EE2211 Lecture 7

EE2211 Introduction to Machine
Learning
Lecture 7
Thomas Yeo
thomas.yeo@nus.edu.sg
Electrical and Computer Engineering Department

National University of Singapore
Acknowledgement: EE2211 development team

Thomas Yeo, Kar-Ann Toh, Chen Khong Tham, Helen Zhou, Robby Tan & Haizhou
© Copyright EE, NUS. All Rights Reserved.

Course Contents
• Introduction and Preliminaries (Haizhou)
– Introduction
– Data Engineering
– Introduction to Probability and Statistics
• Fundamental Machine Learning Algorithms I (Kar-Ann / Helen)
– Systems of linear equations
– Least squares, Linear regression
– Ridge regression, Polynomial regression
• Fundamental Machine Learning Algorithms II (Thomas)
– Over-fitting, bias/variance trade-off
– Optimization, Gradient descent
– Decision Trees, Random Forest
• Performance and More Algorithms (Haizhou)
– Performance Issues
– K-means Clustering
– Neural Networks
2
Fundamental ML Algorithms:
Overfitting, Bias-Variance Tradeoff
Module III Contents

• Overfitting, underfitting & model complexity
• Regularization
• Bias-variance trade-off
• Loss function
• Optimization
• Gradient descent
• Decision trees
• Random forest
3
Regression Review
• Goal: Given feature(s) , we want to predict target
– can be 1-D or more than 1-D
– is 1-D
• Two types of input data
– Training set , from
– Test set { , from
• Learning/Training
– Training set used to estimate regression coefficients
• Prediction/Testing/Evaluation
– Prediction performed on test set to evaluate performance
4
Regression Review: Linear Case
• is 1D & is 1-D
• Linear relationship between &
• Illustration (4 training samples):
• Training/Learning (primal) on training set
• Prediction/Testing/Evaluation on test set
5
Regression Review: Polynomial
• is 1-D (or more than 1-D) & is 1-D
• Polynomial relationship between &
• Quadratic illustration (4 training samples, is 1-D):
6
Regression Review: Polynomial
• is 1-D (or more than 1-D) & is 1-D
• Polynomial relationship between &
• Quadratic illustration (4 training samples, is 1-D):
7
Note on Training & Test Sets
• Linear is special case of polynomial => use “P”
instead of “X” from now on
• There should be zero overlap between training & test

sets
• Important goal of regression: prediction on new
unseen data, i.e., test set
• Why is test set important for evaluation?
8
Overfitting Example
9
Overfitting Example
Training Test Set

Set Fit Fit
Order 9 Good Bad
Order 1 Bad Bad
Order 2 Good Good
10
Overfitting Example
Big Prediction Error
Training Test Set

Set Fit Fit
Order 9 Good Bad
Very Big Prediction Error
Order 1 Bad Bad
Order 2 Good Good
11
Overfitting Example
Big Prediction Error
• If we take one of the blue

lines and compute the
square of its length, this is
called “squared error” for
that particular data point
Very Big Prediction Error • If we average squared
errors across all the red
crosses, it’s called mean
squared error (MSE) in the
test set
12
Underfitting Example
Training Test Set

Set Fit Fit
Order 9 Good Bad
Order 1 Bad Bad
Order 2 Good Good
13
Training Test Set

Set Fit Fit
Order 9 Good Bad
Order 1 Bad Bad
Order 2 Good Good
14
Training Test Set

Set Fit Fit
Order 9 Good Bad
Order 1 Bad Bad
Order 2 Good Good
15
Training Test Set

Set Fit Fit
Order 9 Good Bad
Order 1 Bad Bad
Order 2 Good Good
16
“Just Nice”
Training Test Set

Set Fit Fit
Order 9 Good Bad
Order 1 Bad Bad
Order 2 Good Good
17
“Just Nice”
Training Test Set

Set Fit Fit
Order 9 Good Bad
Order 1 Bad Bad
Order 2 Good Good
18
Overfitting & Underfitting
Training Test Set

Set Fit Fit
Overfitting Order 9 Good Bad
Underfitting Order 1 Bad Bad
Order 2 Good Good
19
• Overfitting occurs when model predicts the training data
well, but predicts new data (e.g., from test set) poorly
• Reason 1
– Model is too complex for the data
– Previous example: Fit order 9 polynomial to 10 data points
• Reason 2
– Too many features but number of training samples too small
– Even linear model can overfit, e.g., linear model with 9 input
features (i.e., is 10-D) and 10 data points in training set =>
data might not be enough to estimate 10 unknowns well
• Solutions
– Use simpler models (e.g., lower order polynomial)
– Use regularization (see next part of lecture)
20
• Underfitting is the inability of trained model to predict the
targets in the training set
• Reason 1
– Model is too simple for the data
– Previous example: Fit order 1 polynomial to 10 data points
that came from an order 2 polynomial
– Solution: Try more complex model
• Reason 2
– Features are not informative enough
– Solution: Try to develop more informative features
21
Overfitting / Underfitting Schematic
Underfitting Overfitting
regime regime
22
Regularization
• Regularization is an umbrella term that includes methods
forcing learning algorithm to build less complex models.
• Motivation 1: Solve an ill-posed problem
– For example, estimate 10th order polynomial with just 5 datapoints
• Motivation 2: Reduce overfitting
• For example, in previous lecture, we added :
• Minimizing with respect to , primal solution is
• For , matrix becomes invertible (Motivation 1)
23
Regularization
• Regularization is an umbrella term that includes methods
forcing learning algorithm to build less complex models.
• Motivation 1: Solve an ill-posed problem
– For example, estimate 10th order polynomial with just 5 datapoints
• Motivation 2: Reduce overfitting
• For example, in previous lecture, we added :
• Minimizing with respect to , primal solution is
• might also perform better in test set, i.e., reduces

overfitting (Motivation 2) – will show example later
24
Regularization
• Consider minimization from previous slide
cost function quantifying Regularization

data fitting error in training
set
25
Regularization
• Consider minimization from previous slide
L2 - Regularization
•
• Encourage to be small (also called shrinkage or
weight-decay) => constrain model complexity
• More generally, most machine learning algorithms can be
formulated as the following optimization problem
• Data-Loss(w) quantifies fitting error to training set given

parameters w: smaller error => better fit to training data
• Regularization(w) penalizes more complex models
26
Regularization Example
Training Test
Set Fit Set Fit
Order 9 Good Bad
Order 9 + = 1 Good Good
27
Training Test
Set Fit Set Fit
Order 9 Good Bad
Order 9, =1 Good
28
Training Test
Set Fit Set Fit
Order 9 Good Bad
Order 9, =1 Good Good
29
Bias versus Variance
• Suppose we are trying to predict red target below:
Low Variance High Variance
Low Bias: blue Low Bias: blue
predictions on average Low Bias predictions on average
close to red target close to red target
Low Variance: low High Variance: large
variability among variability among
predictions predictions
High Bias: blue High Bias: blue

High Bias
predictions on average predictions on average

not close to red target not close to red target
Low Variance: Low High Variance: high
variability among variability among
predictions predictions
30
Bias + Variance Trade off
• Test error = Bias Squared + Variance + Irreducible Noise
High Bias High Variance

Low Variance Low Bias
31
Bias + Variance Example
• Simulate data from order 2 polynomial (+ noise)
• Randomly sample 10 training samples each time
• Fit with order 2 polynomial: low variance, low bias
• Fit with order 4 polynomial: high variance, low bias
4th Order Polynomials 2nd Order Polynomials
32
33
34
35
36
37
38
• Fit with order 2 polynomial: low variance, low bias Order 2
Achieves Lower
• Fit with order 4 polynomial: high variance, low bias Test Error
39
Bias + Variance Trade off Theorem
40
41
42
Summary
• Overfitting, underfitting & model complexity
– Overfitting: low error in training set, high error in test set
– Underfitting: high error in both training & test sets
– Overly complex models can overfit; Overly simple models can underfit
• Regularization (e.g., L2 regularization)
– Solve “ill-posed” problem (e.g., more unknowns than data points)
– Reduce overfitting
• Bias-Variance Tradeoff
– Test error = Bias Squared + Variance + Irreducible Noise
– Interpretation:
• Overly complex models can have high variance, low bias
• Overly simple models can have low variance, high bias
• Interpretation is not always true (see tutorial)
43

EE2211 Lecture 7

Uploaded by

Copyright:

Available Formats

EE2211 Lecture 7

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

EE2211 Lecture 7

Uploaded by

Copyright:

Available Formats

EE2211 Introduction to Machine

Electrical and Computer Engineering Department

Acknowledgement: EE2211 development team

© Copyright EE, NUS. All Rights Reserved.

Module III Contents

• Training/Learning (primal) on training set

• Prediction/Testing/Evaluation on test set

• Training/Learning (primal) on training set

• Prediction/Testing/Evaluation on test set

• Training/Learning (primal) on training set

• Prediction/Testing/Evaluation on test set

• Prediction/Testing/Evaluation on test set

• There should be zero overlap between training & test

Training Test Set

Training Test Set

• If we take one of the blue

Training Test Set

Training Test Set

Training Test Set

Training Test Set

Training Test Set

Training Test Set

Training Test Set

• Minimizing with respect to , primal solution is

• For , matrix becomes invertible (Motivation 1)

• Minimizing with respect to , primal solution is

• might also perform better in test set, i.e., reduces

cost function quantifying Regularization

• Data-Loss(w) quantifies fitting error to training set given

High Bias: blue High Bias: blue

predictions on average predictions on average

High Bias High Variance

You might also like