because

(x − μq )(μq − μp ) is an odd function, and q(x) is symmetric. It is easy to
check if q(x) and p(x) are the same distribution, the KL-divergence becomes
zero, as expected.
Now, if p(x) is standard normal distribution: μp = 0 and σp = 1, we shall
have,

σq2 + μ2q 1
DKL (qp) = − log σq − , (17.9)
2 2

Further, when q(x) is also standard normal distribution, the KL-divergence

becomes zero.

17.6.1.3 KL-divergence for two multi-variable normal distributions

Consider now two normal distributions of p(x) and q(x) of a multi-variable

x ∈ Rk . The dimension of the domain becomes k, and x = [x1 , x2 , . . . , xk ]
in Rk . The probability density functions are given by
 Unsupervised Learning Techniques 621

Figure 17.18: An example of multi-variable normal distribution with k = 2. The

probability distribution of the samples along the red and blue straight lines are,
respectively, given in red and blue Gaussian bell curves. [Image modified based on that of
IkamusumeFan from Wikimedia Commons, under the CC BY-SA 3.0 license].


1 1  −1
p(x) = N (x; μp , Σp ) =  exp − (x − μp ) Σp (x − μp )
k 2
(2π) |Σp |


1 1  −1
q(x) = N (x; μq , Σq ) =  exp − (x − μq ) Σq (x − μq )
2
(2π)k |Σq |
(17.10)

where μ = [μ1 , μ2 , . . . , μk ] is a vector of means in Xk , and Σ is the

covariant matrix in Xk×k . Both p(x) p(x) are still scalar functions, and their
distributions are controlled by their means and covariances. An example of
multi-variable normal distribution with k = 2 is shown in Fig. 17.18. The
KL-divergence from q(x) to p(x) can be written as the following starting
with a multidimensional integral.
 
DKL (qp) = q(x) log(q(x)) − log(p(x)) dx
 
1 det Σp
= q(x) log + (x − μp ) Σ−1
p (x − μp )
2 det Σq

 −1
− (x − μq ) Σq (x − μq ) dx
622 Machine Learning with Python: Theory and Applications


1 det Σp
= Eq log + (x − μp ) Σ−1
p (x − μp )
2 det Σq

 −1
− (x − μq ) Σq (x − μq )

1 det Σp
= log + Eq [(x − μp ) Σ−1
p (x − μp )]
2 det Σq


− Eq [(x − μq ) Σ−1
q (x − μq )]


1 det Σp
= log + Eq [(x − μp ) Σ−1
p (x − μp )] − k
2 det Σq


1 det Σp −1  −1
= log + tr(Σp Σq ) + (μq − μp ) Σp (μq − μp ) − k
2 det Σq
(17.11)

In the 4th step, we used properties of trace and expectation manipulations

on products of matrices:
  
Eq (x − μq ) Σ−1
q (x − μq ) = E q tr (x − μ q ) −1
Σ q (x − μq )
 
= tr Eq (x − μq ) Σ−1 q (x − μq )
  (17.12)
 −1
= tr Eq (x − μq )(x − μq ) Σq
 
= tr Σq Σ−1
q = tr[I] = k

In the last step, we used the matrix version of manipulations similar to

Eq. (17.7).
It is easy to check if q(x) and p(x) are the same distribution, the KL-
divergence becomes zero, as expected.
Now, if p(x) is standard normal distribution, we shall have

1 
DKL (qp) = − log(det Σq ) + tr(Σq ) + μ
q μq − k (17.13)
2

Further, when q(x) is also standard normal distribution, the KL-divergence

becomes zero.
In VAE, we often assume the covariance matrix Σq is diagonal: Σq =
diag(Σq1 , Σq2 , . . . , Σqk ). Therefore, the trace of it will be sum of these
 Unsupervised Learning Techniques 623

variances. The determinant of it will be product of these variances. Therefore,

Eq. (17.13) can be further simplified to
 k k k k

1    
DKL (qp) = − log( Σqi ) + Σqi + μ2qi − 1
2
i i i i
 k k k k

1    
= − log(Σqi ) + Σqi + μ2qi − 1 (17.14)
2
i i i i
k

1  
= − log(Σqi ) + Σqi + μ2qi − 1
2
i

Finally, for numerical stability reason, we often replace Σqi with exp(Σqi ),
and the final equation for KL-divergence often used in VAE in Eq. (17.4) is
given as follows.

k
1  
DKL (qp) = exp(Σqi ) + μ2qi − log(Σqi ) − 1 , (17.15)
2
i

where the means and variances are computed using the current learning
parameters W for q, and W̃ for p for the deterministic part. The standard
normal distribution is used for the stochastic part as shown in Eq. (17.3).
This allows the back-propagation during training by-passing through the
bottleneck layer without the stochastic part effecting the computation of
gradients with respect to W in a VAE model. This is known as the so-called
re-parameterization trick.
With the above formulation, a VAE can now be built with minimum
change to the standard autoencoder. An example code is available at
https://keras.io/examples/generative/vae/ where a step by step how to
create, train and make use of a VAE model.

References

[1] G. Hamerly and C. Elkan, Alternatives to the k-means algorithm that find better
clusterings, 2002. https://doi.org/10.1145/584792.584890.
[2] J.A. Hartigan and M.A. Wong, Algorithm AS 136: A K-Means Clustering Algorithm,
Journal of the Royal Statistical Society. Series C (Applied Statistics), 28(1), 100–108,
1979. http://www.jstor.org/stable/2346830.
[3] P. Fabian, V. Gae, G. Alexandre et al., Scikit-learn: Machine learning in Python,
Journal of Machine Learning Research, 12(85), 2825–2830, 2011. http://jmlr.org/
papers/v12/pedregosa11a.html.
624 Machine Learning with Python: Theory and Applications

[4] D. Comaniciu and P. Meer, Mean shift: A robust approach toward feature space
analysis, IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5),
603–619, 2002.
[5] A. Kramer Mark, Nonlinear principal component analysis using autoassociative
neural networks, AIChE Journal, 37(2), 233–243, 1991. https://aiche.onlinelibrary.
com/doi/abs/10.1002/aic.690370209.
[6] I. Goodfellow, Y. Bengio and A. Courville, Deep Learning, 2016.
[7] P. Diederik and W. Max, An introduction to variational autoencoders, CoRR,
abs/1906.02691, 2019. http://arxiv.org/abs/1906.02691.
[8] G.E. Hinton, A. Krizhevsky and S.D. Wang, Transforming auto-encoders, Artificial
Neural Networks and Machine Learning — ICANN 2011, 2011.
[9] D. Shuyong, H. Zhiping, G.R. Liu et al., A novel inverse procedure via creating
tubenet with constraint autoencoder for feature-space dimension-reduction, Interna-
tional Journal of Applied Mechanics, 13(8), 2150091, 2021.
[10] C. Ranjan, Understanding Deep Learning: Application in Rare Event Prediction,
Connaissance Publishing, 2020. www.understandingdeeplearning.com.
[11] G.R. Liu, S.Y. Duan, Z.M. Zhang et al., Tubenet: A special trumpetnet for explicit
solutions to inverse problems, International Journal of Computational Methods,
18(01), 2050030, 2021. https://doi.org/10.1142/S0219876220500309.
[12] L. Shi, F. Wang, G. Liu et al., Two-way TubeNets Uncertain Inverse methods for
improving Positioning Accuracy of robots Based on Interval, The 11th International
Conference on Computational Methods (ICCM2020), 2020.
[13] D. Shuyong, S. Lutong, G.R. Liu et al., An uncertainty inverse method for parameters
identification of robot arms based on two-way neural network, Inverse Problems in
Science & Engineering, Revised, 2021.
 Chapter 18

Reinforcement Learning (RL)

18.1 Basic Underlying Concept

18.1.1 Problem statement

We shall have a pretty good idea about the supervised and unsupervised
machine learning techniques. This chapter discusses another important
technique called reinforcement learning (RL).
When a child is around the age of 1, he/she starts to learn walking. If one
observes the learning process carefully, the child is trying to adjust his/her
steps forward by moving different parts of muscles, without failing down
or avoiding being punished by gravity that is an environment he/she dost
not know. We may also notice that if a parent is squatted down in front of
him/her, he/she will take bolder steps, and the learning often progresses
faster. This is because he/she sees a higher safety and a reward, which
reinforces the learning process. In a way, machine RL is a mimic of this
process of learning to walk, and the machine-child is called agent. An
agent learns to take actions in an environment with unknown distributed
punishments or awards, and the learning is reinforced through the process
of cumulating rewards and/or avoiding punishments.
The RL differs from the supervised learning mainly in the learning guiding
principle. Supervised learning needs a dataset with predefined input-label
pairs, and the learning objective it to minimize the differences between
the given labels and the predictions using the corresponding inputs. In RL,
however, the agent does not have such a dataset. It is given an environment
that is unknown to the agent. It needs to learn the environment by taking
trial actions that can lead to some awards when proper action is taken at
states in the environment.

625
626 Machine Learning with Python: Theory and Applications

A number of techniques related to RL have been developed in various dis-

ciplines including the game theory, operational research, control theory, and
optimal control. Techniques include the Monte Carlo method, Q-learning,
Deep Q-Network, Deep Deterministic Policy Gradient, and Trust Region
Policy Optimization. [1–4]. There are both model-based and model-free RLs.
If a simulation model for predicting the dynamics of the environment is
available, model-based RL may be a good choice, because the agent can
predict the next state with a probability from the current state when an
action is taken. If such a predictive model is not available, one has to use
model-free RL. In this chapter, we will brief first on the fundamental principle
and theory for reinforcement learning. Our focus will be on the so-called
Q-learning method, which is a basic representative technique and is model-
free. We will also introduce neural network implementation of Q-learning:
Q-Network. In the last section, we will discuss about some currently most
popular policy gradient methods, including the powerful Proximal Policy
Optimization (PPO). Readers may refer to Reinforcement Learning: An
Introduction [1] and the wikipedia page and the reference links therein for
more details on various RL techniques.

18.1.2 Applications in sciences, engineering, and business

Reinforcement learning algorithms can be used to solve problems that arise
in sciences, engineering, and business settings where operation automation
is required.
Service and manufacturing robots: The agent (robot) senses and
interacts with the environment and learns to complete a task in the most
effective manner.
Automatic and optimal control: The agent learns to take the most stable
or optimal positions and movements in an environment, by trial and error.
Delivery drones: The agent (drone) learns to find a safe and most economic
path and navigates to the destination.
Operational research: The agent learns to find the best possible operations
to accomplish tasks related to resource management.
Personalized recommendations: The agent learns to make the best
recommendations of news or advertisements to users, by learning and
predicting the behavior of the users.
 Reinforcement Learning (RL) 627

Games: The agent learns to take the best actions to maximize the score.
Each action will affect the states of the agent in the gaming environment. A
typical example is the widely publicized AlphaGo.

18.1.3 Reinforcement learning approach

Figure 18.1 is a schematic drawing of a reinforcement learning process. An
environment is given with some distributions of rewards and/or punishments.
Such distributions are unknown to the agent. The agent sets to explore over
it. It takes actions and then observes environment. Because the environment
is initially unknown to the agent, it can only take actions at random at
the beginning. By chance, it may make a wrong move and needs to restart.
Or, by chance it makes a good move and gets awarded, and tries again. This
trial-and-error learning circle iterates aiming to achieve optimal accumulated
rewards. In each trial, it needs to interpret the environment and update its
states. An algorithm is needed to ensure that such an updating leads to
the best accumulation of rewards (or effective avoidance of the punishment).
To this end, we need a number theory to govern the construction of such
algorithms.

Figure 18.1: A schematic draw of a Reinforcement Learning (RL) process: an agent

takes actions at a state in an environment with distributed reward (or punishment).
The agent iterates a trial-and-error learning circle, aiming to achieve optimal accumu-
lated rewards. Modified based on the original image by Megajuice under the [CC0]
(https://commons.wikimedia.org/w/index.php?curid=57895741).
628 Machine Learning with Python: Theory and Applications

18.1.4 Actions in discrete time: Solution strategy

We know from the previous chapters that any optimization problem has
an objective function, which is given in terms of the learning parameters.
The learning process is essentially finding the minimum, making use of the
gradient information of the objective function. In a reinforcement learning
process, the objective is to accumulate rewards as much as possible, but the
agent does not know where the award is in the given environment.
Clearly, the agent needs to take actions in steps at different points in time.
It must keep track of how the state of the action-outcomes is evolving over
time. At each time, the current state is the basis for making an action that
updates the state, which shall result in a new state for the next time. Clearly,
the outcomes at each time are partially random and partially depending on
the actions taken. Therefore, the Markov decision process or MDP can be
a proper description for our problem.
At any point in time, we need a rule that tells the agent how to act, given
any possible values of the state. Such a rule, determining the actions as a
function of the states, is called a policy or policy function [5], and the very
fundamental equation known as the Bellman equation.
We naturally want the state to evolve in the best possible way, so
that actions on the evolving states can maximize our objective, which is
accumulating rewards. The best possible value of the accumulated rewards,
written as a function of the state, is called the state-value function. When
it is written as a function of the action (and hence also implicitly of the state),
it is called the action-value function. Our reinforcement learning problem
is now casted as an optimization problem with the action-value function as
the objective function, in discrete time.
The final important strategy is how to find the optimizer (the optimal
action and state) that optimizes the action-value function. For this, we use
the Bellman equation, which gives the relationship between the action-
value function at a time point and that at the next time point. Our
optimization problem in discrete time is stated in a step-by-step recursive
form (known as backward induction). This leads to the so-called value
iterations, known as value-based algorithms.
Alternatively, RL can also be casted as a policy optimization problem
and let the agent to learn an optimal policy. This is because when the
policy is optimized and converged in an MDP process, the value function
is also converged at the optimal. The policy and value function are directly
interconnected in the MDP processes. This leads to policy-based algorithms.
 Reinforcement Learning (RL) 629

For effective implementations, various types of neural networks (NN),

which are familiar to our readers by now, are widely used in both value-
based and policy-based RL algorithms/methods.
The following sections will formulate all these key pieces of strategy:
MDP, policy, action-value function, and the Bellman equation. Examples
and codes of both value-based and policy-based algorithms will be presented
and examined.

18.2 Markov Decision Process

The Markov decision process (MDP) is a discrete-time stochastic decision

process. It is a mathematical model for decision making for problems
where outcomes depend on both random variables and decision variables.
Detailed discussions on MDP can be found in multiple sources, including an
article by Bellman, a book Dynamic Programming and Markov Processes
by Ronald Howard, and wikipedia pages. It is an extension of the Markov
chains.
The MDP considers a general dynamic process. At each time step, the
process is in a state at which the agent takes an action. The process moves
into the next state in responding to the action and some random variables,
while the agent receives a reward. The probability of the process moving
into the next state is determined by the current action taken and also the
random variables at the current time step. The important assumption is that
the dynamic process has the Markov property, meaning that the probability
of moving into the next state from the current state does not conditionally
depend on all the previous states and actions. This assumption offers one
practical means to model our reinforcement learning process, and makes it
possible algorithmically.
Mathematically, an MDP in a RL is defined as follows.

• State space S for the agent is interacting with and observing from.
• Action space A defines possible actions that agent may take.
• Transition probability function P (s → s ; a) defines the probability of tran-
sition from state s to state s under action a.
• Reward function R(s → s ; a) defines the immediate reward after transition
from s to s under action a.
• Discount factor γ ∈ (0, 1) that discounts awards from future steps in the
process.
630 Machine Learning with Python: Theory and Applications

The process is essentially award driven, and convergence is ensured in theory

by the discount factor that bounds the total awards obtainable in the entire
process.
The challenges are:
1. the state of the environment is entirely unknown to the agent initially;
2. the choice of actions (the policy) is basically stochastic (although action
space is deterministic); and
3. the transition from one state to the next is also stochastic. The agent has
to learn the state and the policy in some way (value-optimized, policy-
optimized, or a combination), so that its actions can get the maximized
reward.
Therefore, the learning needs to be done via a series of explorations and
exploitations.

18.3 Policy

Policy, π, defines how the agent makes an action selection a ∈ A in state

s ∈ S. It is defined as the following map.

π : S × A → [0, 1] (18.1)

It is a conditional probability:

π(a, s) = Pr(at = a | st = s) (18.2)

Optimal policy: There can be a large number of possible policies (the com-
binations of variables in the state space and that in the action space). The
agent needs to find an optimal policy that maximizes the expected return in
terms of cumulative reward. In value-based algorithms, the simplest ”opti-
mal” policy for the agent is the greedy policy, which takes the action that
gives the best values among possible choices at each step. However, this at-
single-time best choice may not be the best in the entire MDP process. Thus,
in policy-based algorithms, the goal is set for agent to learn a smoother opti-
mal policy over multiple consecutive states in the MDP process, which can
be done using MLPs (Chapter 13), while the agent explores the environment.

18.4 Value Functions

When the agent takes an action, it shall receive an immediate reward for the
current action. In addition, the agent is going to take more actions in the
coming time steps, for which the agent should also receive rewards. However,
 Reinforcement Learning (RL) 631

the future reward should be discounted when it is taken into consideration

currently. Such a discount not only makes sense realistically but also has
mathematical implications. If these is no discount for the future rewards, the
sum of all these future rewards can be unbounded and hence mathematically
not allowable. Therefore, we need a proper definition for measuring the
return that the agent receives for the current and future rewards.
Return: We now formally define return Rt at state st , considering possible
rewards from the future:
∞

Rt = γ k rt+k+1 (18.3)
k=0

where γ ∈ (0, 1) is the discount rate considering the reward rt+k+1 from the
future, which should be valued less currently. Note also that the return of
the rewards from the far future terminal state (if any) should vanish.
When the policy π is followed by the agent at a state, the state-value
function Vπ (s) can now be defined as the expected return from the future
starting with state st . It estimates the total return from all these future
rewards:
∞ 

Vπ (s) = E [R | st = s, π] = E γ k rt+k+1 | st = s, π (18.4)
k=0

This is the state-value function under the policy π.

The action-value function under policy π, Qπ (s, a), can be defined in
the similar manner:
∞ 

Qπ (s, a) = E [R | st = s, at = a, π] = E γ k rt+k+1 | st = s, at = a, π
k=0
(18.5)

18.5 Bellman Equation

In the MDPs, the action-value function Eq. (18.5) used in reinforcement

learning has a recursive relationship. For the greedy policy, for example, we
have the Bellman equation written as:
 
Q(s, a) = max r(s, a) + γQ(s , a) (18.6)
a∈A

The Bellman equation gives Q(s, a) values at the current time step in
terms of those at the next time step Q = Q(s , a), known as backward
induction. Using Eq. (18.6), we can device the following simple but very
basic algorithm, which is the base for a Q-learning algorithm.
632 Machine Learning with Python: Theory and Applications

• Initialization: At time t = 0, Q(s, a), ∀s ∈ S, a ∈ A is initialized simply

with zeros. The action variable of the agent a∗ is computed following
the optimal policy, which gives a∗ = argmaxa∈A Q(s0 , a), added with a
randomly generated small value. Because Q is null at this initial stage,
a∗ has only the random value. This means that this initial action is a
random one, and the agent gets randomly into a new state s1 . Now, we can
search for the future reward maxa∈A Q(s1 , a), and then update Q values
as Q(s0 , a0 ) using Eq. (18.6). This completes the initialization.
• A typical iteration: Consider at an arbitrary time step t ≥ 1, the
current Q(st , at ) is known. We compute a∗ = argmaxa∈A Q(st , a), added
with a smaller random value. The agent takes action with a∗ , and
gets into the next state st+1 . We can now search for the future reward
maxa∈A Q(st+1 , a), and update the Q values as Q(st , at ) using Eq. (18.6).
This completes one typical iteration, and it moves to the next time step
t+1. We see clearly here the important backward induction mechanism:
the agent at the current state gets into the next state, and then has its
current state updated using Eq. (18.6). If the current state has zero value,
and the next state has a value (or reward), this iteration will give back to
the current state some value! In other words, some of the next-state value
is “induced” back to the current state. Because Q is initialized with all
zeros, the agent has to make sufficient random moves to get into a state
where there is some reward. That reward can then start to be back-induced
gradually to previous states in future iterations. We make a remark that
the essential mechanism of the Q-learning algorithm is that the Q values
are updated backward in state space.
• The iteration continues until the Q values converge. Note that we shall
control the randomly generated values in practice. The random value is
gradually reduced with the increasing number of iteration and vanishes
when the iteration gets to infinity.

Finally, we may have a simple intuitive analysis on the convergence of

the algorithm. First, because γ < 1, the return given in Eq. (18.3) will be
finite for all k, if the reward r is finite for all possible policy π, including the
optimal policy. Therefore, Q value from Eq. (18.6) will always be bounded.
Second, because the optimal policy is used at all time steps, the iteration
using Eq. (18.6) can push the Q values to its upper bound further with the
increase of iterations. Therefore, Q has to converge, because it is pushed to
larger and larger by the algorithm to its upper bound.
The above algorithm based on the Bellman equation will be used in the
code given in the example section.
 Reinforcement Learning (RL) 633

18.6 Q-learning Algorithm

Q-learning algorithm is a basic and useful reinforcement learning process

proposed by Watkins (1989) [2] in his PhD thesis “Learning from Delayed
Rewards”, Cambridge, UK. The convergence proof of this algorithm was
refined in [3]. It is said to be “model-free” because the updating process does
not require a predefined model for the environment. The algorithm is based
on MDP, with an optimal policy, and uses the Bellman equation. The agent
learns iteratively taking improved actions, while updating the environment
value function.
“Q” is the name of the action-value function that is the maximum
expected accumulative rewards used in the algorithm. It is a matrix of shape
(dim S, dim A) and is updated during the iterative process. We shall use the
following example to introduce the Q-algorithm.

18.6.1 Example 1: A robot explores a room with unknown

obstacles with Q-learning algorithm
Consider a robot (as the agent) that wants to get to a goal in a room. There
are, however, obstacles in the room. To make the problem simple, let us use
a uniform regular grid to mesh the floor of the room into 4 × 4 regular cells.
Therefore, the dimension of the state space is 16(= 4 × 4). These obstacles
occupy some of the cells, but the robot does not know which ones. The robot
is capable of walking from any empty cell to the next cell in four directions:
up, down, left, and right. Therefore, the dimension of the action space is 4. It
is the robot’s decision on which direction to go to. It starts at the up-leftmost
corner cell (square-one) and walks to a goal cell where a reward is placed.
The robot will be rewarded if it arrives at the goal cell. It will then go back
to the starting cell and try the same again. The attempt (called episode)
continues. If it hits on an obstacle, this attempt is considered failed, and it
has to go back to the starting cell and try all over again. The robot is allowed
to try as many times as required. The objective is to have the robot learn
by itself that the environment can get the rewards at a higher and higher
probability with an increased number of attempts.
We now design a Q-algorithm so that the robot equipped with it is capable
of finding a valid path of reaching the goal cell. Clearly, the robot needs to
fail and try for multiple times, and the only thing that allows it to have
better trials is getting the rewards. We use a Q as a matrix of variables to
record the accumulative reward received by the robot. This record must be
a discrete function of space in the room for each possible action that the
634 Machine Learning with Python: Theory and Applications

robot can take, and has a value for each cell. It also has to choose one of the
four actions that the robot takes. Therefore, we can build a table of record,
and let us call it Q-table with shape (16, 4). At the beginning, nothing has
happened, and hence the Q-table is initialized with all zeros. At the goal cell,
we give it a value of 1.0 as the reward. All the cells occupied by obstacles
(the robot does not know where it is at the beginning) will send the robot
back to square one.
With this setting, and the formulations that are presented in the previous
sections, the formula for updating Q in the Q-algorithm can be easily
written as
temporal-difference (gradient-like)
 

 optimal value at t+1 

 

Qnew (st , at ) ← Qold (st , at ) + lr · rt + γ · max Q(st+1 , a) −Qold (st , at )

a

 
temporal-target (label-like)
(18.7)

where lr is the learning rate (0 < lr ≤ 1) and it is a predetermined parameter.

The reward rt and the discount rate γ are also predetermined parameters.
Note that Eq. (18.7) is essentially the same as Eq. (18.6). If lr is set to 1.0,
both are the same. Equation (18.7) shows explicitly the learning concept,
which is very similar to the standard loss function used in a supervised
training algorithm. The temporal target is something like the label in the
standard loss function, and the temporal difference acts like the gradient.
When the learning rate is set near 1, Q values change faster. The difference
is that Q-learning algorithm is a dynamic process; when the episode (time
step) advances, the optimal value at t + 1 (that is the discounted maximum
reward from the future state st+1 ) is changing and hence the temporal
target.
With this understanding of Eq. (18.7), we can also construct a neural net-
work and different types of deepnets such as CNN to perform reinforcement
learning. We will discuss this in detail in the next section on a Q-Network.
Based on the theory discussed earlier in this chapter, the discount rate
must be set smaller than 1, so that the value for the action variable a is
finite even if the episode gets infinite. In a Q-learning algorithm, we shall set
a maximum number for the episodes, at which the iteration terminates.
Note that at each episode, an inner loop is used to control the exploration
over the state space S. The limit (299 for example) for this inner loop should
be set to a number that is larger than the dimension of S (that is 16 for
our example). This is because the robot may go back and forth (wondering)
 Reinforcement Learning (RL) 635

during the learning process, and hence when it arrives at the goal cell (if it
does), the number of the steps it takes can be quite large (larger than 16).
Even if it failed to arrive at the goal cell because of hitting on an obstacle
and ends the episode, the number can also be quite large, because of the
possibility of “wondering”.

18.6.2 OpenAI Gym

We are now ready for coding our Q-learning algorithm. We would now need
a module called Gym under the MIT License Copyright (c) 2016 [OpenAI]
(https://openai.com). It is a tool kit for developing and comparing reinforce-
ment learning algorithms. For our example, we will use OpenAI Gym to set
up the environment, the room with obstacles. We will also use TensorFlow
to build the algorithm and matplotlib to plot the results. First, we import
necessary modules. The following code is from awjuliani/DeepRL-Agent
under the MIT License:

# https://github.com/awjuliani/DeepRL-Agents
# awjuliani/DeepRL-Agents is licensed under the MIT License
# import necessary modules
import numpy as np
import gym
import random
import tensorflow as tf
import matplotlib.pyplot as plt
%matplotlib inline

Next, we load the environment, FrozenLake-v0, available at Gym module.

This environment was set for people walking on a frozen lake with a hole.
The settings are fixed by Gym and cannot be changed, but it fits perfectly
with our problem without the need for a change.

env = gym.make('FrozenLake-v0') # create an environment instance

The environment has a simple mesh with 4 × 4 cells:

• SFFF (S: starting cell, safe).

• FHFH (F: free cell, safe to walk).
• FFFH (H: hard obstacle, failed and back to starting point).
• HFFG (G: goal cell, mission accomplished, get a rewards and back to
starting point).
636 Machine Learning with Python: Theory and Applications

Note that the robot does not know this setting before starting.
We shall now examine the Q-learning algorithm for the robot exploring
a room with obstacles.

18.6.3 Define utility functions

# Compute the moving average curve of a curve

env=gym.make('FrozenLake-v0') # create environment instance

def mv_avg(x, n): # move average over past n points

return [sum(x[i:i+n])/n for i in range(len(x)-n)]

def print_state(Q,ns,na):
#Q = (env.observation_space.n,env.action_space.n)
print('Initial Q-Table Values: (',ns,',',na,')')
print('Norm of Q=',np.linalg.norm(Q,np.inf))
for s in range(ns//na):
for a in range(na):
print (f'{Q[4*a,s]:7.4f}{Q[4*a+1,s]:7.4f}'
f'{Q[4*a+2,s]:7.4f}{Q[4*a+3,s]:7.4f}')

def tweak_reward(reward, done):

if reward == 0:
reward = -0.01
if done:
if reward < 1:
reward = -1
return reward

def schedule(p0,n_episodes,beta): # schedule parameter decay

return p0/(np.arange(1,episodes+2))**beta

18.6.4 A simple Q-learning algorithm

# Set learning parameters for Q-learning

np.set_printoptions(formatter={'float': '{: 0.3f}'.format})

episodes = 2000 # maximum episodes (trials)

jwSteps = 299 # maximum walk steps in an episode
 Reinforcement Learning (RL) 637

lr_0 = 0.90 #0.85 # initial learning rate (lr)

beta_lr = 0.01 # lr decay power

gamma_0 = .98 # initial discount rate (back reduction)

beta_g = 0.001 # gamma decay power

eps_0 = 0.01 # initial epsilon (random action)

beta_eps = 0.80 # epsilon decay power

lrv = schedule(lr_0,episodes,beta_lr)
gav = schedule(gamma_0,episodes,beta_g)
epsv = schedule(eps_0,episodes,beta_eps)

print(f'lrv_last={lrv[episodes]:.4e}', f'gav_last=
{gav[episodes]:.4e}',f'epsv_last={epsv[episodes]:.4e}')
# initialize lists for recording the total rewards and steps
jList = [] # walked steps in episodes

rList = [] # total rewards received in episodes

norm_Q = [] # maximum (inf norm) value in Q-table
# Initialize the Q-table with all zeros
Q = np.zeros([env.observation_space.n,env.action_space.n])
# Q: array of shape (env.observation_space.n,env.action_space.)
# both env.observation_space.n and env.action_space are
# attributes from Gym once the evn instance is created.
# For FrozenLake-v0, the values are in (16,4),cells
# For our example: observation space is state space

np.random.seed(1)

for e in range(episodes): # episodes+1: maximum episodes

# reset the environment, which gives an initial state
s = env.reset()
#print('e=',e,' s=',s)
rAll = 0
# to record the total rewards received in an episode
done = False # is the attempt over?
j = 0
lrr = lrv[e]
gam = gav[e]
eps = epsv[e]
#The Q-learning algorithm
while j < jwSteps:
# maximum walk steps over the state space in an episode
#env.render('human') # visualizing the training
638 Machine Learning with Python: Theory and Applications

j+=1
# compute the action values using the greedy policy
# using the current Q-table

a=np.argmax(Q[s,:]+np.random.randn(1,env.
action_space.n)*eps)
# The action values are added with random values
# The random value is a must at the initial episode
# because the robot has no choice but walk randomly
# It should be reduced with increase of episodes.
# It shall vanish when the episode gets infinite.

# Take a step with a, move to a new state s1 and get a

# reward r from the environment, and check where the
# attempt is over (either hits a obstacle or arrive
# at the goal cell)
s1,r,done,_ = env.step(a)
# Tweaking the reward to help the agent learn faster
# r = tweak_reward(r0, done) # one may try this

if done and e%500 == 0: #%100 # monitoring the progress

if r == 1:
print("episode:{}/{},score:{:.2f},goal!".
format(e,episodes,r))
#print_state(Q,env.observation_space.n,
env.action_space.n)
else:
print("episode:{}/{},score:{:.2f}".
format(e, episodes, rAll))
# Update Q-Table using the Bellman equation with a
# learning rate lr
Q[s,a] = Q[s,a] + lrr*(r + gam*np.max(Q[s1,:]) - Q[s,a])

#if np.max(Q[s1,:]) != 0 and e%500 == 0:

# print('e=',e,'j=',j,' s1=',s1,' s=',s,' a=',a)
# print('Q:',Q,'\n Q[s1,:]:',Q[s1,:],' \n qm:',
# np.max(Q[s1,:]))

rAll += r
s = s1 # update the state
if done: break

norm_Q.append(np.linalg.norm(Q,np.inf)) # max value in

# Q-table
 Reinforcement Learning (RL) 639

jList.append(j) # record walked steps in episodes

rList.append(rAll) # record total rewards received
# in episodes
print("Score over time: " + str(sum(rList)/episodes))
#print_state(Q,env.observation_space.n,env.action_space.n)
N_avg = 100
plt.plot(rList,label="Current Reward")
plt.plot(mv_avg(rList, N_avg),label="Average reward over
"+str(N_avg))
plt.title('Reward received')
plt.legend(loc='lower right')
plt.show()

plt.plot(norm_Q, label="Max Q value")

plt.plot(mv_avg(norm_Q, N_avg),label="Averaged max Q over"
+str(N_avg))
plt.title('Maximum Q value')
plt.legend()
plt.show()

lrv last=8.3412e-01 gav last=9.7258e-01 epsv last=2.2856e-05

episode:0/2000,score:0.00
episode:500/2000,score:0.00
episode:1000/2000,score:1.00, goal!
episode:1500/2000,score:0.00
Score over time: 0.4445

Figure 18.2: Reward received by the agent in the FrozenLake-v0 environment through
Q-learning.
640 Machine Learning with Python: Theory and Applications

Figure 18.3: Q value evolution in the FrozenLake-v0 environment during Q-learning.

Figure 18.2 shows the record on rewards received over each episode of
attempt by the robot starting with a random action (randomly generated
Q-table). It is shown that the robot is wondering at the initial few hundred
episodes. It seems to learn its way to reach the goal cell with a good success
rate.
Figure 18.3 shows the evolution of the maximum value in the Q-table
over each episode. Such values become quite close to 1.0 after a few hundred
episodes.

18.6.5 Hyper-parameters and convergence

Note that the Q-learning algorithm above has some level of instability,
meaning that its performance depends on the settings of the hyper-
parameters (learning rate, discount rate, etc.) and the given environment.
This may be partially because of the random process used in the algorithm.
A more sophisticated choice of these parameters or designing a proper time
schedule for these parameters can help to achieve more stable convergence.
This is at the expense of slower convergence to tread for more stable
convergence. For the learning rate, for example, one can let it decay slowly
with episodes. The general rule is given in Eq. (9.41) in Chapter 9. For
Q-learning algorithms, this rule may not always work well, because the ”loss
function” here can be quite far from being convex.
A similar tuning is also possible for the discount rate [3] and the hyper-
parameter controlling the random actions. The basic underlying principle is
that when the Q-algorithm is converged ideally, all these learning, backward
induction, and random actions are no longer needed. All these tuning can
be difficult.
 Reinforcement Learning (RL) 641

18.7 Q-Network Learning

With a good understanding of the Q-learning algorithm, and the knowledge

established in earlier chapters on neural networks, we are now ready to
introduce a neural network for Q-learning, called Q-Network. The major
difference between the Q-Network and the networks we studied earlier is the
loss function, as shown in Eq. (18.7).
We show how a Q-Network can be established and trained for the same
Example 1, but now use a Q-Network for the robot. We will also use OpenAI
Gym to set up the room environment with obstacles. We will use the Tensor-
Flow graph to build the Q-Network structure and execute the training with
TensorFlow sessions. The following code is from awjuliani/DeepRL-Agent
under the MIT License. It has been modified with more detailed explanations
for easy comprehension for beginners. Most of the explanation is given
together with the code lines as comments so that it can be followed with ease.

18.7.1 Example 2: A robot explores a room with unknown

obstacles with Q-Network

# Import necessary modules

from __future__ import division
import gym
import numpy as np
import random
#import tensorflow as tf
import tensorflow.compat.v1 as tf
# This code works only for tf version 1
tf.disable_v2_behavior()
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, InputLayer
import matplotlib.pyplot as plt
%matplotlib inline

# Create the environment instant using Gym module

env = gym.make('FrozenLake-v0')

# a utility function
def mv_avg(x, n): # move average over past n points
return [sum(x[i:i+n])/n for i in range(len(x)-n)]
642 Machine Learning with Python: Theory and Applications

18.7.2 Building TensorFlow graph

tf.reset_default_graph() # define the graph for Q-Network

#tf.compat.v1.reset_default_graph()

The following procedure is for setting up the graph for our Q-Network.
It should have an array of 16 neurons at the input layer, a 16 × 4 weight
matrix, and an array of 4 neurons at the output layer. The structure has a
simple configuration of 16 × 4 × 1.

# Graph for our Q-Network

# First, a placeholder for a tensor to-be-fed. This is for the
# input layer (with 16 inputs) for our Q-Network
inputs1 = tf.placeholder(shape=[1,16],dtype=tf.float32)

# variables in the graph for the Q-Network. Initiated with random

# values and to-be updated during the learning process, when the
# session is run
W = tf.Variable(tf.random_uniform([16,4],0,0.01))

# A matrix-multiply function to produce the output for the

# Q-Network
Qout = tf.matmul(inputs1,W)

# An instance for a prediction following the greedy optimal policy

predict = tf.argmax(Qout,1)

# A placeholder for a tensor to-be-fed. This is for the

# output layer (with 4 outputs) for our Q-Network
nextQ = tf.placeholder(shape=[1,4],dtype=tf.float32)

# A loss function using the tensorflow mean-squared-difference

# between the temporal-target Q and the predicted Q values from
# the Q-Network.
loss = tf.reduce_sum(tf.square(nextQ - Qout))

# A trainer instance using the gradient descent optimizer with lr

trainer = tf.train.GradientDescentOptimizer(learning_rate=0.1)

# A training instance for our Q-Network

updateModel = trainer.minimize(loss)

With the TensorFlow graph defined with necessary arrays (tensors),

variables, instances, and functions for our Q-Network, we are now ready
 Reinforcement Learning (RL) 643

to create and run a session following the graph for training the Q-Network,
based on the Q-learning algorithm.

# Set learning parameters

gamma = .99 # discount rate
epsi = 0.1 # epsilon to control random actions
episodes = 2000 # maximum episodes (trials)
jwSteps = 99 # maximum walk steps in an episode

# initialize lists for recording the total rewards and steps

jList = [] # walked steps in episodes
rList = [] # total rewards received in episodes
jRate = []

init=tf.global_variables_initializer()
# instance for initializing variables one more piece for the graph

with tf.Session() as sess: # Create a session

sess.run(init) # variable initialization for the session

for i in range(episodes):
# Reset the room environment, which gives an initial
# state for robot agent
s = env.reset()

rAll = 0 # initialization variable for the episode

done = False
j = 0

# The algorithm for the Q-Network

while j < jwSteps:
j+=1
# Run the "predict" and 'Qout' defined in the
# graph for the Q-network
a,allQ = sess.run([predict,Qout],
feed_dict={inputs1:np.identity(16)[s:s+1]})
# feed values to placeholder inputs1.
# It is a zero vector but with 1 at entry s.

if np.random.rand(1) < epsi: # sample a small

# random action
a[0] = env.action_space.sample()

# Take the action with a, gets into a new state,

# receive a reward from the room environment,
# and check whether or not this attempt
644 Machine Learning with Python: Theory and Applications

# (episode) is over (done)

s1,r,done,_ = env.step(a[0])

# Compute the Q' values in the new state through

# our network
Q1 = sess.run(Qout,feed_dict={inputs1:np.
identity(16)[s1:s1+1]})

# Obtain the greedy maxQ' and set our target value

# for chosen action.
maxQ1 = np.max(Q1) # Obtain the greedy maxQ'
targetQ = allQ
targetQ[0,a[0]] = r + gamma*maxQ1 # target value
#Train the Q-Network using target and predicted
#Q values
_,W1 = sess.run([updateModel,W],
feed_dict={inputs1:np.identity(16)[s:s+1],
nextQ:targetQ})
rAll += r
s = s1
if done:
#Reduce chance of random action as we train
#the model.
epsi = 1./((i/50) + 10)
break

jList.append(j)
rList.append(rAll)
jRate.append(rAll/j)

print("Score over time: " + str(sum(rList)/episodes))

Score over time: 0.433

18.7.3 Results from the Q-Network

We now plot the results in a similar manner as in Example 1.

N_avg = 100
plt.plot(rList,label="Current Reward")
plt.plot(mv_avg(rList, N_avg),label="Average reward over"
+str(N_avg))
plt.title('Reward received')
plt.legend(loc='lower right')
plt.show()
 Reinforcement Learning (RL) 645

plt.plot(norm_Q, label="Max Q value")

plt.plot(mv_avg(norm_Q, N_avg),label="Averaged max Q over"
+str(N_avg))
plt.title('Maximum Q value')
plt.legend()
plt.show()

Figure 18.4: Award received (top) by the agent and Q value evolution (bottom) in the
FrozenLake-v0 environment during Q-learning using an NN.

The results and learning process are shown in Fig. 18.4. Quite similar
observations can be made as Example 1.
With the understanding of how to build a Q-Network, one should not
have much difficulty in learning to construct deepnets for Q-learning. Our
study has found that when a greedy policy is used at each step, such an at-
single-time best choice may not be the best in the entire process. The agent
may be trapped in local optimal and fail to arrive to the global optimal,
as shown in the above examples. Policy-based algorithms can overcome this
problem.
646 Machine Learning with Python: Theory and Applications

18.8 Policy gradient methods

The Q-learning and DQN studied in the previous sections is known as value-
based aiming to maximize the Q values. Instabilities can often be observed
in this type of algorithms. Alternatively, one can optimize the policy instead,
known as policy-based methods, such as the policy gradient methods. These
methods have been studied quite intensively in the past few years, and
some significant improvements are made. These methods aim to optimize the
policy, instead of the value. Because the transition probability P (s → s ; a)
determines the return R, and is under action a that is in turn controlled
(conditioned) by the policy π. Therefore, if we can find an optimal policy
that maximize the return, our goal is achieved. There are many policy
gradient methods and techniques developed in the past years, and study
all these can be a challenge. In this section, we introduce first one of the
best performers, Proximal Policy Optimization (PPO) [6], and then derive
a formulation procedure that outlines the major path leading to PPO.

18.8.1 PPO with NN policy

The Advantage Actor Critic (A2C), Asynchronous Advantage Actor Critic

(A3C) [7], Trust Region Policy Optimization (TRPO) [7] and the Proximal
Policy Optimization (PPO) [6] are typical type of policy gradient methods.
The PPO is an improvement on TRPO. Readers may take a look at the
excellent on-line lecture series by Pieter Abbeel on this topic.
It is found that policy gradient methods are, in general, also quite
sensitive to hyper-parameters, such as the step-size, which limits the
performance and application, in addition to its sample efficiency issues. We
know that supervised learning has quite good convergence and uses relatively
less hyper-parameters. In addition, reinforcement learning is usually much
less stable because of the absence of firm labels. Careful sampling strategies
are thus required to smooth the searching process. These strategies can be
complicated, expensive and with additional hyperparameters. PPO aims to
find a balance among simple formulation, ease of implementation and sample
efficiency. It computes an update at each step while ensuring the updated
policy is not too far from the old policy to reduce variance and increase
stability in training. This is done via “clipping”. The loss function is designed
as follows.

LCLIP (ŵ) = Et [min (rt (ŵ)At , clip (rt (ŵ), 1 − ε, 1 + ε) At )] (18.8)

 Reinforcement Learning (RL) 647

where
• ŵ is the vector of all parameters used for the policy model.
• Et denotes an empirical expectation over time-steps.
• rt is the ratio of the probabilities under the current and old policies.
• At is the estimated advantage (the expected return of a state with its
baseline subtracted).
• ε is a hyperparameter, usually set as 0.1 or 0.2.
The clip function is defined as
⎧
⎨ 1 − ε if x < 1 − ε
clip(x, 1 − ε, 1 + ε) = 1 + ε if x > 1 + ε (18.9)
⎩
x else

18.8.2 Strategy used in policy gradient methods and PPO

18.8.2.1 Build an NN model for policy
In a typical policy gradient method, such as PPO, we first build a neural
network model (such as MLP, CNN, etc.) that maps dim S → dim A. This
is done by letting the agent to explore the environment first by itself for,
say m time-steps. All the data (states, actions, values, rewards, and dones)
generated during these m steps are stored. The stored dataset is now used
to train one (or two) neural networks for the dim S → dim A mapping. The
training of such NNs is essentially the same as we discussed in previous
chapters for classifications and regressions. Care is needed in defining the
loss functions for various policy gradient methods. The trained NNs are
next used by the agent for further exploration for the next m time-steps,
and a new dataset will be generated which is used to train a new policy of
dim S → dim A mapping. This process continues, we shall have a new and
old policies at every m time-steps. Readers may refer to the codes made
public by philtabor for more details. We denote such an NN model as πŵ
where ŵ stands for the training parameters. The following key formulations
are useful in understanding the policy gradient methods. In particular, both
TRPO and PPO use two (new and old) policies to for a so-called ratio policy
for better controlling stability.

18.8.2.2 P and R formulation

Let τ denotes a state-action sequence, s0 , a0 , s1 , a1 , . . . , sT , aT . Starting from
Eq. (18.4), the problem can be casted as
648 Machine Learning with Python: Theory and Applications

⎡ ⎤
⎢ ⎥
⎢ T ⎥ 
⎢
max E ⎢ R(st , at ) | πŵ ⎥ ⇒ max E [R(τ ) | π ] ⇒ max P (τ ; ŵ)R(τ )
ŵ ⎥ ŵ
ŵ
ŵ
⎣ t=0 ⎦ τ

 
 
R(τ ) U (ŵ)
(18.10)
To solve the above problem, we need to evaluate the gradient of U (ŵ) with
respect to ŵ.
 
∇ŵ U (ŵ) = ∇ŵ P (τ ; ŵ)R(τ ) = ∇ŵ P (τ ; ŵ)R(τ )
τ τ
 ∇ŵ P (τ ; ŵ)
= P (τ ; ŵ) R(τ ) (18.11)
τ
P (τ ; ŵ)

= P (τ ; ŵ) [∇ŵ log P (τ ; ŵ)] R(τ )
τ

Using m sample paths and policy πŵ , the above expectation can be
approximated as a summation, and we obtain,

1    
m
∇ŵ U (ŵ) ≈ ∇ŵ log P (τ (i) ; ŵ) R(τ (i) ) (18.12)
m
i=1

Note here that the gradient is applied only to the P . Thus, R needs not be
smooth and can be discrete.
P →π
We now ready to make a connection between the transition probability
function P to policy π, using the states and actions.
⎡ ⎤
⎢  (i)
T
(i) (i)
 
(i)

(i) ⎥
∇ŵ log P (τ (i) ; ŵ) = ∇ŵ log ⎢
⎣ P s t+1 | s t , at · π ŵ at | s t
⎥
⎦
t=0
 
 
transition probability policy
 T 
   T
  
(i) (i) (i) (i) (i)
= ∇ŵ log P st+1 | st , at + log πŵ at | st
t=0 t=0
T
  
(i) (i)
= ∇ŵ log πŵ at | st (18.13)
t=0
 
policy only

R→A
 Reinforcement Learning (RL) 649

Equation (18.12) can now written as

⎛ ⎡ ⎤⎞
 T 
m
1 ⎜   
(i) (i) ⎢ ⎥⎟
∇ŵ U (ŵ) ≈ ⎝ ∇ŵ log πŵ at | st ⎣R(τ (i) ) − bR ⎦⎠ (18.14)
m
 
i=1 t=0
At

where A is called advantage, which is R subtracted by bR that is a baseline

of the return. It can be calculated for example using
m
1 
bR ≈ R(τ (i) ) (18.15)
m
i=1

Because bR is an expectation, this subtraction does not introduce any bias. It

enhances the effect of returns apart from the baseline, and hence widely used
in policy gradient algorithms. The advantage term can be further improved:
T
 −1  
(i) (i)
At = R(τ (i) ) − bR = R st , at − bR
t=0
t−1 
  T
 −1  
(i) (i) (i) (i)
= R sk , ak + R sk , ak − bR
(18.16)
k=0
  k=t
(i)
independent of at , removed
T
 −1  
(i) (i)
= R sk , ak − bR
k=t

The removal of the terms independent of the current action helps to lower
the variance.
Using Eqs. (18.10) and (18.14), our problem can now re-casted as

max Et [πŵ (at | st )At ] (18.17)

ŵ

A number of outstanding algorithms were developed based on Eq. (18.17),

including the Advantage Actor Critic (A2C), Asynchronous Advantage
Actor Critic or A3C [7], and other Actor-Critic Methods.

18.8.3 Ratio policy

Using now the importance sampling, Eq. (18.17) is modified by making use
of the old model from the previous iteration.
maxŵ Et [rt (ŵ)(at | st )At ]
subject to Et (KL[πŵold (· | st ), πŵ (at | st )]) ≤ δ
650 Machine Learning with Python: Theory and Applications

where the KL-divergence (discussed in Chapter 4, used also in Chapter 17

for VAEs) are used to constrain the loss to prevent the current policy getting
too far apart from the old one, and rt is the ratio of policy π with respect
to policy πold that is the NN model with previous trained parameters.

πŵ (at | st )
rt (ŵ) = (18.18)
πŵold (at | st )

Equation (18.18) is used in the TRPO method.

Now, the KL constraint can be implemented using a penalty method,
which leads the early version of PPO. When the constraint is imposed using
the clip function we obtain the PPO with loss defined in Eq. (18.8).
This objective function with clipping implements a Trust Region update
with first order approximation, which is compatible with the widely used
effective Stochastic Gradient Descent (SGD) algorithm. With such an objec-
tive function defined, an NN can be conveniently used in the implementation.
The PPO with clip demonstrated good performances on continuous control
tasks and many other types of tasks. We demonstrate it in two examples:
CartPole and CarRacing.

18.8.4 PPO: Controlling a pole staying upright

Consider a pole attached to a cart via a hinge joint. The cart can move
along a frictionless one-dimensional track. Our goal is to have the pole stay
vertically up by controlling the cart moving +1 or −1. When an RL agent
learns to achieve this goal, it is placed in an environment that gives a reward
of +1 at every time-step when the pole stays upright. The episode ends when
the pole inclines more than 15 degrees from the vertical position, or the cart
moves more than 2.4 units away from the center. More detailed on this
environment can be found at the OpenAI Gym.
This time we use the stable-baselines3 module to train the RL agent.
Some codes have also made public by philtabor, and Nicholas Renotte
at github to perform these tasks. Examples introduced below uses stable-
baselines3 (which require an install of stable-baselines3).

#!pip install stable-baselines3[extra] #an useful model for RL

import os
import gym
from stable_baselines3 import PPO
 Reinforcement Learning (RL) 651

from stable_baselines3.common.evaluation import evaluate_policy

from stable_baselines3.common.vec_env import DummyVecEnv
import matplotlib.pyplot as plt
%matplotlib inline

#Load and then create an environment

Env_str = "CartPole-v0"
Environment2Solve = Env_str #One of the gym environments
#https://gym.openai.com/envs/#classic_control
Env = gym.make(Environment2Solve)

Figure 18.5: A random agent on the CartPole-v0 environment.

Figure 18.5 shows an agent at a random state on the CartPole-v0 envi-

ronment. The following code produces an animation of samples from the
CartPole environment. It simply repeats falling and being rest.
# Sample some episodes of the environment before training
episodes = 2 # one may try more
print(f"Take look {episodes} samples for {Env_str}")
for episode in range(1, episodes+1):
s = Env.reset() # current state
done = False
rAll = 0

while not done:

Env.render()
action = Env.action_space.sample()
n_state, reward, done, info = Env.step(action)
rAll+=reward

print(f"Episode {episode}: Score={rAll}")

#Env.close() # close env gently when needed

652 Machine Learning with Python: Theory and Applications

Take look 2 samples for CartPole-v0

Episode 1: Score=41.0
Episode 2: Score=20.0

"Understanding the Environment

"action-0: push cart to left, action-1: push to the right
print(f"Type of the action_space:{Env.action_space}")
print(f"A sample from the space:{Env.action_space.sample()}")
Type of the action_space:Discrete(2)
A sample from the space:1

np.set_printoptions(precision=3)
#[cart position, cart velocity, pole angle, pole angular velocity]
print(f"Type of the observation_space:{Env.observation_space}")
print(f"A sample from the space:{Env.observation_space.sample()}")
Type of the observation_space:Box(4,)
A sample from the space:[-4.199e+00 -3.231e+38 2.745e-01
-1.295e+38]

Env = DummyVecEnv([lambda: Env])

"model = PPO('MlpPolicy', Env, verbose = 2)"
Using cpu device

model.learn(total_timesteps=50_000) # let the agent to learn

-----------------------------
| time/ | |
| fps | 995 |
| iterations | 1 |
| time_elapsed | 2 |
| total_timesteps | 2048 |
-----------------------------

------------------------------------------
| time/ | |
| fps | 679 |
| iterations | 2 |
| time_elapsed | 6 |
| total_timesteps | 4096 |
 Reinforcement Learning (RL) 653

| train/ | |
| approx_kl | 0.0033051167 |
| clip_fraction | 0.0277 |
| clip_range | 0.2 |
| entropy_loss | -0.305 |
| explained_variance | 0.123 |
| learning_rate | 0.0003 |
| loss | 196 |
| n_updates | 1240 |
| policy_gradient_loss | -0.00145 |
| value_loss | 169 |
------------------------------------------

......

------------------------------------------
| time/ | |
| fps | 625 |
| iterations | 25 |
| time_elapsed | 81 |
| total_timesteps | 51200 |
| train/ | |
| approx_kl | 0.0022214411 |
| clip_fraction | 0.0128 |
| clip_range | 0.2 |
| entropy_loss | -0.241 |
| explained_variance | 0.391 |
| learning_rate | 0.0003 |
| loss | 83.3 |
| n_updates | 1470 |
| policy_gradient_loss | -0.000732 |
| value_loss | 154 |
------------------------------------------

<stable_baselines3.ppo.ppo.PPO at 0x1aeda592518>

The training is quite fast. It took about three minutes on authors laptop to
train 200k steps.
654 Machine Learning with Python: Theory and Applications

18.8.5 Save and reload the learned model

model_path = os.path.join('Trained_Models','PPO_'+Env_str),
model.save(model_path)

model = PPO.load(model_path, env=Env) # re-load trained model

18.8.6 Evaluate and view the trained model

from stable_baselines3.common.evaluation import evaluate_policy

evaluate_policy(model, Env, n_eval_episodes=2, render=True)

(200.0, 0.0)

Env.close()

To perform a thorough quantitative test on the trained model, we write

the following code. It runs a number of tests of episodes, and in each episode
we put the trained (or untrained) model (agent) in the rest environment for
a maximum number of steps, and receives awards. The total awards received
in each test episode is recorded and plotted in a figure.

test_episodes = 200 # number of tests of episodes

jwSteps = 100 # maximum steps over the state
rList = [] # total rewards received in episodes
for e in range(1, test_episodes+1):
rAll = 0 # to record the total rewards received
done = False
s = Env.reset()
j = 0
while j < jwSteps: # maximum steps over the state
#a=[Env.action_space.sample()] # by untrained samples
a, _ = model.predict(s) # action by trained model
s1, r, done, info = Env.step(a) # s1: the next state
j += 1
if done:
# #print(f"Done! j={j}; info)
break
rAll += r # accumulating rewards
s = s1 # update the state & continue next step
 Reinforcement Learning (RL) 655

rList.append(rAll/jwSteps) # total rewards received

# in an episode

N_avg = 100
x,y=np.array([range(N_avg,len(rList))]).T, mv_avg(rList,N_avg)
plt.plot(rList,label="Current Reward")
plt.plot(x, y, label="Average reward over "+str(N_avg))
plt.title('Reward received, after 10k training')
plt.legend(loc='upper left') #'lower right' )
plt.savefig('images/'+Env_str+'_50k.png', dpi=300)
plt.show()

Figure 18.6 shows the performance of an untrained agent on the CartPole-v0

environment. It acts randomly. The average score is around 0.2.

Figure 18.6: Score of an untrained agent on the CartPole-v0 environment.

After 10k steps of training the performance of the agent is significantly

improved, as shown in Fig. 18.7. The average score is around 0.97. After 50k
steps of training, the agent achieved perfect scores by this test, as shown in
Fig. 18.8.
The following code (or the code given below Section 17.8.5) can generate
animations, which are difficult to put in a book. Readers may run the code,
and should find that a CartPole balances itself quite well after 200k steps of
training.
656 Machine Learning with Python: Theory and Applications

Figure 18.7: An agent trained with 10k steps on the CartPole-v0 environment.

Figure 18.8: An agent trained with 50k steps on the CartPole-v0 environment.

observed = Env.reset()
frames = []
jwSteps = 100
while j < jwSteps: # maximum steps over the state
action, _states = model.predict(observed)
observed, rewards, done, info = Env.step(action)
 Reinforcement Learning (RL) 657

#frames.append(Env.render(mode="rgb_array") #for recording

Env.render() #for viewing the states
if done:
print(f"Done! {info})
break
j += 1

One may also use an alternate algorithm, such as the deep Q-Network (DQN)
to train the model. The code is as simple as follows.

#from stable_baselines3 import DQN

#model = DQN('MlpPolicy',Env,verbose=2)

18.8.7 PPO: Self-driving car

We finally look at an example of a car that learns all by itself to drive on

a track environment. For this example we would need to install SWIG. We
need also to install gym[box2d] and pyglet==1.3.2, in addition to the stable-
baselines3. The codes made public by Nicholas Renotte at github will also
be used.
The environment to be used is CarRacing-v0. Figure 18.9 shows the track
in the CarRacing environment. The state space is a square field of 96 × 96
pixels with an enclosed waving track in it. The car needs to learn to stay
on track. The reward is −0.1 every frame and +1000/N for every track tile
visited by the car, where N is the total number of tiles in the track. For
example, if the car finished in 732 frames, the reward is 1000 − 0.1∗ 732 =
926.8 points. An episode ends when all tiles are visited.

#!pip install gym[box2d] pyglet==1.3.2

import gym
from stable_baselines3 import PPO
from stable_baselines3.common.vec_env import VecFrameStack
from stable_baselines3.common.evaluation import evaluate_policy
import os
import matplotlib.pyplot as plt
%matplotlib inline
658 Machine Learning with Python: Theory and Applications

Figure 18.9: The track in the CarRacing environment.

18.8.8 View samples of the racing car before training

This following code shall produce an animation of samples from the

CarRacing environment with a car running. The car simply moves straight
before training.

# Sample some episodes of the environment before training

episodes = 2 # one may try a few more
print(f"Take look {episodes} samples for {Env_str}")

for episode in range(1, episodes+1):

state = Env.reset()
done = False
rAll = 0
frames = []
while not done:
Env.render() # for in-window real-time render
action = Env.action_space.sample()
n_state, reward, done, info = Env.step(action)
rAll+=reward
print(f"Episode {episode}: Score={rAll}")
#Env.close() #for gentle close the environment

Take look 2 samples for CarRacing-v0

Track generation: 1130..1417 -> 287-tiles track
Episode 1: Score=-30.069930069930493
 Reinforcement Learning (RL) 659

Track generation: 1130..1416 -> 286-tiles track

Episode 2: Score=-33.333333333333826

Env.close() #for gentle close the environment

We can check the environment using the following code.

print(f"action_space:{Env.action_space})
print(f"A sample in action_space:{Env.action_space.sample()})
action_space:Box(3,)
A sample in action_space:[0.5832531 0.9261694 0.27739018]

The action space is a three-dimensional box, allowing the car turning left
and right, and speed.

print(f"observation_space:{Env.observation_space}")
print(f"A part of a sample in observation:\
{Env.observation_space.sample()[0,0:5,:]}")

observation_space:Box(96, 96, 3)
A part of a sample in observation:
[[139 246 145]
[ 18 144 155]
[ 95 226 46]
[215 70 49]
[105 88 1]]

18.8.9 Train the racing car using the CNN policy

log_path = os.path.join('Trained_Models','Logs_'+Env_str)
model_path = os.path.join('Trained_Models','PPO_500k'+Env_str)

model=PPO("CnnPolicy",Env,verbose=2,tensorboard_log=log_path)
Using cpu device
Wrapping the env with a 'Monitor' wrapper
Wrapping the env in a DummyVecEnv.
Wrapping the env in a VecTransposeImage.

model.learn(total_timesteps=500_000)
# try smaller steps first model.save(model_path)
660 Machine Learning with Python: Theory and Applications

Using cpu device

Wrapping the env with a `Monitor` wrapper
Wrapping the env in a DummyVecEnv.
Wrapping the env in a VecTransposeImage.
Track generation: 1041..1324 -> 283-tiles track
retry to generate track (normal if there are not many of
this messages)
Track generation: 1084..1359 -> 275-tiles track
Logging to Trained_Models\Logs_CarRacing-v0\PPO_1
Track generation: 1284..1609 -> 325-tiles track
Track generation: 1129..1418 -> 289-tiles track
retry to generate track (normal if there are not many of
this messages)
Track generation: 1292..1619 -> 327-tiles track
---------------------------------
| rollout/ | |
| ep_len_mean | 1e+03 |
| ep_rew_mean | -54.4 |
| time/ | |

| fps | 39 |
| iterations | 1 |
| time_elapsed | 51 |
| total_timesteps | 2048 |
---------------------------------

18.8.10 Evaluate and view the learned model

The following code shall produce animations of the CarRacing car running
on the track.

evaluate_policy(model,Env,n_eval_episodes=2,render=True)

Track generation: 1214..1522 -> 308-tiles track

Track generation: 1219..1538 -> 319-tiles track
Track generation: 996..1258 -> 262-tiles track

(-93.59801837056875, 0.11267495155334473)

Env.close() #for gentle close the environment

model = PPO.load(model_path, env=Env) # re-load trained model

 Reinforcement Learning (RL) 661

Wrapping the env with a `Monitor` wrapper

Wrapping the env in a DummyVecEnv.
Wrapping the env in a VecTransposeImage.

episodes = 2 #5
for episode in range(1, episodes+1):
s = Env.reset() # initial state
done = False
rAll = 0
frames = []
while not done:
Env.render() # for in-window real-time render
action, _ = model.predict(s.copy())
s1, reward, done, info = Env.step(action)
rAll+=reward
s = s1
print(f"Episode {episode}: Score={rAll}")

Track generation: 1151..1443 -> 292-tiles track

Episode 1: Score=-55.32646048110042
Track generation: 1140..1429 -> 289-tiles track
Episode 2: Score=-54.86111111111186

We have video recorded the trained car using a screen recorder. The video
plays when the following code are executed with the recorded files.

from IPython.display import Video

#Video('./images/PPO_10k_Driving.mp4', width=600, height=600)
Video('./images/PPO_500k_Driving.mp4', width=600, height=600)

<IPython.core.display.Video object>

Our tests found that a not-well-trained car with small training steps (<20k)
runs wildly. When it is trained 500k steps, it runs reasonably well. The
training took a number of hours on the author’s laptop. Interested readers
may run the above codes, or contact the author for these videos. Our tests
found that the PPO is much more stable compared to the Q-algorithm.
662 Machine Learning with Python: Theory and Applications

18.9 Remarks

Finally, we mention that tremendous progress has indeed been made so

far in machine learning, but challenges have always accompanied with
the advancements. Our quest will continue as always in the journey of
developing machine learning techniques for increasing challenging problems,
for the well-being of humanity. As mentioned numerous times in this text,
significant advances require open-minded researchers in all the fields related
to machine learning, including incorporating ideas and techniques from
the well-established and still developing physics-law-based methods. Quests
never end.

References
[1] R. Sutton and A. Barto, Reinforcement Learning: An Introduction, MIT Press,
Cambridge, MA, 2018.
[2] C.J.C.H. Watkins, Learning from Delayed Rewards, PhD thesis, May 1989, Cam-
bridge, UK. http://www.cs.rhul.ac.uk/∼ chrisw/new thesis.pdf.
[3] C.J.C.H. Watkins and P. Dayan, Q-learning, Machine Learning, 8(1), 279–292.
https://doi.org/10.1007/BF00992698.
[4] Duan Shuyong, Zhang Linxin, G.R. Liu et al., A smoothed-shortcut Q-learning
algorithm for optimal robot agent path planning, Journal of Mechanical Engineering,
Accepted, 2021.
[5] R.E. Bellman, Dynamic Programming, Dover, 1957.
[6] Schulman John, “Proximal Policy Optimization Algorithms”, CoRR, vol. abs/1707.
06347, arXiv, 2017. online.
[7] Mnih Volodymyr, “Asynchronous Methods for Deep Reinforcement Learning”, CoRR,
vol. abs/1602.01783, 2016. online.
[8] Schulman John, “Trust Region Policy Optimization”, CoRR, vol. abs/1502.05477,
arXiv, 2015. online.
 Index

A analytical minimization process, 505

abnormality detection, 585 anonymous functions, 86
action-value function, 626, 629 arithmetic operators, 39
actions, 623 attributes of variables, 36
activation functions, 214, 265 augmentation, 7
AdaBoost, 443 augmented dataset, 8
AdaDelta, 378 augmented discrete point, 7
adagrad gradient algorithm, 371 augmented feature space, 7
adam gradient algorithm, 381 autoencoders, 610
adaptive gradient algorithm (AdaGrad), autograd, 307, 327
371 automatic differentiation (AD), 305
adaptive moment estimation (Adam), 381 automorphism, 196
adding axis, 74 average pooling, 544
advantage, 647 awards, 623
affine space, 7, 338, 475 axis, 60
affine transformation, 8, 190, 194, 199,
223, 268, 332, 471 B
affine transformation array (ATA), 190, back-propagation, 323–324
213 backward induction, 630
affine transformation matrices, 9, 195, 472 bad-conditioned, 112
affine transformation unit (ATU), 190, baseline, 647
197, 543 basis functions, 220
affine transformation unit in CNNs, 542 Bayesian statistics, 180
affine transformation weight matrix, 195, Bellman equation, 626, 629
471, 473, 565 better fitted boundary, 503
affinity, 189, 196 binary classification, 223, 254, 421, 438,
agent, 623 448
AlexNet, 548 binary classification problem, 431
algebraic system equations, 109 binary cross-entropy, 177–180, 188
analytic differentiation, 304 binary cross-entropy loss, 429, 432
analytic optimization methods, 332 binary logistic regression, 445

663
664 Machine Learning with Python: Theory and Applications

block structure, 84 convolution operator, 540

boolean operators, 42 convolutional neural network (CNN), 539
boolean values, 41 covariance matrix, 408, 622
boston housing price dataset, 416 cross-entropy, 173–175, 188
bottleneck, 610, 618 cross-entropy loss function, 447, 452
broadcasting, 67 curve fitting, 120, 402
built-in functions, 40 CVXOPT, 249

C D
10-classification, 451 data compression, 133, 211, 585
candidate memory, 565 data encoded parameter, 211
CarRacing, 655 data iterator, 356, 368, 372, 376, 379, 383,
CartPole, 649 409
case, 197 data loss function, 507
chain multiplication, 185 data scaling (normalization), 149
chain of stacked affine transformations, data type, 11
215, 473 data-based model, 3
chained structure, 474 data-parameter converter, 10, 190, 210
chaining ATA, 215 data-point, 5–7, 228, 226
CIFAR images, 552 dataset, 5–7
CIFAR10 dataset, 553 de-noising, 585
class inheritance, 86, 90 decision boundaries for classification
classification, 5 problems, 241, 443, 503, 538
clip function, 645 decision rule, 226
close-form solution, 334, 392 decision tree, 443
closed-form, 619 decoder, 610
CNN policy, 657 deeper (more layers) net, 479
COCO dataset, 552
denoising, 611
code performance, 93
dense layers, 556
column-space of matrix, 476
densely connected, 213
comment line, 21
derivative of the function, 348
comparison operators, 77, 460
diagonal matrix, 129
complete linear bases, 8
different strides and paddings, 542
computation complexity, 258
discount factor, 627
computational graph, 305
discount rate, 629
concatenation, 42, 565
distance function, 239, 258
condition number, 116–117
distance-based, 424
conditional probability, 181
distances, 189
conditions for activation functions, 288
distributed reward, 625
configuration of a CNN, 545–546
doc-strings, 21
confusion matrix, 254
conjugate gradient method, 367, 386 dot-product, 102, 106, 115, 167, 225, 317
controlled random sampling, 158–159 double-star, 54
CONV layers, 546
Conv2D layer, 556 E
convergence process, 347, 589 edge detector, 540–541
convergence theorem, 349 effect of the L2 regularization, 533
convex functions, 343, 348 effects of parallel data-points, 331, 337
ConvNet, 539 eigenvalue decomposition, 113, 136
convolution filter, 540 element-wise computations, 61
 Index 665

element-wise operations, 65, 266, 565 gradient descent in one dimension, 345
encoder, 610 gradient descent momentum (GDM), 363,
encouragement rule, 230 365
enrichment functions, 220 gradient vanishing, 288
entropy of probability, 167 gradient-based algorithms, 354
error exception, 34 gradient-based techniques, 332
expected return, 629 greedy policy, 628
exponential loss, 427
extended affine space, 565, 568 H
extra-forest, 460 handwritten digits recognition, 181
head gradient, 320–321
F hermitian matrix, 116
f -string, 39 Hessian matrix, 248, 364
feature function, 220, 401, 404 hidden layers, 467–468, 478
feature space, 6 high-order polynomial fitting, 401,
features, 5, 7 403–404
feed-forward, 323–324 higher-order curves, 502
feed-forward process, 470 higher-order functions, 218
filter and convolution, 539 higher-order nonlinear latent behavior,
filter matrix, 541 218
fitting to data-points, 502 higher-order polynomial bases, 220
flattened tall vector, 9 highly elliptical shape, 363
forget gate, 563 hinge loss, 425
FrozenLake-v0, 633 Huber loss function, 394
fully connected, 468 hyper-parameters, 638
fully trained model, 501 hyperbolic cosine, 394
functional of prediction functions, 10 hyperbolic tangent function (tanh), 282
functions with conditions, 322 hyperparameters, 495
hypothesis space, 9, 217, 477
G
gated recurrent units (GRUs), 568 I
gauss distribution, 167 image compression, 611
gauss integration, 147 image segmentation, 611
gauss points, 147 import external module, 22
gauss weights, 147 in-place operations, 66
Gaussian basis functions, 404, 406 inception module, 551
Gaussian distribution, 407 inception network, 551
Gaussian kernel, 258 independent connections, 218
Gaussian Naive Bayes, 443 independent features, 218
Gaussian process, 443 independent learning parameters, 478, 565
GD algorithm, 349 indicator function, 424
general guideline, 501 inequality constraints, 249
global minimum, 340 initial data treatment, 148–149
global optimal, 598, 643 initialization of means, 586
global optimum, 249, 588 inner product, 6, 234, 246, 250
GoogLeNet, 551 integrated development environment
gradient boosting regressor, 443, 460, 495 (IDE), 19
gradient descent (GD), 344 interpolation, 120
gradient descent in hyper-dimensions, 347 iris dataset, 260
666 Machine Learning with Python: Theory and Applications

J log-likelihood, 185
joint probability, 180 logistic function, 266
Jupyter Notebook, 19, 24 logistic loss, 426
logistic prediction function, 422
K long short-term memory (LSTM), 480,
563
K-means clustering, 585 loss function codes, 9, 412, 396, 421
kernel density estimation (KDE), 149 loss function comparison, 395
kernel trick, 257 loss functions for classification, 423
KL-divergence, 175–177, 619–620, 648 loss functions for regression, 389
Krylov methods, 110 LSTM layer, 566
LSTM output spaces, 564
L LSTM unit, 564–565
0–1 loss, 424 LSTM-MLP, 577
L1 loss function, 393
L2 loss function, 333, 392 M
L2 norm for the regularization, 506 10-means clustering, 605
L2 regularization, 531 2D-matrix-dot-product, 540
label space, 8 makedown, 19
labeled dataset, 226 margin function, 226, 423, 425
labels (ground truth), 5–6 margin-based, 424
lagrangian multiplier, 243–247 Markov decision process (MDP), 626–627
Lambda functions, 86 mathematical trick, 185
landmark CNNs, 546 matrix inversion, 111
LAPACK, 110 matrix of variables, 319
Lasso, 506 matrix-matrix multiplication, 106
leaky Relu, 284 max pooling, 544
learning direction, 345 MaxPooling2D, 555
learning parameters, 9–10, 196 mean absolute error (MAE) loss function,
learning rate, 345–346, 348, 351, 369 393
least square formulation, 332, 334 mean-shift, 605
least-square solution, 122 mean-squared-error (MSE), 392
left-singular vectors, 129 memory address, 27
LEGB rule, 85 memory size, 35
LeNet-5, 547 min-max scaling, 149
linear algebra, 95 mini-batch algorithm, 354
linear basis functions, 6 mini-batch SGD, 355
linear combination, 245 minimizer of learning parameters, 10
linear function prediction unit, 189, 197, minimum number of data-points, 477
422 MINPACK, 143
linear prediction function, 190, 218 MLPClassifier, 385
linear regression, 368, 390 MNIST dataset, 181, 448, 481, 488, 526,
linear schedule, 352 575
linear SVM, 443 modified function, 247
linearly separable, 224, 230, 234 modified objective function, 242–243
Lipschitz continuity, 351 module in Python, 92
list comprehensions, 76 modules, 91
local derivatives, 305 moment matrix, 8, 335
local minima, 339 momentum parameter, 365
log-cosh loss function, 394 monotonically increasing function, 430
 Index 667

moving vectors, 578 numpy 1D array vs. vector, 74

multi-dimensional array, 62, 67 numpy 2D array vs. matrix, 74
multiclass classification, 260, 445 numpy array, 55–56, 71
multidimensional integral, 621
multidimensional numpy array, 60 O
multilayer perceptron (MLP), 467
2nd-order polynomial bases functions, 220
multinomial regression, 447
multiple channels, 543 3rd-order functions, 218
multiple strides, 541 6th-order polynomial, 403
multipliers method, 242 object detection, 86, 539, 558
must-have nonlinearity, 265 objective function, 242, 249
MXNet, 63 objective tests, 503
MXNet NDArrays, 70–71, 97 one-hot encoding, 149, 452
one-to-one unique transformation, 218
N OpenAI Gym, 633
optimal learning parameters, 219, 475–476
1 → 1 → 1 network, 217
Naive Bayes classification, 181–182 optimal policy, 628
nearest neighbors, 443 optimization techniques, 331
negative direction, 346 outliers, 398
Nesterov accelerated gradient (NAG), 370 output gate, 563
network for k-classifications, 446 output layer, 469
neural network configuration, 390 overfitting, 501
neural network for classification, 432 overly fitted boundary, 503
neural networks (NN), 627 overly simplified boundary, 503
neural-pulse-unit (NPU), 269, 271
neuron basis functions, 275 P
neurons-samples theory, 474–476 p → k mapping, 219, 468
Newton methods, 386 p → 1 neural network, 390–391
NN regulated, 567 p → k network, 213
non-negative definite, 248
p → k neural network, 214
nonlinear activation function, 206, 478
P ∗ -number, 475, 477
nonlinear affine transformations, 219
padding, 542
nonlinear feature functions, 406
paired keys and values, 51
nonlinear kernel functions, 246
parallel data-points, 337
nonlinear kernel models, 263
parallelism, 189
nonlinear PCA, 258
nonlinear regression, 415, 418 parameter encoding, 210
norm, 108 Pascal pyramid, 257
normal distribution (Gaussian PCA formulation, 135
distribution), 165, 618 peephole LSTM, 567
normal equation, 334 perceptron, 223
normal matrix, 130, 334 perceptron algorithm, 230
NP-hard, 586 perceptron convergence proof, 236
null-space, 476 perceptron convergence theorem, 233
number of data-points, 477 physics-law-based model, 1–2, 220
number of layers, 475, 477 point-to-point transformation,
number of neurons, 475 209–210
number of samples, 6, 477 policy function, 626
numerical differentiation, 305, 327 policy gradient algorithms, 647
668 Machine Learning with Python: Theory and Applications

policy gradient methods, 644–645 R

policy-based algorithms, 643 radial basis functions (RBFs), 126
polynomial kernels, 257
random distributions, 164
pooling, 544
random forest, 443, 456, 460
possible norms, 506
random sampling, 158
power activation function, 288
randomly initialized model, 453
power scheduling, 351
rank deficiency, 118
power-linear function, 288
rank of a matrix, 118
power-quadratic function, 288
ratio policy, 647
predictability for constants, 192
rational activation function, 288
predictability for linear functions, 192
ratios of distances, 189
predictability of high-order functions,
RBF SVM, 443
214
recommendation, 585
predictability of the solution, 331
reconstructing the data, 611
predicting complex nonlinear functions,
219 reconstruction, 143
prediction function gradient, 213 recurrent neural networks (RNNs),
563
prediction functions, 189, 193, 421
prediction of linear functions, 193 reduced LSTM, 566
principal component analysis (PCA), 135, regression, 5
585 regularization effects, 501, 526
principal components, 135, 140 regularization loss function, 507
prior distribution, 619 regularization parameter, 505, 533–534
probability and statistics, 157 regularization techniques, 504
proper regularization parameter, 538 regularized model, 533–534
property of a convex function, 348 regularized solution, 506
proximal policy optimization (PPO), 644 reinforcement learning (RL), 623
pseudo-dimension of the affine space, 8, Relu functions, 283
472, 475–476 ResNet, 480, 549–550
pulse location, 271 return, 629
pulse width, 271 reward function, 627
Python code for the perceptron, 226 right-singular vectors, 129
Python environment, 15 RMSProp, 374
Python language, 19 root finding, 143
rotation matrix, 119
Q rule for estimating learning parameters,
478
Q-algorithm, 632
Q-learning, 624
S
Q-learning algorithm, 630–631,
633–634 saddle point, 247–248, 340
Q-Network, 624, 639 sequence feature models, 563
Q-table, 632 sequential dataset, 564
QR decomposition, 130 sigmoid curve shift, 270
QR transformation, 334 sigmoid function (σ(z)), 266, 268
quadratic discriminant analysis (QDA), sigmoid wrapped affine transformation,
443 208
quadratic programming (QP), 249 similarity function, 258
quadratic programming problem, 245 singular value decomposition (SVD), 112,
Quasi-Newton methods, 386 129
 Index 669

skip connection, 480, 549 TensorFlow graph, 464, 539, 640

slack variables, 244 terminal control, 9
slicing, 46 ternary conditionals, 84
small sub-space, 477 testing dataset, 11
softmax activation, 445, 447, 469 Tikhonov regularization, 504
softmax cross-entropy, 483
time sequence, 575
softmax function, 446, 451, 482
softmax regression, 447 total number of learning parameters, 473,
softplus function, 286 506
solution existence theory, 194, 331, 334, total number of neurons, 477
336 trainable parameters, 216, 473
solution predictability, 337 training dataset, 11
space evolution process, 474 training parameters, 475
spaces in machine learning models, 474 transition probability function, 627
sparse model, 507 transpose to 1D array, 74
speech recognition, 584 trapezoid rule, 145
square loss, 427
Trust Region update, 648
squared differences, 333
tunable hyperparameter, 394
squashing function, 266
stable-baselines, 3, 648 typical case of overfitting, 531
stacked affine transformations, 216 typical data structures, 198
standard affine space, 567
standard normal distribution, 166 U
standard scaling, 149
UCI ML hand-written digits, 601
state space, 627
state-value function, 626, 629 UCI repository, 433
states, 623 underscore placeholders, 28, 49
stationary points, 247 uniform distribution, 165
statistical equivalence in expectation, 382 unique encoding, 211–212
step function, 270 unique minimum, 332
stochastic gradient descent (SGD), 353, unitary matrix, 129
412 universal approximation theorem,
strictly positive-definite, 506 274, 479
stride, 542 universal approximation theory, 265
structure of LSTMs, 564
universal functions, 71
sub-gradient, 394
universal prediction theory, 189, 218–219,
subclass, 90
279, 338, 456, 467, 479
sum of the dimensions of all the affine
spaces, 477 unpacking a string, 44
summation function, 317 unsupervised learning techniques, 585
support vector machine (SVM), 223, 237 up sampling, 545
supporting vectors, 245 update gate, 563
SVD algorithm, 130 update of means, 588
SVD decomposition, 130
SVM problem, 248 V

T value functions, 628

tanh prediction function, 423 variable formatting, 38
Taylor series expansion, 347 variable type, 25
tensor, 100 variational autoencoder (VAE), 617
670 Machine Learning with Python: Theory and Applications

vector space, 6 X
versions of Python, 15 xw formulation, 192, 213, 391
VGG-16, 549 xw+b formulation, 192, 214
viewing the weight matrix, 490
Voronoi diagram, 587, 589
Y
W YOLO, 551
YOLOv3 object detection, 559
wide (more neurons in a layer) net,
479
wrapped affine transformation, 209 Z
wrapped with activation functions, zigzag behavior, 364
338 zigzag path, 364