Python Machine Learning

The Ultimate Beginner's Guide to

Learn Python Machine Learning Step
by Step using Scikit-Learn and
Tensorflow
Ryan Turner
© Copyright 2019 – Ryan Turner - All rights reserved.
The content contained within this book may not be reproduced, duplicated or
transmitted without direct written permission from the author or the
publisher.

Under no circumstances will any blame or legal responsibility be held against

the publisher, or author, for any damages, reparation, or monetary loss due to
the information contained within this book. Either directly or indirectly.

Legal Notice:
This book is copyright protected. This book is only for personal use. You
cannot amend, distribute, sell, use, quote or paraphrase any part, or the
content within this book, without the consent of the author or publisher.

Disclaimer Notice:
Please note the information contained within this document is for educational
and entertainment purposes only. All effort has been executed to present
accurate, up to date, and reliable, complete information. No warranties of any
kind are declared or implied. Readers acknowledge that the author is not
engaging in the rendering of legal, financial, medical or professional advice.
The content within this book has been derived from various sources. Please
consult a licensed professional before attempting any techniques outlined in
this book.

By reading this document, the reader agrees that under no circumstances is

the author responsible for any losses, direct or indirect, which are incurred as
a result of the use of information contained within this document, including,
but not limited to, — errors, omissions, or inaccuracies.
Table of Contents
Getting Started
What is Machine Learning?
Classification of Machine Learning Algorithms
Supervised Learning
Unsupervised Learning
Reinforcement Learning
What is Deep Learning?
What is TensorFlow?
Chapter 1: History of Machine Learning
Chapter 2: Theories of Machine Learning
Chapter 3: Approaches to Machine Learning
Philosophies of Machine Learning
Supervised and Semi-supervised Learning Algorithms
Unsupervised Learning Algorithms
Reinforcement Learning
Chapter 4: Environment Setup
Installing Scikit-Learn
Installing TensorFlow
Chapter 5: Using Scikit-Learn
Loading Datasets
Regression
Chapter 6: k-Nearest Neighbors Algorithm
Splitting the Dataset
Feature Scaling
Training the Algorithm
Evaluating the Accuracy of the Algorithm
Comparing K Value with the Error Rate
Chapter 7: K-Means Clustering
Data Preparation
Visualizing the Data
Creating Clusters
Chapter 8: Support Vector Machines
Importing the Dataset
Preprocessing the Data
Training the Algorithm
Making Predictions
Evaluating the Accuracy of the Algorithm
Chapter 9: Machine Learning and Neural Networks
Feedforward Neural Networks
Recurrent Neural Networks
Chapter 10: Machine Learning and Big Data
Chapter 11: Machine Learning and Regression
Chapter 12: Machine Learning and the Cloud
Benefits of Cloud-Based Machine Learning
Chapter 13: Machine Learning and the Internet of Things (IoT)
Consumer Applications
Commercial Applications
Industrial Applications
Infrastructure Applications
Trends in IoT
Chapter 14: Machine Learning and Robotics
Examples of Industrial Robots and Machine Learning
Neural Networks with Scikit-learn
Chapter 15: Machine Learning and Swarm Intelligence
Swarm Behavior
Applications of Swarm Intelligence
Chapter 16: Machine Learning Models
Chapter 17: Applications of Machine Learning
Chapter 18: Programming and (Free) Datasets
Limitations of Machine Learning
 The Philosophical Objections: Jobs, Evil, and Taking Over the World
Chapter 19: Machine Learning and the Future
Conclusion
Getting Started
If you’ve worked with another type of programming, you’re in luck.
Translating programming languages is a cinch. Many programs use the same
or similar codes and coding techniques. Not only does further learning and
adaptations to various programming languages help you to understand
machinery, but you will also be able to move to and from languages with
veritable ease.
If you’ve never used Python before, it’s common to experience some
difficulties translating what you know into Python’s language. Python,
however, is one of the easiest forms of programming and is a great starter for
anyone who wants to pick up on various languages. Quite often, Python can
be found inexpensively or free, if you’re willing to look for it.
Programming provides the means for electronics and various technology to
operate and learn based on your behaviors. If you are familiar with devices
such as Apple’s Siri, Google Home, and others, you’ll recognize that these
devices are based on programming. Data acquisition through direct
interaction with humans gives these devices the means to grow and enhance.
If you understand this kind of programming or are eager to start learning how
you can make and improve your own devices, continue reading. Now we’ll
concentrate on attaining the skills to work with programming on all levels.
What is Machine Learning?

The learning process begins with data or observations, like instruction, direct
experience, or examples of extracting patterns from the data and using these
patterns to make predictions in the future (Bose, 2019). The primary goal of
machine learning is to allow computers to learn automatically without
intervention by humans and adjust accordingly.
With machine learning, we can analyze large quantities of data. Machine
learning gives us profitable results, but we may need a number of resources
to reach this point. Additional time may be needed to train the machine
learning models.
Classification of Machine Learning Algorithms

The machine learning algorithms can fall in the category of supervised,

unsupervised, or reinforced learning.
Supervised Learning

For the case of supervised learning, the human is expected to provide both
the inputs and the outputs which are desired and furnish the feedback based
on the accuracy of the predictions during training. After completion of the
training, the same algorithm will have to apply that was applied to the next
data.
The concept of supervised learning can be seen as similar to learning under a
teacher’s supervision in human beings. The teacher gives some examples to
the student, and the student then derives new rules and knowledge from these
examples so as to apply this somewhere else.
It is also good for you to know the difference between the regression
problems and classification problems. In regression problems, the target is a
numeric value, while in classification, the target is a class or a tag. A
regression task can help determine the average cost of all houses in London,
while a classification task will help determine the types of flowers based on
the length of their sepals and petals.
Unsupervised Learning

For the case of unsupervised learning, the algorithms do not expect to be

provided with the output data. An approach called deep learning, which is an
iterative approach, is used so as to review the data and arrive at new
conclusions. This makes them suitable for use in processing tasks which are
complex compared to the supervised learning algorithms. This means that the
unsupervised learning algorithms learn solely from examples without
responses to these. The algorithm finds patterns from the examples on its
own.
Supervised learning algorithms work similarly to how humans determine any
similarities between two or more objects. Majority of recommender systems
you encounter when purchasing items online work based on unsupervised
learning algorithms. In this case, the algorithm derives what to suggest to you
to purchase from what you have purchased before. The algorithm has to
estimate the kind of customers whom you resemble, and a suggestion is
drawn from that.
Reinforcement Learning

This type of learning occurs when the algorithm is presented with examples
which lack labels, as is the case with unsupervised learning. However, the
example can be accompanied by positive or negative feedback depending on
the solution which is proposed by the algorithm. It is associated with
applications in which the algorithm has to make decisions, and these
decisions are associated with a consequence. It is similar to trial and error in
human learning.
Errors become useful in learning when they are associated with a penalty
such as pain, cost, loss of time, etc. In reinforced learning, some actions are
more likely to succeed compared to others.
Machine learning processes are similar to those of data mining and predictive
modeling. In both cases, searching through the data is required so as to draw
patterns and adjust the actions of the program accordingly. A good example
of machine learning is the recommender systems. If you purchase an item
online, you will get an ad which is related to that item, and that is a good
example of machine learning.
What is Deep Learning?

It is through deep learning that a computer is able to learn to perform

classification tasks directly from text, images, or sound. Deep learning
models are able to achieve a state-of-the-art accuracy, which in some cases
exceeds human-level performance. Large sets of labeled data and neural
network architectures are used to train models in deep learning.
What is TensorFlow?

TensorFlow is a framework from Google used for the creation of deep

learning models. TensorFlow relies on data-flow graphs for numerical
computation. TensorFlow has made machine learning easy. It makes the
processes of acquiring data, training machine learning models, making
predictions, and modifying future results easy.
The library was developed by Google’s Brain team for use in large-scale
machine learning. TensorFlow brings together machine learning and deep
learning algorithms and models, and it makes them much useful via a
common metaphor. TensorFlow uses Python to give its users a front-end API
that can be used for building applications, with the applications being
executed in high-performance C++.
TensorFlow can be used for building, training, and running deep neural
networks for image recognition, handwritten digit classification, recurrent
neural networks, word embedding, natural language processing, etc.
Chapter 1: History of Machine Learning

The history of Machine Learning begins with the history of computing. And
this history began before computers were even invented. The function of
computers is actually math, in particular, Boolean logic. So before anyone
could create a computing machine, the theory about how such a machine
could work had to be figured out by mathematicians. It was the growth of this
theory from hundreds of years ago that made possible the revolution in
computer software that we see today. In a way, the present and future of
computing and machine intelligence belong to great minds from our past.
In 1652, Blaise Pascal, a 19-year-old teenager, created an arithmetic machine
that could add, subtract, multiply, and divide.
In 1679, German mathematician Gottfried Wilhelm Leibniz created a system
of binary code that laid the groundwork for modern computing.
In 1834, English inventor Charles Babbage conceived a mechanical device
that could be programmed with punch cards. While it was never actually
built, its logical structure, Boolean logic, is what nearly every modern
computer relies on to function.
In 1842, Ada Lovelace became the world’s first computer programmer. At 27
years of age, she designed an algorithm for solving mathematical problems
using Babbage’s punch-card technology.
In 1847, George Boole created algebra capable of reducing all values to
Boolean results. Boolean logic is what CPUs use to function and make
decisions.
In 1936, Alan Turing discussed a theory describing how a machine could
analyze and execute a series of instructions. His proof was published and is
considered the base of modern computer science.
In 1943, a neurophysiologist, Warren McCulloch, and a mathematician,
Walter Pitts, co-wrote a paper theorizing how neurons in the human brain
might function. Then they modeled their theory by building a simple neural
network with electrical circuits.
In 1950, Alan Turing proposed the notion of a “learning machine.” This
machine could learn from the world like human beings do and eventually
become artificially intelligent. He theorized about intelligence and described
what was to become known as the “Turing Test” of machine intelligence.
Intelligence, Turing mused, wasn’t well defined, but we humans seem to
have it and to recognize it in others when we experience them using it. Thus,
should we encounter a computer and can’t tell it is a computer when we
interact with it, then we could consider it intelligent too.
In 1951, Marvin Minsky, with the help of Dean Edmonds, created the first
artificial neural network. It was called SNARC (Stochastic Neural Analog
Reinforcement Calculator).
In 1952, Arthur Samuel began working on some of the first machine learning
programs at IBM's Poughkeepsie Laboratory. Samuel was a pioneer in the
fields of artificial intelligence and computer gaming. One of the first things
they built was a machine that could play checkers. More importantly, this
checkers-playing program could learn and improve its game.
Machine Learning was developed as part of the quest in the development of
Artificial Intelligence. The goal of Machine Learning was to have machines
learn from data. But despite its early start, Machine Learning was largely
abandoned in the development of Artificial Intelligence. Like work on the
perceptron, progress in Machine Learning lagged as Artificial Intelligence
moved to the study of expert systems.
Eventually, this focus on a logical, knowledge-based approach to Artificial
Intelligence caused a split between the disciplines. Machine Learning systems
suffered from practical and theoretical problems in representation and
acquiring large data sets to work with. Expert systems came to dominate by
1980, while statistical and probabilistic systems like Machine Learning fell
out of favor. Early neural network research was also abandoned by Artificial
Intelligence researchers and became its own field of study.
Machine Learning became its own discipline, mostly considered outside the
purview of Artificial Intelligence, until the 1990s. Practically, all of the
progress in Machine Learning from the 1960s through to the 1990s was
theoretical, mostly statistics and probability theory. But while not much
seemed to be accomplished, the theory and algorithms produced in these
decades would prove to be the tools needed to re-energize the discipline. At
this point, in the 1990s, the twin engines of vastly increased computer
processing power and the availability of large datasets brought on a sort of
renaissance for Machine Learning. Its goals shifted from the general notion
of achieving artificial intelligence to a more focused goal of solving real-
world problems, employing methods it would borrow from probability theory
and statistics, ideas generated over the previous few decades. This shift and
the subsequent successes it enjoyed brought the field of Machine Learning
back into the fold of Artificial Intelligence, where it resides today as a sub-
discipline under the Artificial Intelligence umbrella.
However, Machine Learning was, continues to be, and might remain a form
of Specific Artificial Intelligence. SAI are software algorithms able to learn a
single or small range of items, which cannot be generalized to the world at
large. To this date, GAI remains an elusive goal, one many believe will never
be reached. Machine Learning algorithms, even if they are specific artificial
intelligence (SAI), are changing the world and will have an enormous effect
on the future.
Chapter 2: Theories of Machine Learning

The goal of Machine Learning stated simply is to create machines capable of

learning about the world so they can accomplish whatever tasks we want
them to do. This sounds simple enough. This is something every human
being accomplishes without any effort. From birth until adulthood, human
beings learn about the world until they are able to master it. Why should this
goal be so difficult for machines?
But even after 70 years of dedicated effort, the goal of general artificial
intelligence remains elusive. As of today, there is nothing even approaching
what could be called a Generalized Artificial Intelligence (GAI). That is an
AI capable of learning from experience with the world, and from that
learning becoming capable of acting in the world with intelligence. Like a
child transitions from a relatively helpless being at birth, spends the next
decade or two experiencing its world, language, culture, and physical reality
surrounding it, then finally becomes an intelligent adult with an astounding
general intelligence. A human being can learn to solve crossword puzzles in
seconds, merely by a quick examination of the wording of the questions and
the design of the field of squares. Drawing on knowledge of language,
puzzles, and the fact we often put letters in little boxes, a human being can
infer how to do a crossword from previous, unrelated experiences. This
simple task would absolutely baffle the most powerful artificial intelligence
on Earth if it had not been specifically trained to do crossword puzzles.
Early Artificial Intelligence started out as an attempt to understand and
employ the rules of thought. The assumption was that thinking is a chain of
logical inference. Such a rules-based system might work like this:
Assertion: Birds can fly.
Fact: A robin is a bird.
Inference: Robins can fly.
Therefore, early Artificial Researchers claimed, encoding thousands of such
inferences in massive databases would allow a machine to make intelligent
claims about the world. A machine equipped with such a database should be
able to make inferences about the world and display intelligence to rival
human intelligence.
Unfortunately, researchers using this rules-based approach very quickly ran
into problems. To continue with our bird example, what about flightless birds
like the ostrich, or penguin, or emu? What about a bird with a broken wing
that, at least temporarily, cannot fly? What do we call a bird in a small cage,
where it cannot fly because of the bars? To human minds confronted with
these exceptions, these questions are simple — of course, they are birds
because flight, although available to most birds, does not define bird-ness.
Rules-based Artificial Intelligence researchers developed language to
accommodate their rules of inference, and these languages were flexible
enough to accept and deal with such deviations as flightless birds. But writing
down and programming all these distinctions, the deviations from the rule
proved far more complicated than anticipated. Imagine attempting to code
every exception to every rule of inference. The task quickly becomes
Herculean and impossible for mere mortals.
The rise of probabilistic systems was in response to this choke point for these
inference methods of artificial intelligence. Probabilistic systems are fed
enormous amounts of relevant data about a subject like birds and are left to
infer what they have in common (sometimes guided with classifying labels –
supervised or semi-supervised, sometimes with no labels at all –
unsupervised). These commonalities are used to modify the algorithm itself
until its output approaches the desired output for the data in question.
This system works okay for birds or cats, but it does not work well for a more
abstract idea like flight. Being fed images of thousands of things flying might
allow an Artificial Intelligence to understand the concept of flight, but this
learning tends to also label clouds as flying, or the sun, or a picture of an
antenna on the top of a building with the sky as a backdrop. And the concept
of flight is much more concrete than something truly abstract, like love or
syntax.
So probabilistic Artificial Intelligence, what we are calling Machine
Learning, switched from attempting to mimic or match human thought and to
deal with concrete, real-world problems that it could solve. This transition in
goals sparked the Machine Learning revolution. By focusing on concrete,
real-world problems rather than generalized ideas, the algorithms of Machine
Learning were able to achieve very powerful, very precise predictions of
future states and to apply those predictions to new data.
Machine Learning today is considered a subset of Artificial Intelligence. The
purpose of Machine Learning is to develop software that can learn for itself.
Essentially, researchers present a set of data with labels to a piece of
software. The software ingests these examples and the relevant labels
provided and attempt to extrapolate rules from the data and labels in order to
make decisions on data it has never seen before. This rules-based system to
make new matches is called a classifier.
The name for this approach to handling data is Computational Learning
Theory. It is highly theoretical and mathematical, using statistics in
particular. Computation Learning Theory is the study and analysis of
machine learning algorithms. Learning Theory deals with inductive learning,
a method known as supervised learning.
The goal of supervised learning is to correctly label new data as it is
encountered, in the process reducing both the number of mislabels and the
speed of the pattern matching process itself. However, there are hard limits to
the performance of Machine Learning algorithms. The facts are data sets used
for training are always finite, and the future is uncertain. These facts mean
that the results of Machine Learning systems will be probabilistic — it is very
difficult to guarantee a particular algorithm’s accuracy. Instead of guarantees,
researchers will place upper and lower bounds on the probability of an
algorithm’s success at classification.
Chapter 3: Approaches to Machine Learning

There is no formal definition of the approaches to Machine Learning.

However, it is possible to separate the approaches into five loose groups. This
chapter will identify these groups and describe their individual philosophical
approaches to Machine Learning. Each of these philosophies has been
successful in solving different categories of problems that humans need to
deal with.
Finally, we will examine three real-world methods employed in actual
Machine Learning, where some of these philosophies are turned into working
systems.
Philosophies of Machine Learning

Artificial Intelligence has always been inspired by human cognition. And

even though the philosophical methods described below rely heavily on
complex math and statistics, each of them is an attempt to mimic, if not the
process of human cognition, then at least the means whereby human
cognition seems to function. This is an artificial intelligence that does not
concern itself with how people or machines think, but only with whether the
results of such thinking produce meaningful results.
Inverse Deduction
Human beings are hard-wired to fill in gaps. From the attachment of the
ocular nerve to the retina (creating a black spot in our vision where we can’t
actually see) to creating faces where none exist (the face in the moon, for
example), our brains are designed to fill in details in order to make patterns.
This is a form of pattern induction. Deduction is the opposite. It is the process
of moving from general experience to individual details. Inductive Artificial
Intelligence is a more mathematical approach to this gap in the pattern filling.
For example, if our algorithm knows 5 + 5 = 10, then it still is able to induce
an answer if we give it 5 and ask it what needs to be added to this 5 to end up
with 10. The system must answer the question “what knowledge is missing?”
The required knowledge to answer this question is generated through the
ingestion of data sets.
Making Connections
Another philosophy of Artificial Intelligence attempts to mimic some
functions of the human brain. Scientists have been studying how human
brains function for decades, and some researchers in Artificial Intelligence,
beginning in earnest in the 1950s, have attempted to represent,
mathematically, the neural architecture of the human brain. We call this
approach the neural network. They are composed of many mathematical
nodes, analogous to neurons in the brain. Each node has an input, a weight,
and an output. They are often layered, with the output of some neurons
becoming the input of others. These systems are trained on data sets, and with
each iteration, the weights attributed to each node is adjusted as required in
order to more closely approximate the desired output. The neural network
approach has had much of the success in machine learning, lately in
particular, because it is suited for dealing with large amounts of data. It is
important to understand, however, that the term neural net is more of an
inspiration than an attempt to replicate how human brains actually function.
The complex interconnected web of neurons in the human brain is not
something a neural net attempts to emulate. Instead, it is a series of
mathematical, probabilistic inputs and outputs that attempts to model the
results of the human neural activity.
Evolution
This approach is to mimic biological evolution. A variety of algorithms are
tasked with solving a problem and exposed to an environment where they
compete to produce the desired output. Each individual algorithm has a
unique set of characteristics with which it can attempt to accomplish the task.
After each iteration, the most successful algorithms are rewarded just like a
successful animal in nature with higher fitness, meaning they are more likely
to pass their features on to the next generation. Those that do continue to the
next iteration can cross their feature sets with other successful algorithms,
mimicking sexual reproduction. Add in some random “mutations,” and you
have the next generation, ready to iterate over the data again. This method
can produce some powerful learning in its algorithms.
Bayesian
The Bayesian approach uses probabilistic inference to deal with uncertainty.
It is a method of statistical inference where Bayes’ theorem is employed to
update hypothesis probabilities as new evidence becomes available. After a
number of iterations, some hypotheses become more likely than others. A
real-world example of this approach is Bayesian spam filtering. How it works
is fairly simple. It has a collection of words that it compares to in the content
of emails. It compares these words to the contents of email messages,
increasing or decreasing the probability of the email being spam according to
what it finds. Bayesian algorithms have been quite successful over the last
two decades because they are adept at building structured models to real-
world problems which have previously been considered intractable.
By Analogy
The fifth philosophical approach to Machine Learning involves using an
analogy. This is yet another tactic borrowed from human cognition.
Reasoning by analogy is a powerful tool humans use to incorporate new
information by comparing it to similar information that is already known. By
comparing this new information to established categories in our mental
repertoire, human beings are able to classify and incorporate new information
even when very little detail about this new information has been provided.
The label for this approach is the “nearest neighbor” principle – basically
asking what this new information is more similar to and categorizing it based
on its similarity to known items. This approach is used in Support Vector
Machines and until recently was likely the most powerful Machine Learning
philosophy in use. An example of a real-world Support Vector Machine can
be found in features like movie recommendations. If you and a stranger have
given 5 stars to one movie and only one star to a different movie, the
algorithm will recommend another movie that this stranger has rated
favorably, assuming by analogy that your tastes are similar, and therefore,
you will appreciate this movie recommendation as well.
What these philosophical approaches to Machine Learning tell us is twofold
— the discipline, despite its roots in the mid-20 century, is still a nascent
th

discipline that continues to grow into its own understanding. After all,
Machine Learning is attempting to model the capacity of the most complex
device in the known universe, the human brain.
As the history of artificial intelligence demonstrates, we can see there is no
“best” approach to Machine Learning. The “best” approach is only the best
approach for now. Tomorrow, a new insight might surpass everything that’s
been accomplished to date, and pave a new path for artificial intelligence to
follow.
How do collections of computer code and mathematics thus far managed to
learn? Here are the three most common methods currently in use to get
computer algorithms to learn.
Supervised and Semi-supervised Learning
Algorithms

In the supervised approach to Machine Learning, researchers first construct a

mathematical model of a data set which includes the expected inputs and the
required outputs. The data this produces is called training data and consists of
sets of training examples (input and desired output). Each of these training
examples is comprised of one or more inputs, along with the desired output –
the supervisory signal. In semi-supervised learning, some of the desired
output values are missing from the training examples.
Through iteration, that is, running the training data through the learning
algorithm repeatedly, the learning algorithm develops a function to match the
desired outputs from the inputs of the training data. During the iteration, this
function is optimized by the learning algorithm. When this system is deemed
ready, new data sets are introduced that are missing the desired outputs –
these are known as the testing sets. At this point, errors or omission in the
training data may become more obvious, and the process can be repeated
with new or more accurate output requirements. An algorithm that
successfully modifies itself to improve the accuracy of its predictions or
outputs can be said to have successfully learned to perform a task.
Supervised and semi-supervised learning algorithms include two classes of
data handling — classification and regression. When outputs are limited by a
constrained set of possible values, classification learning algorithms are
employed. But when the outputs can be returned as a numerical value in a
range of possible values, regression learning algorithms are the best fit.
Finally, there is similarity learning. While closely related to regression and
classification learning, the goal of similarity learning is to examine data sets
and determine how similar or different the information sets are. Similarity
learning can be used in tasks such as ranking, recommendation systems
(Netflix recommended anyone?), visual identity tracking, facial
recognition/verification, and speaker (audio) verification.
Unsupervised Learning Algorithms

In unsupervised learning, algorithms are simply given large data sets

containing only inputs. Their goal is to find grouped or clustered data points
that can be compared to new data sets. Unsupervised learning algorithms
learn from data that has no labels and has not been organized or classified
before being submitted. Instead of attempting to produce a required output
prompted by supervisory systems, these algorithms attempt to find
commonalities in the inputs they receive, which they then apply to new data
sets. They react when these commonalities are found, missing, or broken in
each new data set. Unsupervised learning algorithms are used in diverse
fields including density estimation in statistics and the summarizing and
explanation of data features. This type of learning can also be useful in fraud
detection, where the goal is to find anomalies in input data.
Cluster analysis is when unsupervised learning algorithms break down a set
of observations about data in clusters (subsets) so that the information within
each cluster is similarly based on one or more predefined criteria. Information
drawn from other clusters will be internally similar, while dissimilar from
each other. There are different approaches to data clustering, which are
derived from making alternative assumptions about the structure of the data.
Reinforcement Learning

Reinforcement learning is an approach to Machine Learning that attempts to

“reward” systems for taking actions in their environment that are in
alignment with the objectives of the system. In this way, software trains itself
using trial and error over sets of data until it induces the reward state. This
field is actually quite large and does not solely come the purview of Machine
Learning. It is also employed in disciplines like game theory, swarm
intelligence, simulation-based optimization, information theory, and more.
Reinforcement learning for Machine Learning often employs an environment
characterized by a Markov Decision Process – that is, a mathematical model
to deal with a situation where the cause of the outcome is part random and
part under control of the software decision maker. In addition, reinforcement
learning algorithms typically employ dynamic programming techniques. This
is a method whereby tasks are broken down into sub-tasks, where possible, to
be regrouped once each sub-task has been accomplished into a solution for
the main task. Reinforcement learning algorithms do not need to assume
there is an exact model of the Markov Decision Process for them to employ.
Instead, these algorithms are successfully employed when such exact models
are not feasible.
Typical uses for Machine Learning reinforcement learning algorithms are in
learning to play games against human opponents or in self-
driving/autonomous vehicles.
Chapter 4: Environment Setup

Before getting into the practical part of machine learning and deep learning,
we need to install our two libraries, that is, Scikit-Learn and TensorFlow.
Installing Scikit-Learn

Scikit-learn is supported in Python 2.7 and above. Before installing Scikit-

learn, ensure that you have Numpy and SciPy libraries already installed.
Once you have installed these, you can go ahead to install Scikit-learn on
your machine.
The installation of these libraries can be done using pip. Pip is a tool that
comes with Python, which means that you get pip after installing Python. To
install sci-kit-learn, run the following command on the terminal of your
operating system:
pip install sci-kit-learn
The installation should run and come to completion.
You can also use conda to install scikit-learn.
conda install scikit-learn
Once the installation of scikit-learn is complete, you need to import it into
your Python program in order to use its algorithms. This can be done by
using the import statement as shown below:
import sklearn
If the command runs without an error, know that the installation of scikit-
learn was successful. If the command generates an error, note that the
installation was not successful.
You can now use scikit-learn to create your own machine learning models.
Installing TensorFlow

TensorFlow comes with APIs for programming languages like C++, Haskell,
Go, Java, Rust, and it comes with a third-party package for R known as
tensorflow. On Windows, TensorFlow can be installed with pip or
Anaconda.
The native pip will install the TensorFlow on your system without having to
go through a virtual environment. However, note that the installation of
TensorFlow with pip may interfere with other Python installations on your
system. However, the good thing is that you only have to run a single
command and TensorFlow will be installed on your system. Also, when
TensorFlow is installed via pip, users will be allowed to run the TensorFlow
programs from the directory they want.
To install TensorFlow with Anaconda, you may have to create a virtual
environment. However, within the Anaconda itself, it is recommended that
you install TensorFlow via the pip install command rather than the conda
install command.
Ensure that you have installed Python 3.5 and above on your Windows
system. Python3 comes with a pip3 program which can be used for
installation of TensorFlow. This means we should use the pip3 install
command for installation purposes. The following command will help you
install the CPU-only version for TensorFlow:
pip3 install --upgrade tensorflow
The command should be run from the command line:

If you need to install a GPU version for TensorFlow, run this command:
 pip3 install --upgrade tensorflow-gpu
This will install TensorFlow on your Windows system.
You can also install TensorFlow with the Anaconda package. Pip comes
installed with Python, but Anaconda doesn’t. This means that to install
TensorFlow with Anaconda, you should first install the Anaconda. You can
download Anaconda from its website and find the installation instructions
from the same site.
Once you install Anaconda, you get a package named conda, which is good
for the management of virtual environments and the installation of packages.
To get to use this package, you should start the Anaconda.
On Windows, click Start, choose “All Programs,” expand the “Anaconda
…” folder then click the “Anaconda Prompt.” This should launch the
anaconda prompt on your system. If you need to see the details of the conda
package, just run the following command on the terminal you have just
opened:
conda info
This should return more details regarding the package manager.
There is something unique with Anaconda. It helps us create a virtual Python
environment using the conda package. This virtual environment is simply an
isolated copy of Python with the capability of maintaining its own files,
paths, and directories so that you may be able to work with specific versions
of Python or other libraries without affecting your other Python projects
(Samuel, 2018). Virtual environments provide us with a way of isolating
projects and avoid problems that may arise as a result of version requirements
and different dependencies across various components. Note that this virtual
environment will remain separate from your normal Python environment
meaning that the packages installed in the virtual environment will not affect
the ones you have in your Python’s normal environment.
We need to create a virtual environment for the TensorFlow package. This
can be done via the conda create command. The command takes the syntax
given below:
 conda create -n [environment-name]
In our case, we need to give this environment the name tensorenviron. We
can create it by running the following command:
conda create -n tensorenviron
Allow the process of creating the environment to continue or not. Type “y”
and hit the enter key on the keyboard. The installation will continue
successfully.

After creating an environment, we should activate it so that we may be able

to use it. The activation can be done using the activate command followed by
the name of the environment as shown below:
activate tensorenviron
Now that you have activated the TensorFlow environment, you can go ahead
and install TensorFlow package in it. You can achieve this by running the
following command (Newell, 2019):
conda install tensorflow
You will be presented with a list of packages that will be installed together
with the TensorFlow package. Just type “y,” then hit the enter key on your
keyboard. The installation of the packages will begin immediately. The
process may take a number of minutes, so remain patient. However, the speed
of your internet connection will determine how long this process takes. The
progress of the installation process will also be shown on the prompt
window.
After some time, the installation process will complete, and it will be time for
you to verify whether the installation was successful. We can simply do this
by running the Python’s import statement. The statement should be run from
Python’s terminal. While on the Anaconda prompt, type python and hit the
enter key. This should take you to the Python terminal. Now run the
following import statement:
import tensorflow as tf
If the package was not installed successfully, you would get an error;
otherwise, the installation of the package was successful.
Chapter 5: Using Scikit-Learn

Now that you are done with the installations, you can begin to use the
libraries. We will begin with the Scikit-Learn library.
To be able to use scikit-learn in your code, you should first import it by
running this statement:
import sklearn
Loading Datasets

Machine learning is all about analyzing sets of data. Before this, we should
first load the dataset into our workspace. The library comes loaded with a
number of datasets that we can load and work with. We will demonstrate this
by using a dataset known as Iris. This is a dataset of flowers.
Regression

Linear regression
It is the most popular type of predictive analysis. Linear regression involves
the following two things:
1. Do the predictor variables forecast the results of an outcome variable
accurately?
2. Which particular variables are key predictors of the final variable, and
in what standard does it impact the outcome variable?
Naming variables
The regression’s dependent variable has many different names. Some names
include outcome variable, criterion variable, and many others. The
independent variable can be called exogenous variable or repressors.
Functions of the regression analysis
1. Trend Forecasting
2. Determine the strength of predictors
3. Predict an effect
Breaking down regression
There are two basic states of regression-linear regression and multiple
regression. Although there are different methods for complex data and
analysis, linear regression contains an independent variable to help forecast
the outcome of a dependent variable. On the other hand, multiple regression
has two or more independent variables to assist in predicting a result.
Regression is very useful to financial and investment institutions because it is
used to predict the sales of a particular product or company based on the
previous sales and GDP growth among many other factors. The capital
pricing model is one of the most common regression models applied in
finance. The example below describes the formulae used in linear and
multiple regression.

Choosing the best regression model

Selecting the right linear regression model can be very hard and confusing.
Trying to model it with a sample data cannot make it easier. This section
reviews some of the most popular statistical methods which one can use to
choose models and challenges that you might come across. It also lists some
practical advice to use to select the correct regression model.
It always begins with a researcher who would like to expand the relationship
between the response variable and predictors. The research team that is
accorded with the responsibility to perform investigation essentially measures
a lot of variables but only has a few in the model. The analysts will make
efforts to reduce the variables that are different and apply the ones which
have an accurate relationship. As time moves on, the analysts continue to add
more models.
Statistical methods to use to find the best regression model
If you want a great model in regression, then it is important to take into
consideration the type of variables which you want to test as well as other
variables which can affect the response.
Modified R-squared and Predicted R-squared.
Your model should have higher modified and predicted R-squared values.
The statistics are shown below help eliminate critical issues which revolve
around R-squared.
 The adjusted R-squared increases once a new term improves the
model.
Predicted R-squared belongs to the cross-validation that helps define
the manner in which your model can generalize remaining data sets.
P-values for the Predictors
When it comes to regression, a low value of P denotes statistically significant
terms. The term “Reducing the model” refers to the process of factoring in all
candidate predictors contained in a model.
Stepwise regression
This is an automated technique which can select important predictors found
in the exploratory stages of creating a model.
Real World Challenges
There are different statistical approaches for choosing the best model.
However, complications still exist.
The best model happens when the variables are measured by the study.
The sample data could be unusual because of the type of data collection
method. A false positive and false negative process happens when you
handle samples.
If you deal with enough models, you’ll get variables that are significant
but only correlated by chance.
P-values can be different depending on the specific terms found in the
model.
Studies have discovered that the best subset regression and stepwise
regression can’t select the correct model.
Finding the correct Regression Model
Theory
Study research done by other experts and reference it in your model. It is
important that before you start regression analysis, you should develop ideas
about the most significant variables. Developing something based on
outcome from other people eases the process of collecting data.
Complexity
You may think that complex problems need a complex model. Well, that is
not the case because studies show that even a simple model can provide an
accurate prediction. Once there is a model with the same explanatory
potential, the simplest model is likely to be a perfect choice. You just need to
start with a simple model as you slowly advance the complexity of the
model.
How to calculate the accuracy of the predictive model
There are different ways in which you can compute the accuracy of your
model. Some of these methods are listed below:
1. You divide the data sets into test and training data sets. Next, build the
model based on the training set and apply the test set as a holdout
sample to measure your trained model with the test data.
2. Another method is to calculate the “Confusion Matrix” to the computer
False Positive Rate and False Negative Rate. These measures will
allow a person to choose whether to accept the model or not. If you
consider the cost of the errors, it becomes a critical stage of your
decision whether to reject or accept the model.
3. Computing Receiver Operating Characteristic Curve (ROCC) or the
Lift Chart or Area under the curve (AUC) are other methods that you
can use to decide on whether to reject or accept a model.
Chapter 6: k-Nearest Neighbors Algorithm

The KNN algorithm is highly used for building more complex classifiers. It is
a simple algorithm, but it has outperformed many powerful classifiers. That is
why it is used in numerous applications of data compression, economic
forecasting, and genetics.
KNN is a supervised learning algorithm, which means that we are given a
labeled dataset made up of training observations (x, y), and our goal is to
determine the relationship between x and y (Karbhari, 2019). This means
that we should find a function that x to y such that when we are given an
input value for x, we are able to predict the corresponding value for y.
The concept behind the KNN algorithm is very simple. It calculates the
distance of the new data point to all the other training data points. The
distance can be of various types including Manhattan, Euclidean, etc. The K-
nearest data points are chosen, in which K can be any integer. Finally, the
data point is assigned to the class in which most of the K data points belong
to.
We will use a dataset named Iris. We had explored it previously. We will be
using this to demonstrate how to implement the KNN algorithm. This dataset
is made up of four attributes namely sepal-width, sepal-length, petal-width
and petal-length. Each type of the Iris plant has certain attributes. Our goal
is to predict the class to which a plant belongs to. The dataset has three
classes Iris-setosa, Iris-versicolor, and Iris-virginica.
We now need to load the dataset into our working environment. Download it
from the following URL:
https://archive.ics.uci.edu/ml/machine-learning-databases/iris/iris.data
Store it in a Pandas data frame. This following script will help us achieve
this:
# create a variable for the dataset url
iris_url = "https://archive.ics.uci.edu/ml/machine-learning-
databases/iris/iris.data"
# Assign column names to the dataset
iris_names = ['Slength', 'Swidth', 'Plength', 'Pwidth', 'Class']
# Load the dataset from the url into a pandas dataframe
dataset = pd.read_csv(iris_url, names=iris_names)
We can have a view of the first few rows of the dataset:
print(dataset.head())

The S is for Sepal while P is for Petal. For example, Slength represents Sepal
length while Plength represents Petal length.
import sys
sys.__stdout__ = sys.stdout
url = "https://archive.ics.uci.edu/ml/machine-learning-databases/iris/iris.data"
# Assign colum names to our dataset
names = ['Slength', 'Swidth', 'Plength', 'Pwidth', 'Class']
# Read the dataset to a pandas dataframe
dataset = pd.read_csv(url, names=names)
As usual, we should divide the dataset into attributes and labels:
X = dataset.iloc[:, :-1].values
y = dataset.iloc[:, 4].values
The variable X will hold the first four columns of the dataset which are the
attributes while the variable y will hold the labels.
Splitting the Dataset

We need to be able to tell how well our algorithm performed. This will be
done during the testing phase. This means that we should have training and
testing data. We need to split the data into two such parts. 80% of the data
will be used as the training set while 20% will be used as the test set.
Let us first import the train_test_split method from Scikit-Learn:
from sklearn.model_selection import train_test_split
We can then split the two as follows:
Feature Scaling

Before we can make the actual predictions, it is a good idea for us to scale the
features. After that, all the features will be evaluated uniformly. Scikit-Learn
comes with a class named StandardScaler, which can help us perform the
feature scaling. Let us first import this class:
from sklearn.preprocessing import StandardScaler
We then instantiate the class, and use it to fit a model based on it:
feature_scaler = StandardScaler()
feature_scaler.fit(X_train)
X_train = feature_scaler.transform(X_train)
X_test = feature_scaler.transform(X_test)
The instance was given the name feature_scaler.
Training the Algorithm

With the Scikit-Learn library, it is easy for us to train the KNN algorithm. Let
us first import the KNeighborsClassifier from the Scikit-Learn library:
from sklearn.neighbors import KNeighborsClassifier
The following code will help us train the algorithm:
knn_classifier = KNeighborsClassifier(n_neighbors=5)
knn_classifier.fit(X_train, y_train)
Note that we have created an instance of the class we have created and named
the instance knn_classifier. We have used one parameter in the instantiation,
that is, n_neighbors. We have used 5 as the value of this parameter, and this
basically denotes the value of K. Note that there is no specific value for K,
and it is chosen after testing and evaluation. However, for a start, 5 is used as
the most popular value in most KNN applications.
We can then use the test data to make predictions. This can be done by
running the script given below:
pred_y = knn_classifier.predict(X_test)
Evaluating the Accuracy of the Algorithm

Evaluation of the KNN algorithm is not done in the same way as evaluating
the accuracy of the linear regression algorithm. We were using metrics like
RMSE, MAE, etc. In this case, we will use metrics like confusion matrix,
precision, recall, and f1 score.
We can use the classification_report and confusion_matrix methods to
calculate these metrics. Let us first import these from the Scikit-Learn library:
from sklearn.metrics import confusion_matrix, classification_report
Run the following script:
print(confusion_matrix(y_test, pred_y))
print(classification_report(y_test, pred_y))

The results given above show that the KNN algorithm did a good job in
classifying the 30 records that we have in the test dataset. The results show
that the average accuracy of the algorithm on the dataset was about 90%. This
is not a bad percentage.
Comparing K Value with the Error Rate

We earlier said that there is no specific value of K that can be said to give the
best results on the first go. We chose 5 because it is the most popular value
used for K. The best way to find the best value of K is by plotting a graph of
K value and the corresponding error for the dataset.
Let us create a plot using the mean error for predicted values of the test set
for all the K values that range between 1 and 40. We should begin by
calculating the mean of error for the predicted value with K ranging between
1 and 40. Just run the script given below:
error = []
# K values range between 1 and 40
for x in range(1, 40):
knn = KNeighborsClassifier(n_neighbors=x)
knn.fit(X_train, y_train)
pred_x = knn.predict(X_test)
error.append(np.mean(pred_x != y_test))
The code will run the loop from 1 to 40. In every iteration, the mean error for
the predicted value of the test set is calculated and the result is appended to
error list.
We should now plot the values of error against the values of K. The plot can
be created by running the script given below:
plt.figure(figsize=(12, 6))
plt.plot(range(1, 40), error, color='blue', linestyle='dashed', marker='o',
markerfacecolor='blue', markersize=10)
plt.title('Error Rate for K')
plt.xlabel('K Values')
plt.ylabel('Mean Error')
plt.show()

The code generates the plot given below:

The graph shows that we will get a Mean Error of 0 when we use values of K
between 1 and 17. It will then be good for you to play around with the value
of K and see its impact on the accuracy of the predictions.
Chapter 7: K-Means Clustering

Clustering falls under the category of unsupervised machine learning

algorithms. It is often applied when the data is not labeled. The goal of the
algorithm is to identify clusters or groups within the data.
The idea behind the clusters is that the objects contained in one cluster are
more related to one another than the objects in the other clusters. The
similarity is a metric reflecting the strength of the relationship between two
data objects. Clustering is highly applied in exploratory data mining. It has
many uses in diverse fields such as pattern recognition, machine learning,
information retrieval, image analysis, data compression, bio-informatics, and
computer graphics.
The algorithm forms clusters of data based on the similarity between
data values. You are required to specify the value of K, which is the
number of clusters that you expect the algorithm to make from the data.
The algorithm first selects a centroid value for every cluster. After that,
it performs three steps in an iterative manner:
1. Calculate the Euclidian distance between every data instance and the
centroids for all clusters.
2. Assign the instances of data to the cluster of centroid with the nearest
distance.
3. Calculate the new centroid values depending on the mean values of the
coordinates of the data instances from the corresponding cluster.
Let us manually demonstrate how this algorithm works before implementing
it on Scikit-Learn:
Suppose we have two dimensional data instances given below by the name
D[1]:
D = { (5,3), (10,15), (15,12), (24,10), (30,45), (85,70), (71,80), (60,78),
(55,52), (80,91) }
Our goal is to divide the data into two clusters, namely C1 and C2 depending
on the similarity between the data points.
We should first initialize the values for the centroids of both clusters, and
this should be done randomly. The centroids will be named c1 and c2 for
clusters C1 and C2 respectively, and we will initialize them with the
values for the first two data points, that is, (5,3) and (10,15). It is after
this that you should begin the iterations.
Anytime that you calculate the Euclidean distance, the data point should
be assigned to the cluster with the shortest Euclidean distance. Let us
take the example of the data point (5,3):
Euclidean Distance from the Cluster Centroid c1 = (5,3) = 0
Euclidean Distance from the Cluster Centroid c2 = (10,15) = 13
The Euclidean distance for the data point from point centroid c1 is shorter
compared to the distance of the same data point from centroid c2. This
means that this data point will be assigned to the cluster C1 .
Let us take another data point, (15,12):
Euclidean Distance from the Cluster Centroid c1 = (5,3) is 13.45
Euclidean Distance from the Cluster Centroid c2 = (10,15) is 5.83
Centroid c2 is shorter. Thus, it will be assigned to the cluster C2.
Now that the data points have been assigned to the right clusters, the next
step should involve calculations of the new centroid values. The values
should be calculated by determining the means of the coordinates for the data
points belonging to a certain cluster.
If, for example, for C1, we had allocated the following two data points to
the cluster,
(5, 3) and (24, 10), the new value for x coordinate will be the mean of the
two:
x = (5 + 24) / 2
x = 14.5
The new value for y will be:
y = (3 + 10) / 2
y = 13/2
y = 6.5

The new centroid value for the c1 will be (14.5, 6.5).

This should be done for c2, and the entire process should be repeated. The
iterations should be repeated until when the centroid values do not update
anymore. This means if for example, you do three iterations, you may find
that the updated values for centroids c1 and c2 in the fourth iterations are
equal to what we had in iteration 3. This means that your data cannot be
clustered any further.
You are now familiar with how the K-Means algorithm works. Let us discuss
how you can implement it in the Scikit-Learn library.
Let us first import all the libraries that we need to use:
import matplotlib.pyplot as plt[2]
import numpy as np
from sklearn.cluster import KMeans
Data Preparation

We should now prepare the data that is to be used. We will be creating a

numpy array with a total of 10 rows and 2 columns. So, why have we
chosen to work with a numpy array? It is because Scikit-Learn library can
work with the numpy array data inputs without the need for preprocessing.
Let us create it:
[80,91],])
Visualizing the Data

Now that we have the data, we can create a plot and see how the data points
are distributed. We will then be able to tell whether there are any clusters at
the moment:
plt.show()
The code gives the following plot:

If we use our eyes, we will probably make two clusters from the above data,
one at the bottom with five points and another one at the top with five points.
We now need to investigate whether this is what the K-Means clustering
algorithm will do.
Creating Clusters

We have seen that we can form two clusters from the data points; hence the
value of K is now 2. These two clusters can be created by running the
following code:
kmeans_clusters = KMeans(n_clusters=2)
kmeans_clusters.fit(X)
We have created an object named kmeans_clusters, and 2 have been used as
the value for the parameter n_clusters. We have then called the fit() method
on this object and passed the data we have in our numpy array as the
parameter to the method.
We can now have a look at the centroid values that the algorithm has created
for the final clusters:
print (kmeans_clusters.cluster_centers_)
This returns the following:

The first row above gives us the coordinates for the first centroid, which is,
(16.8, 17). The second row gives us the coordinates of the second centroid,
which is, (70.2, 74.2). If you followed the manual process of calculating the
values of these, they should be the same. This will be an indication that the
K-Means algorithm worked well.
The following script will help us see the data point labels:
print(kmeans_clusters.labels_)
This returns the following:

The above output shows a one-dimensional array of 10 elements which

correspond to the clusters that are assigned to the 10 data points. You clearly
see that we first have a sequence of zeroes which shows that the first 5 points
have been clustered together while the last five points have been clustered
together. Note that the 0 and 1 have no mathematical significance, but they
have simply been used to represent the cluster IDs. If we had three clusters,
then the last one would have been represented using 2s.
We need to plot the data points alongside their assigned labels to be able to
distinguish the clusters. Just execute the script given below:
plt.scatter(X[:,0],X[:,1], c=kmeans_clusters.labels_, cmap='rainbow')[3]
plt.show()
The script returns the following plot:

We have simply plotted the first column of the array named X against the
second column. At the same time, we have passed kmeans_labels_ as the
value for parameter c which corresponds to the labels. Note the use of the
parameter cmap='rainbow'. This parameter helps us to choose the color type
for the different data points.
As you expected, the first five points have been clustered together at the
bottom left and assigned a similar color. The remaining five points have been
clustered together at the top right and assigned one unique color.
We can choose to plot the points together with the centroid coordinates for
every cluster to see how the positioning of the centroid affects clustering. Let
us use three clusters to see how they affect the centroids. The following script
will help you to create the plot:
plt.scatter(X[:,0], X[:,1], c=kmeans_clusters.labels_, cmap='rainbow')[4]
plt.scatter(kmeans_clusters.cluster_centers_[:,0]
,kmeans_clusters.cluster_centers_[:,1], color='black')
plt.show()
The script returns the following plot:

We have chosen to plot the centroid points in black color.

Chapter 8: Support Vector Machines

SVMs fall under the category of supervised machine learning algorithms and
are highly applied classification and regression problems. It is known for its
ability to handle nonlinear input spaces. It is highly applied in applications
like intrusion detection, face detection, classification of news articles, emails
and web pages, handwriting recognition and classification of genes.
The algorithm works by segregating the data points in the best way possible.
The distance between the nearest points is referred to as the margin. The goal
is to choose a hyperplane with the maximum possible margin between the
support vectors in a given dataset.
To best understand how this algorithm works, let us first implement it in
Scikit-Learn library. Our goal is to predict whether a bank currency note is
fake or authentic. We will use the attributes of the note including variance of
the image, the skewness of the wavelet transformed image, the kurtosis of the
image and entropy of the image. Since this is a binary classification
algorithm, let us use the SVM classification algorithm.
If we have a linearly-separable data with two dimensions, the goal of a
typical machine learning algorithm is to identify a boundary that will divide
the data so as to minimize the misclassification error. In most cases, one gets
several lines with all these lines correctly classifying the data.
SVM is different from the other classification algorithms in the way it selects
the decision boundary maximizing the distance from the nearest data points
for all classes. The goal of SVM is not to find the decision boundary only, but
to find the most optimal decision boundary.
The most optimal decision boundary refers to the decision boundary with the
maximum margin from the nearest points of all classes. The nearest points
from the decision boundary maximizing the distance between the decision
boundary and the points are known as support vectors. For the case of
support vector machines, the decision boundary is known as maximum
margin classifier or maximum margin hyperplane.
Complex mathematics is involved in the calculation of the support vectors.
Determine the margin between the decision boundary and support vectors
and maximize the margin.
Let us begin by importing the necessary libraries:
import numpy as np[5]
import pandas as pd
import matplotlib.pyplot as plt
Importing the Dataset

We will use the read_csv method provided by the Pandas library to read the
data and import it into our workspace (Agile Actors, 2018). This can be done
as follows:
dataset = pd.read_csv("bank_note.csv")
Let us call the shape method to print the shape of the data for us:
print(dataset.shape)
This returns the following:

This shows that there are 1372 columns and 5 columns in the dataset. Let us
print the first 5 rows of the dataset:
print(dataset.head())
Again, this may return an error because of the lack of the output information.
Let us solve this using the Python’s sys library. You should now have the
following code:
import numpy as np[6]
import pandas as pd
import matplotlib.pyplot as plt
import sys
sys.__stdout__=sys.stdout
dataset = pd.read_csv("bank_note.csv")
print(dataset.head())
The code returns the following output:
All attributes of the data are numeric as shown above. Even the last attribute
is numeric as its values are either 0 or 1.
Preprocessing the Data

It is now time to subdivide the above data into attributes and labels as well as
training and test sets. The following code will help us subdivide the data into
attributes and labels:
X = dataset.drop('Class', axis=1)
y = dataset['Class']
The first line above helps us store all the columns of the dataset into variable
X, except the class column. The drop() function has helped us exclude the
Class column from this. The second line has then helped us store the Class
column into variable y. The variable X now has attributes while the variable y
now has the corresponding labels.
We have achieved the goal of dividing the dataset into attributes and labels.
The next step is to divide the dataset into training and test sets. Scikit-learn
has a library known as model_selection which provides us with a method
named train_test_split that we can use to divide the data into training and
test sets.
First, let us import the train_test_split method:
from sklearn.model_selection import train_test_split
Training the Algorithm

Now that the data has been split into training and test sets, we should now
train the SVM on the training set. Scikit-Learn comes with a library known as
SVM which has built-in classes for various SVM algorithms.
In this case, we will be doing a classification task. Hence, we will use the
support vector classifier class (SVC). The takes a single parameter, that is,
the kernel type. For a simple SVM, the parameter should be set to “linear”
since the simple SVMs can only classify data that is linearly separable.
We will call the fit method of SVC to train the algorithm on our training set.
The training set should be passed as a parameter to the fit method. Let us first
import the SVC class from Scikit-Learn:
from sklearn.svm import SVC
Now run the following code:
svc_classifier = SVC(kernel='linear')
svc_classifier.fit(X_train, y_train)
Making Predictions

We should use the SVC class for making predictions. Note that the
predictions will be made on the test data. Here is the code for making
predictions:
pred_y = svc_classifier.predict(X_test)
Evaluating the Accuracy of the Algorithm

In classification tasks, we use confusion matrix, recall, precision, and F1 as

the metrics. Scikit-Learn has the metrics library which provides us with the
confusion_matrix and classification_report methods which can help us find
the values of these metrics. The following code can help us find the value for
these metrics:
First, let us import the above methods from the Scikit-Learn library:
from sklearn.metrics import confusion_matrix, classification_report
Here is the code that can help in doing the evaluation:
print(confusion_matrix(y_test,pred_y))
print(classification_report(y_test,pred_y))
The code returns the following:

The output given above shows that the algorithm did a good task; an average
of 99% for the above metrics is not bad.
Let us give another example of how to implement SVM in Scikit-Learn using
the Iris dataset. We had already loaded the Iris dataset, a dataset that shows
details of flowers in terms of sepal and petal measurements, that is, width and
length. We can now learn from the data, and then make a prediction for
unknown data. These call for us to create an estimator, and then call its fit
method.
This is demonstrated in the script given below:
from sklearn import svm
from sklearn import datasets
# Loading the dataset
iris = datasets.load_iris()
clf = svm.LinearSVC()
# learn from the dataset
clf.fit(iris.data, iris.target)
# predict unseen data
clf.predict([[ 6.2, 4.2, 3.5, 0.35]])
# Changing model parameters using the attributes ending with an underscore
print(clf.coef_ )
The code will return the following output:

We now have the predicted values for our data. Note that we imported both
datasets and SVM from the scikit-learn library. After loading the dataset, a
model was fitted/created by learning patterns from the data. This was done by
calling the fit() method. Note that the LinearSVC()method helps us to create
an estimator for the support vector classifier, on which we are to create the
model. We have then passed in new data for which we need to make a
prediction.
Chapter 9: Machine Learning and Neural Networks

Neural networks were first developed in the late 1950s as a means to build
learning systems that were modeled on our understanding of how the human
brain functions. Despite their name, however, there is little resemblance
between a neural network and a human brain. Mostly, the name serves as a
metaphor and as inspiration. Each “neuron” in a Neural Network is consists
of a “node,” a piece serial processing code designed to iterate over a problem
until coming up with a solution at which point, the result is passed on to the
next neuron in the layer, or to the next layer if the current layer’s processing
is complete. In contrast, the human brain is capable of true parallel
processing, by nature of the complex interconnections of its neurons and the
fact its processing specialties are located in different areas of the brain
(vision, hearing, smell, etc.), all of which can process signals simultaneously.
Neural networks are a sub-set of Machine Learning. They consist of software
systems that mimic some aspects of human neurons. Neural Networks pass
data through interconnected nodes. The data is analyzed, classified, and then
passed on to the next node, where further classification and categorization
may take place. The first layer of any Neural Network is the input layer,
while the last is the output layer (these can be the same layer, as they were in
the first, most primitive neural networks). Between these neural layers is any
number of hidden layers that do the work of dealing with the data presented
by the input layer. Classical Neural Networks usually contain two to three
layers. When a Neural Network consists of more layers, it is referred to as
Deep Learning. Such Deep Learning systems can have dozens or even
hundreds of layers.
Feedforward Neural Networks

The first and simplest form of Neural Networks is called feedforward. As the
name implies, data flows through a feedforward network in one direction,
from the input, through the node layers containing the neurons and exits
through the output layer. Unlike more modern neural networks, feedforward
networks do not cycle or iterate over their data. They perform a single
operation on the input data, and they provide their solution in an output
stream.
Single-layer perceptron
This is the simplest form of a feedforward neural network with one layer of
nodes. The input is sent to each node already weighted, and depending on
how the node calculates the input and its weight, a minimal threshold may or
not be met, and the neuron either fires (taking the “activated” value) or does
not fire (taking the “deactivated” value).
Multi-layer perceptron
Multi-layer perceptrons, as the name suggests, consist of two or more
(sometimes many more) layers, with the output of the upper layer becoming
the input of the lower layer. Because there are many layers, this form of
neural network often takes advantage of backpropagation, where the
produced output is compared with the expected output, and the degree of
error is fed back through the network to adjust the weights on the appropriate
nodes, all with the intention of producing an output closer to the desired
output state. Each error correction is tiny, so often a great number of
iterations are required to achieve a “learned” state. At this point, the neural
network is no longer considered a feedforward system proper. It is algorithms
such as these multi-layer networks with backpropagation that have become
some of the most successful and powerful Machine Learning devices in the
field.
Recurrent Neural Networks

Recurrent Neural Networks propagate data in both directions, forward like

feedforward networks but also backward from later to earlier processing
stages. One of the main strengths of recurrent neural networks is the ability to
remember previous states. Before recurrent neural networks, a neural network
would forget previous tasks once it was asked to begin a new one. Imagine
reading a comic but forgetting the content of the previous cells as you read
the next one. Recurrent neural networks allow information to persist,
meaning they can take input from subsequent events and “remember” their
experience in previous ones. Put simply, recurrent neural networks are a
series of input and output networks with the output of the first becoming the
input of the second, the second’s output the input of the third, and so on. This
cycling allows Recurrent Neural Networks to develop closer and closer
approximations to the desired output.
Backpropagation
Theoretical development of backpropagation came from Paul Werbos’ Ph.D.
thesis in 1974. Unfortunately, due to complications and difficulties
encountered by those attempts to create Neural Networks (both theoretical
and hardware-related), Neural Networks did not really catch on until the
1980s, when processor power and large, numerous data sets became available
for training and testing.
The application of backpropagation is highly mathematical but can be
summarized like this. Each iteration of a neural network produces a
mathematical degree of error between the desired output and the actual output
of the network. Each neuron in a neural network has a weight attached to it
for the purposes of modifying its calculation on the input it receives.
Backpropagation uses mathematics (probability statistics) when it is possible,
to calculate the derivative of the error between expected and actual outputs.
This derivative is then used on the next iteration to adjust the weight applied
to each of the neurons. Each subsequent iteration produces a smaller error.
Imagine a basketball dropped somewhere in a skateboard half-pipe. It will
roll down towards the center and up the other side, then reverse direction and
roll towards the center again. Each time, it will roll less distance up the side
before reversing direction. Eventually, gravity will bring the ball to a stable
state in the center. In the same way, backpropagation reduces the error
produced in each iteration as it brings the actual output closer to the desired
output.
We have seen above that the goal of Machine Learning is to make software
able to learn from the data it experiences. This is the goal of Neural Networks
as well, but while Machine Learning makes decisions based on what data it
has seen before, Neural Networks are designed to learn and make intelligent
decisions on their own. This is particularly useful when the patterns being
searched for are too numerous or complex for a software programmer to
extract and submit as part of an input training data set.
Chapter 10: Machine Learning and Big Data

Big Data is pretty much what it sounds like — the practice of dealing with
large volumes of data. And by large, we are talking about astoundingly huge
amounts of data — gigabytes, terabytes, petabytes of data. A petabyte, to put
this size into perspective, is 10 to the 15 bytes. Written out that is 1 PB =
1,000,000,000,000,000 bytes. When you consider that a single byte is
equivalent in storage to a single letter like an ‘a’ or ‘x,’ the scale of the data
sets being dealt with my Big Data is truly awe-inspiring. And these sizes are
increasing every day.
The term Big Data comes from the 1990s, although computer scientists have
been dealing with large volumes of data for decades. What sets Big Data
apart from data sets before is the fact the size of data sets began to
overwhelm the ability of traditional data analytics software to deal with it.
New database storage systems had to be created (Hadoop for example) just to
hold the data and new software written to be able to deal with so much
information in a meaningful way.
Today the term Big Data brings with it a series of assumptions and practices
that have made it a field all its own. Most Big Data discussions begin with
the 3 V’s. Big data is data containing more variety arriving in increasing
volumes and with increasing velocity (acceleration would be an accurate term
to use here, but then we’d lose the alliteration).
Volume
The term volume refers to the vast amount of data available. When the term
Big Data was coined in the early 2000s, the amount of data available for
analysis was overwhelming. Since then, the volume of data created has
grown exponentially. In fact, the volume of data produced has become so
vast that new storage solutions have to be created just to deal with it. This
increase in available data shows no sign of slowing, and data is, in fact,
increasing geometrically by doubling every two years.
Velocity
Along with the rise in the amount of data being created, there has been an
increase in the speed at which it is produced. Things like smartphones, RFID
chips, and real-time facial recognition produce not only enormous amounts of
data; this data is produced in real time and must be dealt with as it is created.
If not processed in real time, it must be stored for later processing. The
increasing speed of this data arriving strains the capacity of bandwidth,
processing power, and storage space to contain it for later use.
Variety
Data does not get produced in a single format. It is stored numerically in
detailed databases, produced in structure-less text and email documents, and
stored digitally in streaming audio and video. There is stock market data,
financial transactions, and so on, all of it uniquely structured. So not only
must large amounts of data be handled very quickly, it is produced in many
formats that require distinct methods of handling for each type.
Lately, two more V’s have been added:
Value
Data is intrinsically valuable, but only if you are able to extract this value
from it. Also, the state of input data, whether it is nicely structured in a
numeric database or unstructured text message chains, affects its value. The
less structure a data set has, the more work needs to be put into it before it
can be processed. In this sense, well-structured data is more valuable than
less-structured data.
Veracity
Not all captured data is of equal quality. When dealing with assumptions and
predictions parsed out of large data sets, knowing the veracity of the data
being used has an important effect on the weight given to the information
studying it generates. There are many causes that limit data veracity. Data can
be biased by the assumptions made by those who collected them. Software
bugs can introduce errors and omission in a data set. Abnormalities can
reduce data veracity like when two wind speed sensors next to each other
report different wind directions. One of the sensors could be failing, but there
is no way to determine this from the data itself. Sources can also be of
questionable veracity — in a company’s social media feed are a series of very
negative reviews. Were they human or bot created? Human error can also be
present such as a person signing up to a web service entering their phone
number incorrectly. And there are many more ways data veracity can be
compromised.
The point of dealing with all this data is to identify useful detail out of all the
noise — businesses can find ways to reduce costs, increase speed and
efficiency, design new products and brands, and make more intelligent
decisions. Governments can find similar benefits in studying the data
produced by their citizens and industries.
Here are some examples of current uses of Big Data.
Product Development
Big Data can be used to predict customer demand. Using current and past
products and services to classify key attributes, they can then model these
attributes’ relationships and their success in the market.
Predictive Maintenance
Buried in structured data are indices that can predict the mechanical failure of
machine parts and systems. Year of manufacture, make and model, and so on,
provide a way to predict future breakdowns. Also, there is a wealth of
unstructured data in error messages, service logs, operating temperature, and
sensor data. This data, when correctly analyzed, can predict problems before
they happen so maintenance can be deployed preemptively, reducing both
cost and system downtime.
Customer Experience
Many businesses are nothing without their customers. Yet acquiring and
keeping customers in a competitive landscape is difficult and expensive.
Anything that can give a business an edge will be eagerly utilized. Using Big
Data, businesses can get a much clearer view of the customer experience by
examining social media, website visit metrics, call logs, and any other
recorded customer interaction to modify and improve the customer
experience, all in the interests of maximizing the value delivered in order to
acquire and maintain customers. Offers to individual customers can become
not only more personalized but more relevant and accurate. By using Big
Data to identify problematic issues, businesses can handle them quickly and
effectively, reducing customer churn and negative press.
Fraud & Compliance
While there may be single rogue bad actors out there in the digital universe
attempting to crack system security, the real threats are from organized, well-
financed teams of experts, sometimes teams supported by foreign
governments. At the same time, security practices and standards never stand
still but are constantly changing with new technologies and new approaches
to hacking existing ones. Big Data helps identify data patterns suggesting
fraud or tampering. Aggregation of these large data sets makes regulatory
reporting much faster.
Operation Efficiency
Not the sexiest topic, but this is the area in which Big Data is currently
providing the most value and return. Data helps companies analyze and
assess production systems, examine customer feedback and product returns,
and examine a myriad of other business factors to reduce outages and waste,
and even anticipate future demand and trends. Big Data is even useful in
assessing current decision-making processes and how well they function in
meeting demand.
Innovation
Big Data is all about relations between meaningful labels. For a large
business, this can mean examining how people, institutions, other entities,
and business processes intersect, and use any interdependencies to drive new
ways to take advantage of these insights. New trends can be predicted, and
existing trends can be better understood. This all leads to understanding what
customers actually want and anticipate what they may want in the future.
Knowing enough about individual customers may lead to the ability to take
advantage of dynamic pricing models. Innovation driven by Big Data is
really only limited by the ingenuity and creativity of the people curating it.
Machine Learning is also meant to deal with large amounts of data very
quickly. But while Big Data is focused on using existing data to find trends,
outliers, and anomalies, Machine Learning uses this same data to “learn”
these patterns in order to deal with future data proactively. While Big Data
looks at the past and present data, Machine Learning examines the present
data to learn how to deal with the data that will be collected in the future. In
Big Data, it is people who define what to look for and how to organize and
structure this information. In Machine Learning, the algorithm teaches itself
what is important through iteration over test data, and when this process is
completed, the algorithm can then go ahead to new data it has never
experienced before.
Chapter 11: Machine Learning and Regression

Machine Learning can produce two distinct output types — classification and
regression. A classification problem involves an input that requires an output
in the form of a label or category. Regression problems, on the other hand,
involve an input that requires an output value as a quantity. Let’s look at each
form in more detail.
Classification Problems
Classification problems are expected to produce an output that is a label or
category. That is to say, a function is created from an examination of the
input data that produces a discrete output. A familiar example of a
classification problem is whether a given email is spam or not spam.
Classification can involve probability, providing a probability estimate along
with its classification. 0.7 spam, for example, suggests a 70% chance an
existing email is spam. If this percentage meets or exceeds the acceptable
level for a spam label, then the email is classified as spam (we have spam
folders in our email programs, not 65% spam folders). Otherwise, it is
classified as not spam. One common method to determine classification
problem’s algorithm probability of accuracy is to compare the results of its
predictive model against the actual classification of the data set it has
examined. On a data set of 5 emails, for example, where the algorithm has
successfully classified 4 out of 5 of the emails in the set, the algorithm could
be said to have an accuracy of 80%. Of course, in a real-world example, the
training data set would be much more massive.
Regression problems
In a regression problem, the expected output is in the form of an unlimited
numerical range. The price of a used car is a good example. The input might
be the year, color, mileage, condition, etc., and the expected output is a dollar
value, such as $4,500 - $6,500. The skill (error) of a regression algorithm can
be determined using various mathematical techniques. A common skill
measure for regression algorithms is to calculate the root mean squared error,
RMSE.
Although it is possible to modify each of the above methods to produce each
other’s result (that is, turn a classification algorithm into a regression
algorithm and vice versa), the output requirements of the two define each
algorithm quite clearly:
1. Classification algorithms produce a discrete category result which can
be evaluated for accuracy, while regression algorithms cannot.
2. Regression algorithms produce a ranging result and can be evaluated
using root mean squared error, while classification algorithms cannot.
So, while Machine Learning employs both types of methods for problem-
solving (classification and regression), what method is employed for any
particular problem depends on the nature of the problem and how the solution
needs to be presented.
Chapter 12: Machine Learning and the Cloud

Cloud computing has entered the public consciousness. As a metaphor, it is

powerful because it takes a complex activity and turns it into a simple idea.
But cloud computing is a complex and expensive undertaking. It involves the
integration of hundreds, thousands, or even tens of thousands of computer
servers in a massive data center (a single Amazon data center can host as
many as 80,000 servers). This “raw iron” serves as the backbone of the cloud
service. The servers operate virtual machine managers called hypervisors,
which can be either software or hardware that uses the raw processing power
of the existing data center servers to simulate other software, be that
individual programs or entire operating systems.
The advantage of cloud computing is the fact the cost associated with all that
hardware in the data center is absorbed by the company providing the cloud
services. This includes the building and infrastructure to operate the
hardware, the electricity to power it, and the staff to maintain, service,
update, and repair hardware and software systems.

These companies monetize their hardware and virtualized software by selling

it on demand to whoever wants to use it, be that other companies,
governments, universities, or individuals. And even though the cloud service
is more expensive than purchasing and using their own servers, the advantage
is in the on-demand nature of the service. A company may have a spike in its
retail website use around Christmas, for example, where demand surges and
would overwhelm the servers that are able to operate without issue in the rest
of the year. Instead of having to purchase more hardware to manage this
spike in use, hardware that would remain essentially dormant the rest of the
year (though still consuming electricity, requiring maintenance, and updates
by trained staff, and so on), this company could instead acquire a cloud
services account and “virtualize” part of its website business using load
balancing to make sure its servers are not overwhelmed. If demand is even
more than expected, in a matter of moments, more servers can be spun up to
carry the demand. As traffic decreases late at night, or for whatever reason,
these extra virtual servers can be shut down, no longer costing the company
money to operate.
Previously, we’ve discussed the forces that allowed Machine Learning to
come into its own, the historical development of the mathematical algorithms
being used, the rise of very large, numerous sets of data, and the increasing
power of computer processors. But even as these factors merged to make
Machine Learning viable, it is extremely expensive. There are enormous
resources required, from qualified software developers who know the math
behind Machine Learning algorithms, to gaining access to large data sets
either through business operations or purchase, to being able to afford the
massive processing power to crunch all that data and allow an algorithm to
learn. Access to Machine Learning has for the most part been only for
governments, research organizations like universities, and large businesses.
The Cloud promises to change this. As Machine Learning systems are
developed and refined by researchers and business ahead of the curve in the
industry, they can often end up being available to consumers in cloud-based
services provided by companies like Amazon AWS, Google Cloud, and
Microsoft Azure. Because these services can provide these algorithms on
demand, the cost per use for them can be quite low. At the same time, cloud-
based storage has become very reasonably priced, meaning those massive
data sets needed for training and testing Machine Learning algorithms can be
stored in the cloud as well, reducing cost for storage, the infrastructure
required to carry the bandwidth needed to move such massive amounts of
data, and the time required to make such transfers. The software and data can
reside in the same cloud account, providing instant access, without any
bandwidth costs.
Companies using cloud computing to implement Machine Learning do not
need an IT department capable of creating and managing the AI
infrastructure, as this is taken care of by the companies offering the cloud
services.
Just as cloud-based services offer SaaS (software as a service) solutions,
Machine Learning cloud services offer SDKs (software developer kits) and
APIs (application programming interfaces) to embed Machine Learning
functions into applications. These connections offer support for most
programming languages, meaning developers working for the companies
using the cloud-based solutions do not need to learn any new languages to
harness the Machine Learning infrastructure. Developers can harness the
power of Machine Learning processes directly in their applications, which is
where the most value lies because most real-world use of Machine Learning
systems today are transaction and operations-focused such as real-time loan
applications, fraudulent transactions, mapping, and route planning, facial
recognition, voice authorization, and so on.
Benefits of Cloud-Based Machine Learning

It is possible to use one of the free open source Machine Learning

frameworks such as CNTK, TensorFlow, and MXNet to create your own AI
solutions, but even though the frameworks are free, the barriers to do it
yourself approach can be prohibitive. We’ve touched on them above —
overall cost for hardware and hardware maintenance, the cost for software
maintenance, and the cost of acquiring and maintaining a team of AI
specialists capable of designing and maintaining Machine Learning systems.
Cloud-based Machine Learning projects make it easy for a company or
organization to use limited levels of resources during experimentation and
training while providing a seamless ability to scale up for production and, in
the future, as demand increases. Whereas with an in-house Machine Learning
option, after training has taken place and the algorithm is ready for
production deployment, an entire infrastructure must be created around the
software for outside systems to use it. With a cloud solution, production is
often as simple as a few mouse clicks away.
Amazon AWS, Google Cloud Platform, and Microsoft Azure have many
Machine Learning services that do not require knowledge of Artificial
Intelligence, theories of Machine Learning, or even a data science team.
At this point, the big cloud service providers provide two basic kinds of
Machine Learning offerings — general and specific. A specific offering, for
example, is Amazon’s Rekognition, an image-recognition Machine Learning
implementation that can be run with a single command. On the other hand, if
what you require is not available from one of these specific user options, all
three services offer more generalized Machine Intelligence solutions
requiring the user to create custom code and to be run on general-purpose
services.
There has been an attempt to make services that offer more general-purpose
Machine Learning services that are simpler to use, but as with many attempts
at generalized software frameworks, the market hasn’t embraced the concept
because while being easier to use, the simplicity of the interface means it is
not possible to get the custom requirements customers are looking for.
Chapter 13: Machine Learning and the Internet of

Things (IoT)
Perhaps the first Internet of Things was at Carnegie Mellon University in
1982 when a Coke machine was connected to the internet so it could report
on its current stock levels, as well as the temperature of newly-loaded drinks.
But at that time, computer miniaturization had not progressed to the point it
has today, so it was difficult for people at that time to conceive of very small
devices connecting to the internet. Also, consumer wireless internet would
not be available for another 15 years, meaning any internet connected device
had to be fairly large and wired to its connection.
Today, the Internet of Things (IoT) refers to all the Internet-connected
devices in the world. Some are obvious like your smartphone, and some are
not so obvious like a smart thermostat in your house. The term also applies to
internet-connected devices that are part of a device like a sensor on an
industrial robot’s arm or the jet engine from a commercial jetliner.
A relatively simple explanation, but it is hard to conceive just how many IoT
devices are around us, never mind how many will be in the near future. The
following is a breakdown of some of the Internet of Things devices currently
in use, which is often broken down into the following categories:
Consumer Applications

The number of Internet of Things devices created for consumer use is

expanding rapidly. From smartphones to connected vehicles to wearable
health technology and health monitoring devices, the Internet of Things is
entering our lives at a rapid pace. Many of these applications are the result of
simple programming methods, like those devised in Python.
Smart Home
Home automation is the concept of having your house manage its resources
for you. These resources include lighting, fridge, air conditioning, security,
media, and so on. The Internet of Things is critical to this philosophy because
the means to manage and automate home resources come through the ability
of the, until now, separate devices being able to connect to the internet and
possibly to each other. These connections allow the owner to adjust the
lighting or temperature or view areas of the home when outside the home. At
the same time, these devices are able to alert the owner, sending a warning
that an external perimeter has been breached and offering a real-time video
display of the culprit – more likely a raccoon than a burglar, but in either
case, the devices can provide a feeling of control. Another example might be
when your smart fridge notices your milk is 90% empty and automatically
adds it to your electronic shopping list.
Clearly, we could come up with the examples of many more devices
currently being used, as well as all those not yet created that will be coming
in the future, that employ various kinds of Machine Learning. The security
system above needs to determine something is moving in the backyard.
Perhaps more advanced Machine Learning would even be able to distinguish
a person from a small animal and not bother you at all. At the same time, it
could possibly use facial recognition to determine it is only your neighbor
who came around to fetch a Frisbee accidentally thrown over the adjoining
fence.
The smart fridge not only needs to be able to classify the items it is holding,
but to judge when enough of an item has been consumed to justify ordering
more – clearly a Machine Learning problem since 70% consumed might
mean less than a day left for a family of three, but 2 days’ supply for a single
person. These details will be learned, as a Machine Learning algorithm in
your fridge first observes the occupants using the fridge and determines the
correct quantities for reorder. When another person moves back in for the
summer after college, a Machine Learning algorithm can observe and adjust
its estimates accordingly. Over time, these predictive models can aggregate
their data, allowing new instances of the algorithm to auto-correct for added
or removed residents, based on the data sets of previous move ins and move
outs.
Elder/Disabled Care
One of the Smart Home’s strongest selling features is its ability to help care
for people with disabilities. Among assistive technologies available is voice
control to activate/deactivate or set features around the house — lighting,
temperature, etc., and the ability to identify medical emergencies such as
seizures or falls. No doubt, these features will expand, including an option for
diagnosis and treatment suggestions, as well as the ability to interface with
emergency services and local hospitals to notify and assist in dealing with
medical emergencies.
Machine Learning in the home is already available in terms of voice
recognition and fall monitoring. It will be instrumental in more extensive
medical monitoring, including the ability to “read” emotions and monitor our
health via external and even internal sensors.
Here again, we see how the integration of the IoT and the health care system
can provide a benefit, but behind many of these qualities of life-improving
devices is the predictive and anomaly detection capacity of Machine Learning
algorithms.
Commercial Applications

Healthcare
We are entering the world of “Smart Healthcare,” where computers, the
internet, and artificial intelligence are merging to improve our quality of life.
The Internet of Medical (or Health) Things (IoHT) is a specific application of
the Internet of Things designed for health and medically related purposes. It
is leading to the digitization of the healthcare system, providing connectivity
between properly equipped healthcare services and medical resources.
Some Internet of (Heath) Things applications enable remote health
monitoring and operate emergency notification systems. This can range from
blood pressure monitoring and glucose levels to the monitoring of medical
implants in the body. In some hospitals, you will find “smart beds” that can
detect patients trying to get up and adjust themselves to maintain healthy and
appropriate pressure and support for the patient, without requiring the
intervention of a nurse or other health professional. One report estimated
these sorts of devices and technology could save the US health care system
$300 billion in a year in increased revenue and decreased costs. The
interactivity of medical systems has also led to the creation of “m-health,”
which is used to collect and analyze information provided by different
connected resources like sensors and biomedical information collection
systems.
Rooms can be equipped with sensors and other devices to monitor the health
of senior citizens or the critically ill, ensuring their well-being, as well as
monitor that treatments and therapies are carried out to provide comfort,
regain lost mobility, and so on. These sensing devices are interconnected and
can collect, analyze, and transfer important and sensitive information. In-
home monitoring systems can interface with a hospital or other health-care
monitoring stations.
There have been many advances in plastics and fabrics allowing for the
creation of very low-cost, throw away “wearable” electronic IoT sensors.
These sensors, combined with RFID technology, are fabricated into paper or
e-textiles, providing wirelessly-powered, disposable sensors.
Combined with Machine Learning, the health care IoHT ecosphere around
each one of us can improve quality of life, guard against drug errors,
encourage health and wellness, and respond to and even predict responses by
emergency personnel. As a collection of technology and software, this future
“smart” health care will cause a profound shift in medical care, where we no
longer wait for obvious signs of illness to diagnose, but instead use the
predictive power of Machine Learning to detect anomalies and predict future
health issues long before human beings even know something might be
wrong.
Transportation
The IoT assists various transportation systems in the integration of control,
communications, and information processing. It can be applied throughout
the entire transportation system — drivers, users, vehicles, and infrastructure.
This integration allows for inter and even intra-vehicle communication,
logistics and fleet management vehicle control, smart traffic control,
electronic toll collection systems, smart parking, and even safety and road
assistance.
In case of logistics and fleet management, an IoT platform provides
continuous monitoring of cargo and asset locations and conditions using
wireless sensors and send alerts when anomalies occur (damage, theft, delays,
and so on). GPS, temperature, and humidity sensors can return data to the IoT
platform where it can be analyzed and set on to the appropriate users. Users
are then able to track in real-time the location and condition of vehicles and
cargo and are then able to make the appropriate decisions based on accurate
information. IoT can even reduce traffic accidents by providing drowsiness
alerts and health monitoring for drivers to ensure they do not drive when they
need rest.
As the IoT is integrated more and more with vehicles and the infrastructure
required to move these vehicles around, the ability for cities to control and
alleviate traffic congestion, for businesses to control and respond to issues
with transportation of their goods, and for both of these groups to work
together, increases dramatically. In unforeseen traffic congestion due to an
accident, for example, sensitive products (a patient in an ambulance on the
way to Emergency or produce or other time-sensitive items) could be put in
the front of the queue to ensure they are delayed as little as possible, all
without the need for traffic direction by human beings.
The potential rewards of such an integrated system are many, perhaps only
limited by our imagination. Using IoT enabled statistics from traffic will
allow for optimum traffic routing, which in term will reduce travel time and
CO2 emissions. Smart stop lights and road signs with variable-speed and
information displays will communicate more and more with the onboard
systems of vehicles, providing routing information to reduce travel time. And
all of this technology is made possible by dedicated Machine Learning
algorithms with access to the near endless flow of data from thousands and
thousands of individual IoT sensors.
Building/Home Automation
As discussed above, IoT devices can be used in any kind of building, where
they can monitor and control the electrical, mechanical, and electronic
systems of these buildings. The advantages identified are:
By integrating the internet with a building’s energy management
systems, it is possible to create “smart buildings” where energy
efficiency is driven by I0T.
Real-time monitoring of buildings provides a means for reducing
energy consumption and the monitoring of occupant behavior.
IoT devices integrated into buildings provide information on how smart
devices can help us understand how to use this connectivity in future
applications or building designs.
When you read “smart” or “real-time” above, you should be thinking
Machine Learning because it is these algorithms that can tie all of the sensor
input together into something predictive and intelligent.
Industrial Applications

As we’ve mentioned above, programming and machines that are commonly

used by society are on an upswing. But what about industrial applications?
You’ve probably had a friend or family member who suffered from a disease
or condition treatable only in the hospital. These types of diseases are best
understood through programmed technology.
Other media, such as university programming, is essential to the research of
various scientists. Data acquisition is far from a thing of the past, but the
method by which we interpret and record information has improved
exponentially. These are just a few means by which we program equipment
and provide data for people all over the world.
Manufacturing
Manufacturing equipment can be fitted with IoT devices for sensing,
identification, communication, actuation monitoring, processing, and
networking, providing seamless integration between the manufacturing and
control systems. This integration has permitted the creation of brand-new
business and marketing options in manufacturing. This network control and
management of manufacturing process control, asset and situation
management, and manufacturing equipment set the IoT in the industrial
application and smart manufacturing space, allowing features such as[7]:
Rapid manufacturing of new products.
Real-time product and supply chain optimization by networking
machine sensors and their control systems together.
Allowing dynamic responses to product or market demand
The above is achieved by networking manufacturing machinery, control
systems, and sensors together.
The IoT can be applied to employ digital control systems for automating
process controls, manage service information systems in the optimization of
plant safety and security, and maintain and control operator tools.
Machine Learning, when integrated into these networked systems, allows
asset management and surveillance to maximize reliability by employing
predictive maintenance and statistical evaluation. Real-time energy
optimization is possible through the integration of the various sensor systems.
Machine Learning allows IoT to maximize reliability through asset
management using predictive maintenance, statistical evaluation, and
measurements.
The manufacturing industry has its own version of the Internet of Things, the
industrial Internet of Things (IIoT). The estimated effect IIoT could have on
the manufacturing sector is truly astounding. Growth due to the integration of
the Internet of Things with manufacturing is estimated at a $12 trillion dollar
increase in global GDP by 2030.
Data acquisition and device connectivity are imperative for IIoT. This is not
the goal, but rather a necessary condition for developing something much
greater. At the core of this will be interfacing Machine Learning algorithms
with local IIoT devices. This has already been done in some instances with
predictive maintenance, and this success will hopefully pave the way for new
and much more powerful ways to allow these intelligent maintenance
systems to reduce unforeseen downtime and provide increased productivity.
This term refers to the interface between human beings and the cyber world,
and this system will eventually allow data collected from it into actionable
information, decreasing costs, increasing safety, and boosting efficiency. As
with so many implementations of Machine Learning in our world, it would
seem the only limit to the advantages it can provide in the manufacturing
sector relies on the limits of our imagination to take advantage of it.
A real-world example of this integrated approach was conducted by the
University of Cincinnati in 2014. Industrial band-saws have belts that
degrade and eventually, if not caught by routine inspection, break, presenting
a hazard to the operators. In this experiment, a feedback system was created
in which Industrial Internet of Things sensors monitored the status of the belt
and returned this data to a predictive system, which could alert users when
the band was likely to fail and when the best time to replace it would be. This
system showed it was possible to save on costs, increase operator safety, and
improve the user experience by integrating IIoT devices with Machine
Learning algorithms.
Agriculture
The IoT allows farmers to collect data on all sorts of parameters, things like
temperature, humidity, wind speed/direction, temperature, soil composition,
and even pest infestations.
The collection of this data when combined with Machine Learning
algorithms, allows farmers to automate some farm techniques such as making
informed decisions to improve crop quality and quantity, minimize risk,
reduce required crop management efforts, and minimize waste. Farmers can
monitor soil temperature and moisture and use this data to decide on an
optimum time to water or fertilize fields. In the future, Machine Learning can
use historical and current weather data in combination with data returned
from the Internet of Things devices embedded in the farm to further
maximize these decision-making processes.
An example of a real-world combination of Machine Learning and IoT in
agriculture took place in 2018, when Toyota Tsusho (member of the Toyota
group which is parent to, among many other things, Toyota Motors)
partnered with Microsoft and created fish farming tools employing the
Microsoft Azure Internet of Things application suite for water management.
The water pump in this facility uses Machine Learning systems provided by
Azure to count fish on a conveyor belt and use this result to determine how
effective the water flow was in the system. This one system employed both
the Microsoft Azure IoT Hub and Microsoft Azure Machine Learning
platforms.
Infrastructure Applications

The Internet of Things can be used for control and monitoring of both urban
and rural infrastructure, things like bridges, railways, and off and on-shore
wind farms. The IoT infrastructure can be employed to monitor changes and
events in structural conditions that might threaten safety or increase risks to
users.
The construction industry can employ the Internet of Things and receive
benefits like an increase in productivity, cost savings, paperless workflows,
time reduction, and better-quality work days for employees, while real-time
monitoring and data analytics can save money and time by allowing faster
decision-making processes. Coupling the Internet of Things with Machine
Learning predictive solutions allows for more efficient scheduling for
maintenance and repair, as well as allowing for better coordination of tasks
between users and service providers. Internet of Things deployments can
even control critical infrastructure by allowing bridge access control to
approaching ships, saving time and money. Large-scale deployment of IoT
devices for the monitoring and operation of infrastructure will probably
improve the quality of service, increase uptime, reduce the costs of
operations, and increase incident management and emergency response
coordination. Even waste-management could benefit from IoT deployments,
allowing for the benefits of optimization and automation.
City Scale Uses
The Internet of Things has no upper limit to what it can encompass. Today,
buildings and vehicles are routine targets for the Internet of Things
integration, but tomorrow, it will be cities that are the wide-scale target of
this digital revolution. In many cities around the world, this integration is
beginning with important effects on utilities, transportation, and even law
enforcement. In South Korea, for example, an entire city is being constructed
as the first of its kind, a fully equipped and wired city. Much of the city will
be automated, requiring very little or even no human intervention. Songdo is
a nearly finished construction.
The city of Santander in Spain is taking a different approach. It does not have
the benefit of being built from scratch. Instead, it has produced an app that is
connected with over 10,000 sensors around the city. Things like parking
search, a digital city agenda, environmental monitoring, and more have been
integrated by the Internet of Things. Residents of the city download an app to
their smartphones in order to access this network.
In San Jose California, an Internet of Things deployment has been created
with several purposes — reducing noise pollution, improving air and water
quality, and increasing transportation efficiency. In 2014, the San Francisco
Bay Area partnered with Sigfox, a French company to deploy an ultra-narrow
band wireless data network, the first such business-backed installation in the
United States. They planned to roll out over 4000 more base stations to
provide coverage to more than 30 U.S. cities.
In New York, all the vessels belonging to the city were connected, allowing
24/7 monitoring. With this wireless network, NYWW can control its fleet and
passengers in a way that was simply not possible to imagine even a decade
ago. Possible new applications for the network might include energy and fleet
management, public Wi-Fi, security, paperless tickets, and digital signage
among others.
Most of the developments listed above will rely on Machine Learning as the
intelligent back-end to the forward-facing Internet of Things used to monitor
and update all the components of any smart city. It would seem the next step,
after the creation of smart cities through the integration of Internet of Things
sensors and monitors with Machine Learning algorithms, would be larger,
regional version.
Energy Management
So many electrical devices in our environment already boast an internet
connection. In the future, these connections will allow communication with
the utility from which they draw their power. The utility, in turn, will be able
to use this information to balance energy generation with load usage and to
optimize energy consumption over the grid. The Internet of Things will allow
users to remotely control (and schedule) lighting, HVAC, ovens, and so on.
On the utility’s side, aspects of the electrical grid, transformers, for example,
will be equipped with sensors and communications tools, which the utility
will use to manage the distribution of energy.
Once again, the power of Machine Learning will provide the predictive
models necessary to anticipate and balance loads, aided by the constant flow
of information from the Internet of Things devices located throughout the
power grid. At the same time, these IoT monitors will also provide service
levels and other information, so artificial intelligence systems can track and
identify when components are reaching the end of life and need repair or
replacement. No more surprise power outages, as components getting close to
failure will be identified and repaired before they can threaten the power grid.
Environmental Monitoring
IoT can be used for environmental monitoring and usually, that application
means environmental protection. Sensors can monitor air and water quality,
soil and atmospheric conditions, and can even be deployed to monitor the
movement of wildlife through their habitats. This use of the Internet of
Things over large geographic areas means it can be used by tsunami or
earthquake early-warning systems, which can aid emergency response
services in providing more localized and effective aid. These IoT devices can
include a mobile component. A standardized environmental Internet of
Things will likely revolutionize environmental monitoring and protection.
Here, we can begin to see the geographic scalability of Machine Learning as
it interfaces with the Internet of Things. While the use in cities can be
impressive due to the number of sensors and devices deployed throughout
them, in rural or natural environmental settings, the density of IoT devices
drops sharply, yet these huge geographical areas can come under the scrutiny
of Machine Learning algorithms in ways yet to be conceived.
Trends in IoT

The trend for the Internet of Things is clear — explosive growth. There were
an estimated 8.4 billion IoT devices in 2017, a 31% increase from the year
before. By 2020, worldwide estimates are that there will be 30 billion
devices. At this point, the market value of these devices will be over $7
trillion. The amount of data these devices will produce is staggering. And
when all this data is collected and ingested by Machine Learning algorithms,
control, and understanding of so many aspects of our lives, our cities, and
even our wildlife will increase dramatically.
Intelligence
Ambient Intelligence is an emerging discipline in which our environment
becomes sensitive to us. It was not meant to be part of the Internet of Things,
but the trend seems to be that Ambient Intelligence and autonomous control
are two disciplines into which the Internet of Things will make great inroads.
Already research in these disciplines is shifting to incorporate the power and
ubiquity of the IoT.
Unlike today, the future Internet of Things might be comprised of a good
number of non-deterministic devices, accessing an open network where they
can auto-organize. That is to say, a particular IoT device may have a purpose,
but when required by other Internet of Things devices, it can become part of a
group or swarm to accomplish a collective task that overrides its individual
task in priority. Because these devices will be more generalized with a suite
of sensors and abilities, these collectives will be able to organize themselves
using the particular skills of each individual to accomplish a task, before
releasing its members when the task is complete so they can continue on with
their original purpose. These autonomous groups can form and disperse
depending on the priority given to the task, and considering the
circumstances, context, and the environment in which the situation is taking
place.
Such collective action will require coordination and intelligence, which of
course will need to be provided, at least in part, by powerful Machine
Learning algorithms tasked with achieving these large-scale objectives. Even
if the Internet of Things devices that are conscripted to achieve a larger goal
have their own levels of intelligence, it will be assisted by artificial
intelligence systems able to collate all the incoming data from the IoT devices
responding to the goal and finding the best approach to solve any problems.
Although swarm intelligence technology will also likely play a roll (see later
in the book in the chapter on The Swarm), swarms cannot provide the
predictive ability that crunching massive data sets by Machine Learning
backed by powerful servers have.
Architecture
The system architecture of IoT devices is typically broken down into three
main tiers. Tier one includes devices, Tier two is the Edge Gateway, and Tier
three is the Cloud.
Devices are simply network-connected things with sensors and transmitters
using a communication protocol (open or proprietary) to connect with an
Edge Gateway.
The Edge Gateway is the point where incoming sensor data is aggregated and
processed through systems to provide functionality including data pre-
processioning, securing cloud connections, and running WebSockets or even
edge analytics.
Finally, Tier three includes the applications based in the cloud built for the
Internet of Things devices. This tier includes database storage systems to
archived sensor data. The cloud system handles communication taking place
in all tiers and provides features like event queuing.
As data is passed from the Internet of Things, a new architecture emerges, the
web of things, the application layer which processes the disparate sensor data
from the Internet of Things devices into web applications, driving innovation
in use-cases.
It is no great leap to assume this architecture does, and will even more in the
future, require massive scalability in the network. One solution being
deployed is fog (or edge) computing, where the edge network, tier two above,
takes over many of the analytical and computational tasks, sorting the
incoming data and only providing vetted content to the tier three cloud layer.
This reduces both latency and bandwidth requirements in the network. At the
same time, if communications are temporarily broken between the cloud and
edge layers, the edge layer’s fog computing capabilities can take over to run
the tier one devices until communications can be re-established.
Complexity and Size
Today’s Internet of Things is more a collection of proprietary devices,
protocols, languages, and interfaces. In the future, as there is more
standardization deployed, this collection will be studied as a complex system,
given its fast numbers of communication links, autonomous actors, and its
capacity to constantly integrate new actors. Today, however, most elements
of the Internet of Things do not belong to any massive group of devices.
Devices in a smart home, for example, do not communicate with anything
other than the central hub of the home. These subsystems are there for
reasons of privacy, reliability, and control. Only the central hub can be
accessed or controlled from outside the home, meaning all sensor data,
positions, and the status of the devices inside are not shared. Until such time
as there is an assurance of security from more global networks about access
and control of these private networks of IoT devices, and at the same time,
until non-standard architectures are developed to allow the command and
control of such vast numbers of internet connected devices, the IoT will
likely continue as a collection of non-communicating mini-networks. SND,
Software Defined Networking, appears promising in this area as a solution
that can deal with the diversity and unique requirements of so many IoT
applications.
For Machine Learning to take advantage of all this sensor data, ways will
need to be devised to collect it into training and data sets that can be iterated
over, allowing the software to learn about and be able to predict important
future states of the IoT network. For the time being, unfortunately, most IoT
sensor data produced in these private networks is simply being lost.
Space
Today, the question of an Internet of Thing’s location in space and time has
not been given priority. Smartphones, for example, allow the user to opt out
of allowing an application gaining access to the phone’s capacity for
geolocation. In the future, the precise location, position, size, speed, and
direction of every IoT device on the network will be critical. In fact, for IoT
devices to be interconnected and provide coordinated activity, the current
issues of variable spatial scales, indexing for fast search and close neighbor
operations, and indeed just the massive amount of data these devices will
provide all have to be overcome. If the Internet of Things is to become
autonomous from human-centric decision-making, then in addition to
Machine Learning algorithms being used to command and coordinate them,
space and time dimensions of IoT devices will need to be addressed and
standardized in a way similar to how the internet and Web have been
standardized.
Jean-Louis Gassée and a “basket of remotes”
Gassée worked at Apple from 1981 to 1990 and after that was one of the
founders of Be Incorporated, creator of the BeOS operating system. He
describes the issue of various protocols among Internet of Things providers
as the “basket of remotes” problem. As wireless remote controls gained
popularity, consumers would find they ended up with a basket full of remotes
— one for the TV, one for the VCR, one for the DVD player, one for the
stereo, and possibly even a universal remote that was supposed to replace all
the others but often failed to match at least one device. All these devices used
the same technology, but because each manufacturer used a proprietary
language and/or frequency to communicate with their appliance, there was no
way to easily get one category of device to speak the same language as the
others. In the same way, what Gassée sees happening is, as each new
deployment of some Internet of Things tech reaches the consumer market, it
becomes its own remote and base station, unable to communicate with other
remotes or with the other base station to which other groups of IoT tech can
communicate. If this state of proprietary bubbles of the Internet of Things
collections is not overcome, many of the perceived benefits of a vast IoT
network will never be realized.
Security
Just as security is one of the main concerns for the conventional internet,
security of the Internet of Things is a much-discussed topic. Many people are
concerned that the industry is developing too rapidly and without an
appropriate discussion about the security issues involved in these devices and
their networks. The Internet of Things, in addition to standard security
concerns found on the internet, has unique challenges — security controls in
industry, Internet of Things businesses processes, hybrid systems, and end
nodes.
Security is likely the main concern over adopting Internet of Things tech.
Cyber-attacks on components of this industry are likely to increase as the
scale of IoT adoption increases. And these threats are likely to become
physical, as opposed to merely a virtual threat. Current Internet of Things
systems have many security vulnerabilities including a lack of encrypted
communication between devices, weak authentication (many devices are
allowed to run in production environments with default credentials), lack of
verification or encryption of software updates, and even SQL injection. These
threats provide bad actors the ability to easily steal user credentials, intercept
data, and collect Personally Identifiable Information, or even inject malware
into updated firmware.
Many internet-connected devices are already spying on people in their own
homes, including kitchen appliances, thermostats, cameras, and televisions.
Many components of modern cars are susceptible to manipulation should a
bad actor gain access to the vehicle’s onboard systems including dashboard
displays, the horn, heating/cooling, hood and trunk releases, the engine, door
locks, and even braking. Those vehicles with wireless connectivity are
vulnerable to wireless remote attacks. Demonstration attacks on other
internet-connected devices have been made as well, including cracking
insulin pumps, implantable cardioverter defibrillators, and pacemakers.
Because some of these devices have severe limitations on their size and
processing power, they may be unable to use standard security measures like
strong encryption for communication or even employing firewalls.
Privacy concerns over the Internet of Things have two main thrusts —
legitimate and illegitimate uses. In legitimate uses, governments and large
corporations may set up massive IoT services which by their nature collect
enormous amounts of data. To a private entity, this data can be monetized in
many different ways, with little or no recourse for the people whose lives and
activities are swept up in the data collection. For governments, massive data
collection from the Internet of Things networks provide the data necessary to
provide services and infrastructure, to save resources and reduce emission,
and so on. At the same time, these systems will collect enormous amounts of
data on citizens, including their locations, activities, shopping habits, travel,
and so on. To some, this is the realization of a true surveillance state. Without
a legal framework in place to prevent governments from simply scooping up
endless amounts of data to do with as they wish, it is difficult to refute this
argument.
Illegitimate uses of the massive Internet of Things networks include
everything from DDOS (distributed denial of service) attacks through
malware attacks on one or more of the IoT devices on the network. Even
more worrying, security vulnerabilities in even one device on an Internet of
Things network can, by nature of the fact it is capable of full communication
with all devices nearby because it has access to the encryption requirements
to present itself as a legitimate device on the network, means an infected
device may not only provide its illegitimate host with the data it provides, but
metadata of other devices in the network, and possibly even access to the
edge systems themselves.
In 2016, a DDOS attack powered by an Internet of Things device
compromised by a malware infection led to over 300,000 device infections
and brought down both a DNS provider and several major websites. This
Mirai Botnet was able to single out for attack devices that consisted mostly of
IP cameras, DVRs, printers, and routers.
While there are several initiatives being made to increase security in the
Internet of Things marketplace, many argue that government regulation and
inter-governmental cooperation around the world must be in place to ensure
public safety.
Chapter 14: Machine Learning and Robotics

First, we need to define a robot. They are machines, often programmable by a

computer, that are able to carry out a series of complex actions without
intervention. A robot can have its controls systems embedded or be
controlled by an external device. In popular literature and film, many robots
are designed to look like people, but in fact, they are usually designed to
perform a task, and that requirement determines how they appear.
Robots have been with us for almost as long as computers. George Devol
invented the first digitally operated and programmable one in 1954, called
Unimate. In 1961, General Motors bought it to use for lifting hot metal die
castings. And like computers, robots have changed our society. Their
strength, agility, and ability to continue to execute the same repetitive tasks
perfectly have proved an enormous benefit. And while they did cause some
serious disruption to the manufacturing industries, putting many people out of
work, their ascension in our societies has provided far more employment
opportunities than they have taken.
Robots in current use can be broken down into several categories:
Industrial Robots/Service Robots
These robots are probably familiar. You have likely seen them on the
television or streaming video of automated factories. They usually consist of
an “arm” with one or more joints, which ends with a gripper or manipulating
device. They first took hold in automobile factories. They are fixed in one
location and are unable to move about. Industrial robots will be found in
manufacturing and industrial locations. Service robots are basically the same
in design as industrial robots but are found outside of manufacturing
concerns.
Educational Robots
These are robots used as teacher aids or for educational purposes on their
own. As early as the 1980s, robots were introduced in classrooms with the
turtles, which were used in classrooms where students could train them using
the Lego programming language. There are also robot kits available for
purchase like the Lego Mindstorm.
Modular Robots
Modular robots consist of several independent units that work together. They
can be identical or have one or more variations in design. Modular robots are
able to attach together to form shapes that allow them to perform tasks. The
programming of modular robotic systems is of course, more complex than a
single robot, but ongoing research in many universities and corporate settings
is proving that this design approach is superior to single large robots for
many types of applications. When combined with Swarm Intelligence (see
below), modular robots are proving quite adept at creative problem-solving.
Collaborative Robots
Collaborative robots are designed to work with human beings. They are
mostly industrial robots that include safety features to ensure they do not
harm anyone as they go about their assigned tasks. An excellent example of
this kind of collaborative robot is Baxter. Introduced in 2012, Baxter is an
industrial robot designed to be programmed to accomplish simple tasks but is
able to sense when it comes into contact with a human being and stops
moving.
Of course, all the examples above do not require any artificial intelligence.
When robots are coupled with machine learning, researchers use the term
“Robotic Learning.” This field has a contemporary impact in at least four
important areas:
Vision
Machine Learning has allowed robots to sense their environment visually and
to make sense of what they are seeing. New items can be understood and
classified without the need to program into the robot ahead of time what it is
looking at.
Grasping
Coupled with vision, Machine Learning allows robots to manipulate items in
their environment that they have never seen before. In the past, in an
industrial factory, each time a robot was expected to interact with a different-
shaped object, it would have to be programmed to know how to manipulate
this new object before it could be put to work. With Machine Learning, the
robot comes equipped with the ability to navigate new item shapes and sizes
automatically.
Motion Control
With the aid of Machine Learning, robots are able to move about their
environments and avoid obstacles in order to continue their assigned tasks.
Data
Robots are now able to understand patterns in data, both physical and
logistical, and act accordingly on those patterns.
Examples of Industrial Robots and Machine
Learning

One example of the benefit of applying Machine Learning to robots is of an

industrial robot which receives boxes of frozen food along a conveyor.
Because it is frozen, these boxes often have frost, sometimes quite a lot of
frost. This actually changes the shape of the box randomly. Thus, a
traditionally-trained robot with very little tolerance for these shape changes
would fail to grasp the boxes correctly. With Machine Learning algorithms,
the robot is now able to adapt to different shapes, random as they are and in
real time, and successfully grasp the boxes.
Another industrial example includes a factory with over 90,000 different
parts. It would not be possible to teach a robot how to manipulate these many
items. With Machine Learning, the robot is able to be fed images of new
parts it will be dealing with, and it can determine its own method to
manipulate them.
In 2019, there will be an estimated 2.6 million robots in service on the planet.
That’s up a million from 2015. As more and more robots are combined with
Machine Learning algorithms, this number is sure to explode.
Neural Networks with Scikit-learn

Neural networks are a machine learning framework that tries to mimic the
way the natural biological neural networks operate. Humans have the
capacity of identifying patterns with a very high degree of accuracy. Anytime
you see a cow, you can immediately recognize that it is a cow. This also
applies to when you see a goat. The reason is that you have learned over a
period of time how a cow or a goat looks like and what differentiates between
the two.
Artificial neural networks refer to computation systems that try to imitate the
capabilities of human learning via a complex architecture that resembles the
nervous system of a human being.
Chapter 15: Machine Learning and Swarm

Intelligence
Swarm Intelligence (SI) is defined as collaborative behavior, natural or
artificial, of decentralized, self-organized systems. That is, Swarm
Intelligence can refer to an ant colony or a “swarm” of autonomous mini-
drones in a lab.
In artificial intelligence, a swarm is typically a collection of agents that
interact with each other and their environment. The inspiration for Swarm
Intelligence comes from nature, from the collaboration of bees to the flocking
of birds to the motions of herd animals, groups of animals that appear to act
intelligently even when no single individual has exceptional intelligence, and
there is no centralized decision-making process.
Swarm Behavior

One of the central tenets gleaned from swarm research has been the notion of
emergent behavior. When a number of individuals are given simple rules for
complex behavior, some behaviors seem to arise despite there being no rule
or instruction to create them. Consider the artificial life program created by
Craig Reynolds in 1986, which simulates bird flocking. Each individual bird
was given a simple set of rules:
Avoid crowding local flockmates (separation).
Steer towards the heading average of local flockmates (alignment).
Steer to travel in the direction of the average center of local flockmates
(cohesion).
When he let the birds free, his experimental birds behaved like a real bird
flock. He discovered he could add more rules to make more complex flocking
behavior like goal seeking or obstacle avoidance.
Applications of Swarm Intelligence

Swarm Intelligence can be applied in many areas. Military applications

include research into techniques for unmanned vehicle control. NASA is
considering swarm tech for mapping planets. In a 1992 paper, George Bekey
and M. Anthony Lewis discussed using swarm intelligence in nanobots
inserted into the human body to attack and destroy cancer tumors.
Ant-based routing
In the telecommunication industry, the use of Swarm Intelligence has been
researched using a routing table where small control packets (ants) are
rewarded for successfully traversing a route. Variations on this research
include forwards, backward, and bi-directional rewards. Such systems are not
repeatable because they behave randomly, so commercial uses have thus far
proved elusive.
One promising application for Swarm Intelligence is wireless communication
networks. In this case, the network relies on a limited number of locations
that are expected to provide adequate coverage for users. In this case, the
application of a different kind of ant-based swarm intelligence, stochastic
diffusion search (SDS) has modeled this problem with great success.
Airlines have experimented with ant-based Swarm Intelligence. Southwest
Airlines uses software that employs swarm theory to manage its airlines on
the ground. Each pilot acts like an “ant” in the swarm, discovering through
experience what gate is best for him or her. This behavior turns out to be the
best for the airline as well. The pilot colony uses the gates each one is
familiar with and so can arrive at and leave from quickly, with the software
providing feedback should a particular gate or route be predicted to suffer a
back-up.
Crowd Simulation
The movies are using Swarm Intelligence simulations for depicting animal
and human crowds. In Batman Returns, Swarm Intelligence was employed to
create a realistic bat simulation. For The Lord of the Rings movies, Swarm
Intelligence simulations were employed to depict the massive battle scenes.
Human Swarms
When combined with mediating software, a network of distributed people can
be organized into swarms by implementing closed-loop, real-time control
systems. These systems, acting out in real-time, allow human actors to act in
a unified manner, a collective intelligence that operates like a single mind,
making predictions or answering questions. Testing in academic settings
suggests these human swarms can out-perform individuals in a wide variety
of real-world situations.
Swarm Intelligence and Machine Intelligence are both forms of artificial
intelligence. It’s debatable whether Swarm Intelligence is a sub-set of
Machine Intelligence. It is a different approach toward the goal of smart
machines, modeling the behavior of particular kinds of animals to achieve
desired results. But however they are classified, Swarm Intelligence and
Machine Intelligence can complement each other. In an attempt to determine
emotions from a text, a Swarm Intelligence approach will differ from a
monolithic approach. Instead of one Machine Learning algorithm to detect
emotion in text, a swarm approach might be to create many simple Machine
Learning algorithms, each designed to detect a single emotion. These
heuristics can be layered in hierarchies to avoid any emotion-detector fouling
up the end result. Let’s look at an example. Imagine a Machine Learning
swarm meant to detect emotion in written text examining this sentence:
“When I pulled back the covers of my bed this morning, a giant spider ran
over my leg, and I ran out of the bedroom screaming.”
This is a complex sentence and very difficult for natural language Machine
Learning algorithms to parse for emotion. However, a swarm of simple
Machine Learning algorithms dedicated to detecting one kind of emotion
would likely have the terror algorithm scoring high, while fun and happiness
scored low.
So far our imaginary system is working well. But take a more difficult
example:
“I watched the game yesterday and saw us kick butt in the first period, but by
the third, I was terrified we would lose.”
Human beings understand hyperbole. The writer is not terrified but anxious
the team will lose the game. Our swarm Machine Intelligence algorithms
could have the fear/terror algorithm scoring high, but this would be
inaccurate. Because swarm models can be hierarchical, one model’s output
could be the input of another. In this case, a master model that detects
emotion could filter through the outputs of each individual emotion
algorithm, noting that “excitement” had been triggered by “kick butt,” parse
that the subject of the sentence is a sport, and determine that anxiety is a
better fit than terror.
It seems fair to define Swarm Intelligence as an application of Machine
Learning with an extremely narrow focus.
Chapter 16: Machine Learning Models

There are many models of Machine Learning. These theoretical models

describe the heuristics used to accomplish the ideal, allowing the machines to
learn on their own.
Decision Tree
Just about everyone has used the decision tree technique. Either formally or
informally, we decide on a single course of action from many possibilities
based on previous experience. The possibilities look like branches, and we
take one of them and reject the others.
The decision tree model gets its name from the shape created when its
decision processes are drawn out graphically. A decision tree offers a great
deal of flexibility in terms of what input values it can receive. Also, a tree’s
outputs can take the form of a category, binary, or numerical. The strength of
decision trees is how the degree of influence of different input variables can
be determined by the level of decision node in which they are considered.
A weakness of decision trees is the fact that every decision boundary is a
forced binary split. There is no nuance. Each decision is either yes or no, one
or zero. Moreover, the decision criteria can consider only a single variable at
a time. There cannot be a combination of more than one input variable.
Decision trees cannot be updated incrementally. That is to say, once a tree
has been trained on a training set, it must be thrown out and a new one
created to tackle new training data.
Ensemble Methods address many tree limitations. In essence, the ensemble
method uses more than one tree to increase output accuracy. There are two
main ensemble methods — bagging and boosting.
The bagging ensemble method (known as Bootstrap Aggregation) is mean to
reduce decision tree variance. The training data is broken up randomly into
subsets, and each subset is used to train a decision tree. The results from all
trees are averaged, providing a more robust predictive accuracy than any
single tree on its own.
The boosting ensemble method resembles a multi-stage rocket. The main
booster of a rocket supplies the vehicle with a large amount of inertia. When
its fuel is spent, it detaches, and the second stage combines its acceleration to
the inertia already imparted to the rocket and so on. For decision trees, the
first tree operates on the training data and produces its output. The next tree
uses the earlier tree’s output as its input. When the input is in error, the
weighting it is given makes it more likely the next tree will identify and at
least partially mitigate this error. The end result of the run is a strong learner
emerging from a series of weaker learners.
Linear Regression
The premise of linear regression methods rests on the assumption that the
output (numeric value) may be expressed as a combination of the input
variable set (also numeric). A simple example might look like this:
x = a1y1, a2y2, a3y3…
Where x is the output, a1...an are the weights accorded to each input, and
y1...yn are the inputs.
A weakness of the linear regression model is the fact that it assumes linear
input features, which might not be the case. Inputs must be tested
mathematically for linearity.
K-Means Clustering Algorithm
K-means is an unsupervised machine learning algorithm for cluster analysis.
It is an iterative, non-deterministic method. The algorithm operates on data
sets using predefined clusters. You can think of clusters like categories. For
example, consider K-Means Clustering on a set of search results. The search
term “jaguar” returns all pages containing the word Jaguar. But the word
“Jaguar” has more than one classification. It can refer to a type of car, it can
refer to an operating system created by the Apple Computers, and it can refer
to the animal. K-Means clustering algorithms can be used to group results
that talk about similar concepts[8]. So, the algorithm will group all results that
discuss jaguar as an animal into one cluster, discussions of Jaguar as a car
into another cluster, and discussions of Jaguar as an operating system into a
third. And so on.
K-Means Clustering Applications
K-Means Clustering algorithms are used by most web search engines to
cluster web pages by similarity and to identify the relevance of search results.
But in any application where unstructured data needs to be divided into
meaningful categories, K-Means Clustering is a valuable tool to accomplish
this task.
Neural Network
We have covered neural networks in detail above. The strengths of neural
networks are their ability to learn non-linear relationships between inputs and
outputs.
Bayesian Network
Bayesian networks produce probabilistic relationships between outputs and
inputs. This type of network requires all data to be binary. The strengths of
the Bayesian network include high scalability and support for incremental
learning. We discussed Bayesian models in more detail earlier in the book. In
particular, this Machine Learning method is particularly good at classification
tasks such as detecting if an email is or is not spam.
Support Vector Machine
Support Vector Machine algorithms are supervised Machine Learning
algorithms that can be used for classification and regression analysis,
although they are most often used for classification problems, and so, we will
focus on those.
Support Vector Machines work by dividing categories using a hyperplane
(basically a flat sheet in 3 dimensions). On one side of the hyperplane will be
one category, on the other side the other category. This works well if the
categories to be separated are clearly divisible, but what if they overlap?
Imagine a series of blue and red dots spread out randomly within a square.
There is nowhere to place a hyperplane between the blue and red dots
because they overlap each other on the two-dimensional sheet. Support
Vector Machines deal with this problem by mapping the data into a higher
dimension. Imagine these dots being lifted into the air in a cube, where the
red dots are “heavier” than the blue dots, and therefore remain closer to the
bottom and the blue dots closer to the top. If there is still an overlap, the
Support Vector Machine algorithm maps them into yet another dimension,
until a clear separation appears and the two categories of dots can be
separated by the hyperplane.
This is an incredibly powerful method for classifying data, but there are
issues, of course. One is the fact that looking at the data once it has been
mapped into higher dimensions isn’t possible. It has become gibberish.
Support Vector Machines are sometimes referred to as black boxes for this
reason. Also, Support Vector Machine training times can be high, so this
algorithm isn’t as well suited to very large data sets, as the larger the data set,
the longer the training required.
Chapter 17: Applications of Machine Learning

You are probably using real-world applications of Machine Learning every

day. Some are easily recognized, others are hidden and operating in the
background, making our lives easier, or more convenient, or even safer.
Though you might not recognize it, there are many technological
advancements in your home that work through programming and training to
become more accessible. Your devices’ abilities to learn to provide the basis
for modern technology.
Virtual Personal Assistants
Google Assistant, Apple’s Siri, Amazon’s Alexa, and Microsoft’s Cortana –
these are just a few of the virtual assistants we interact with almost every day.
These are probably the most popular and well-known examples of applied
Machine Learning. When you use these personal assistants, there is more
than one kind of Machine Learning working behind the scenes. First, you can
speak to them using ordinary language, and they answer back the same way.
Speech recognition is a Machine Learning skill these assistants use to
understand you. Then, of course, once they have recognized what you said
and understood what you are asking for, they usually use yet more Machine
Learning to search for answers to your questions. Finally, they respond with
answers and track your response, compare it to previous responses, and use
all this data to be more accurate and efficient for your future requests.
Commuting Predictions
If you’ve used Google or Apple maps, you’ve used Machine Learning to get
yourself to your destination. These apps store your location, direction, and
speed in a central location to lead you to your destination, providing turn
details and so on before you actually reach the turning point. At the same
time, they are aggregating the data from all of the users nearby who are using
their service and using all this information to predict and track traffic
congestion and to modify your route in real-time to avoid it. When there are
not enough people using the app at any given time, the system relies on
predictive Machine Learning from previous traffic data from different days at
the same location to anticipate the traffic if it is unable to collect in real-time.
Uber and Lyft also rely on Machine Learning to decide on the cost of a ride,
basing their decisions on current demand, as well as predictions concerning
rider demand at any given time. These services would become technically
impossible without Machine Learning crunching massive amounts of data,
both historical and in real-time.
Online Fraud Detection
Securing cyberspace is one of the many goals of Machine Learning. Tracking
online fraud is an important tool for achieving this goal. PayPal employs
Machine Learning in its efforts to prevent money laundering. Their software
can track the millions of transactions taking place and separate legitimate
from illegitimate transactions going on between buyers and sellers.
Data Security
Another form of security in cyberspace is combating malware. This is a
massive problem and getting worse. 1 in 13 web requests led to malware on
the internet in 2018, and that’s up 3% from 2016. The problem is so vast only
something as powerful as Machine Learning is able to cope. Fortunately,
most new malware is really just older malware with minor changes. This is
something Machine Learning can search for quickly and with a high degree
of accuracy, meaning new malware is caught almost as quickly as it can be
created.
Machine Learning is also very good at monitoring data access and finding
anomalies, which might predict possible security breaches.
Personal Security
Do you face long lines at the airport? Does it take forever to get into that
concert? Those long lines are for your security. Everyone getting on a plane
or entering a large event is screened to ensure the safety of the flight or event.
Machine Learning is starting to provide assistance to the manual checking of
people, spotting anomalies human screenings might miss, and helping to
prevent false alarms. This help promises to speed up security screening in a
substantial way, while at the same time ensuring safer events through the
more powerful screening processes Machine Learning can provide.
Video Surveillance
How many movies have you watched where a video feed is interrupted with a
loop of a parade or cell, so the good guys (or bad guys) can move through the
monitored area without being seen? Well, Machine Learning may not be able
to defeat this tactic any better than the hapless security guards in the movies,
but it is able to take over a lot of the drudgery of monitoring live video feeds.
The benefit is Machine Learning never gets tired, or needs a break, or has its
attention wander, so nothing is ever missed. Machine learning can focus on
anomalous behavior like standing for a long time without moving, sleep on
benches, or stumbling, meaning human beings can be freed up to do what
they do much better, deal with these people as required.
Social Media Services
Social media platforms utilize many forms of Machine Learning. If you’ve
been using these platforms, then you’ve been using Machine Learning.
Facebook, for example, constantly offers you friend suggestions of “people
you may know.” This is a simple concept of learning through experience.
Facebook’s technology watches who your friends are (and who those friend’s
friends are), what profiles you visit, and how often, what articles you follow,
and pages you visit. By aggregating this information, Machine Learning is
able to provide people you are likely to enjoy interacting with and so
recommends them to you as Friend suggestions.
Or when you upload a picture to Facebook of yourself and some friends,
Facebook uses Machine Learning to identify those people in your photo by
comparing them to images your friends have uploaded of themselves. But
this is not all. These photos are scanned for the poses people are making, as
well as any unique features in the background, Geo-locating the image if it
can.
Online Customer Support
Many websites offer a service to text chat with customer service while you
browse their pages. You’ve probably chatted with one or more of them. But
what you might not know is, not all of those text chats are with real people.
Many companies cannot afford to pay a person to provide this service, so
instead a chatbot will answer your questions, itself relying on details from the
website to provide its answers. Machine Learning means these chatbots get
better over time, learning how and how not to answer questions, and to seem
more human in their interactions.
Search Engine Results, Refined
Search engine providers rely on Machine Learning to refine their responses to
search terms. They monitor your activity after you’ve been given some
results to examine. Do you click on the first link and stay on this page for a
long time? Then their results to your query were probably good. Or do you
navigate to page 2 or 3 or further down the results they have provided,
clicking some links and immediately returning? This means the results
returned did not satisfy your query as well as they could. Machine Learning
uses this feedback to improve the search algorithms used.
Product Recommendations/Market Personalization
Have you ever been shopping online for something you didn’t intend to buy?
Or were you considering buying it, but not for a while? So you look online at
a product, maybe at the same product in more than one online store,
comparing prices, customer reviews, and so on. You might even put one of
them in your shopping cart to checkout taxes and shipping fees. Then, you
close these tabs and get back to work.
Over the next few days, you realize you see this product and similar products
all over web pages you visit. Advertisements for the product seem to be
everywhere for the next few days.
Using the kind of information gathered from your shopping/browsing
practices, Machine Learning is used to tailor everything from specific email
campaigns aimed at you personally, to similar product advertisements, to
coupons or deals by the retailers offering the product you’d been looking at.
As the amount of data for these Machine Learning algorithms grows, expect
this personal approach to advertising to become even more focused and
accurate.
Financial Trading
Being able to “predict the market” is something of a holy grail in financial
circles. The truth is humans have never been much better at this than chance.
However, Machine Learning is bringing the grail within the human grasp.
Using the power of massive supercomputers crunching enormous quantities
of data, large financial firms with proprietary predictive systems make a large
volume of high-speed transactions. With high speed and enough volume,
even low-probability trades can end up making these firms enormous
amounts of money. What about those of us who can’t afford massive
supercomputers? We’re better off doing our stock trading the old-fashioned
way — research specific companies and invest in businesses with solid
business plans and room in the market for growth. No doubt one day soon,
there will be a Machine Learning algorithm to manage this for you.
Healthcare
Machine Learning is revolutionizing the health care system. With its ability
to ingest enormous amounts of data and use that to spot patterns human
simply cannot see, Machine Learning is able to diagnose some diseases up to
a year before a human diagnosis. At the same time, by crunching large data
sets of populations, Machine Learning is able to identify groups or
individuals who are more likely to need hospitalization due to diseases like
diabetes. The predictive power of Machine Learning is likely the most fruitful
avenue in healthcare since much of disease-fighting rests on early diagnosis.
Another inroad to healthcare by Machine Learning is robotic assistants in the
operating room. While many of these robots are simply tools used by doctors,
some are semi-autonomous and aid in the surgery themselves. In the not too
distant future, Machine Learning will allow robot surgeons to perform
operations with complete autonomy. This will free up surgeons to perform
the far more complex and experimental procedures that Machine Learning is
not prepared to perform.
Natural Language Processing
Being able to understand natural human speech has been anticipated for a
long time. We’ve been talking to our robots in movies and cartoons for more
than 70 years. And now, thanks to Machine Learning, that ability is being
realized all around us. We can talk to our smartphones, and we can talk to a
company’s “customer support” robot and tell it what we need help with.
Natural language processing is even able to take complex legal contract
language and translate it into plain language for non-lawyers to understand.
Smart Cars
The autonomous car is very close. But it won’t just be autonomous; it will be
smart. It will interconnect with other cars and other internet-enabled devices,
as well as learn about its owner and passengers, providing internal
adjustments to the temperature, audio, seat settings, and so on. Because
autonomous cars will communicate with each other as they become more
numerous on the road, they will become safer. Accidents will drop to next to
zero when people are no longer behind the wheel. This is the power of
Machine Learning really flexing its muscles, and only one of the many, many
ways it will change our human future.
Chapter 18: Programming and (Free) Datasets

For those interested in the world of programming in Machine Learning, there

are many, many resources on the web. Programming for small devices, like
Arduino, are also available. Leaning programming on these machines
provides easier adaptation to other devices. This chapter will identify and
describe some excellent free data sets available for download and use in
Machine Learning programs, as well as suggest an option for learning
beginning level Machine Learning software development using the Python
programming language.
Limitations of Machine Learning

There are many problems and limitations with Machine Learning. This
chapter will go over the technical issues that are currently or may in the
future limit the development of Machine Learning. Finally, it will end with
some philosophical concerns about possible issues Machine Learning may
bring about in the future.
Concerns about Machine Learning limitations have been summarized in a
simple phrase, which outlines the main objects to Machine Learning. It
suggests Machine Learning is greedy, brittle, opaque and shallow. Let’s
examine each one in turn.
Greedy
By calling it greedy, critics of Machine Learning point to the need for
massive amounts of training data to be available in order to successfully train
Machine Learning systems to acceptable levels of error. Because Machine
Learning systems are trained not programmed, their usefulness will be
directly proportional to the amount (and quality) of the data sets used to train
them.
Related to the size of training data required is the fact that, for supervised and
semi-supervised Machine Learning training, the raw data used for training
must first be labeled so that it can be meaningfully employed by the software
to train. In essence, the task of labeling training data means to clean up the
raw content and prepare it for the software to ingest. But labeling data can
itself be a very complex task, as well as often a laborious and tedious one.
None or improperly labeled data fed into a supervised Machine Learning
system will produce nothing of value. It will just be a waste of time.
Brittle
To say Machine Learning is brittle is to highlight a very real and difficult
problem in Artificial Intelligence. The problem is, even after a Machine
Learning system has been trained to provide extremely accurate and valuable
predictions on data it has been trained to deal with, asking that trained system
to examine a data set even slightly different from the type it was trained on
will often cause a complete failure of the system to produce any predictive
value. That is to say, Machine Learning systems are unable to contextualize
what they have learned and apply it to even extremely similar circumstances
to those on which they have been trained. At the same time, attempting to
train an already trained Machine Learning algorithm with a different data set
will cause the system to “forget” its previous learning, losing, in turn, all the
time and effort put in to preparing the previous data sets and training the
algorithm with it.
Bias in Machine Learning systems is another example of how Machine
Learning systems can be brittle. In fact, there are several different kinds of
bias that threaten Artificial Intelligence. Here are a few:
Bias in Data:
Machine Learning is, for the foreseeable future at least, at the mercy of the
data used to train it. If this data is biased in any way, whether deliberately or
by accident, those biases hidden within it may be passed onto the Machine
Learning system itself. If not caught during the training stage, this bias can
taint the work of the system when it is out in the real world doing what it was
designed to do. Facial recognition provides a good example, where facial
recognition systems trained in predominantly white environments with
predominantly white samples of faces, have trouble recognizing people with
non-white facial pigmentation.
Acquired Bias:
It is sometimes the case that, while interacting with people in the real world,
Machine Learning systems can acquire the biases of the people they are
exposed to. A real-world example was Microsoft’s Tay, a chatbot designed to
interact with people on Twitter using natural language. Within 24 hours, Tay
became a pretty foul troll spouting racist and misogynist statements.
Microsoft pulled the plug (and set about white-washing twitter). Many of the
truly offensive things Tay said were the result of people telling it to say them
(there was an option to tell Tay to repeat a phrase you sent it), but there were
some very troubling comments produced by Tay that it was not instructed to
repeat. Microsoft was clearly aware of how nasty Twitter can get, and I think
it’s fair to say creating a racist, misogynist chatbot was just about the last
thing on their whiteboard. So if a massive, wealthy company like Microsoft
cannot train an artificial intelligence that doesn’t jump into the racist, woman-
hating camp of nasty internet trolls, what does that say about the dangers
inherent in any artificial intelligence system we create to interact in the real
world with real people?
Emergent Bias:
An echo chamber is a group or collection of people who all believe the same
things. This could be a political meeting in a basement or a chat room on the
internet. Echo chambers are not tolerant of outside ideas, especially those that
disagree with or attempt to refute the group’s core beliefs. Facebook has
become, in many ways, the preeminent echo chamber producer on Earth. But
while the echo chamber phrase is meant to describe a group of people with
similar ideas, Facebook’s artificial intelligence has taken this idea one step
further: to an echo chamber of one. The Machine Learning system Facebook
deploys to gather newsfeeds and other interesting bits of information to show
to you can become just such an echo chamber. As the artificial intelligence
learns about your likes and dislikes, about your interests and activities, it
begins to create a bubble around you that, while it might seem comforting,
will not allow opposing or offensive views to reach you. The area where this
is most alarming is news and information. Being constantly surrounded by
people who agree with you, reading the news that only confirms your beliefs,
is not necessarily a good thing. How we know we are correct about a
particular issue is by testing our beliefs against who believe otherwise. Do
our arguments hold up to their criticism or might we be wrong? In a
Facebook echo chamber of one, that kind of learning and growth becomes
less and less possible. Some studies suggest that spending time with people
who agree with you tends to polarize groups, making the divisions between
them worse.
Goals Conflict Bias:
Machine Learning can often support and reinforce biases that exist in the real
world because doing so increases their reward system (a system is
“rewarded” when it achieves a goal). For example, imagine you run a college,
and you want to increase enrollment. The contract you have with the
advertising company is that you will pay a certain amount for each click you
get, meaning someone has seen your advertisement for the college and was
interested enough to at least click on the link to see what you are offering. Of
course, a simple generic ad like “Go to College!” wouldn’t work so well, so
you run several ads simultaneously, offering degrees in engineering,
teaching, mathematics, and nursing.
It is in the advertiser’s best interest to get as many clicks to your landing
pages as possible. So they employ a Machine Learning algorithm to track
who clicks on what advertisement, and it begins to target those groups with
the specific ads to gain more clicks. So far, from the outside, this seems like a
win-win. The advertiser is making more revenue, and your college is
receiving more potential students examining your courses. But then you
notice something in the aggregate data of link clicks. Most of the clicks you
are receiving for math and engineering are from young men, while most of
the clicks for nursing and teaching are from young women. This aligns with
an unfortunate cultural stereotype that still exists in western culture, and your
college is perpetuating it. In its desire to be rewarded, the Machine Learning
assigned maximizing clicks for your advertising campaign found an existing
social bias and exploited it to increase revenue for its owner. These two
goals, increasing your enrollment and reducing gender bias in employment,
have come into conflict and Machine Learning sacrificed one to achieve the
other.
Opaque
One of the main criticisms of Machine Learning, and in particular, against
Neural Networks, is that they are unable to explain to their creators why they
arrive at the decisions they do. This is a problem for two reasons: one, more
and more countries are adopting internet laws that include a right to an
explanation. The most influential of these laws is the GDPR (The EU General
Data Protection Regulation), which guarantees EU citizens the right to an
explanation why an algorithm that deals with an important part of their lives
made the decision it did. For example, an EU citizen turned down for a loan
by a Machine Learning algorithm has a right to demand why this happened.
Because some artificial intelligence tools like neural networks are often not
capable of providing any explanation for their decision, and the fact this
decision is hidden in layers of math not readily available for human
examination, such an explanation may not be possible. This has, in fact,
slowed down the adoption of artificial intelligence in some areas. The second
reason it is important that artificial intelligence be able to explain its
decisions to its creators is for verifying that the underlying process is in fact
meeting expectations in the real world. For Machine Learning, the decision-
making process is mathematical and probabilistic. In the real world, decisions
often need to be confirmed by the reasoning used to achieve them.
Take, for example, a self-driving car involved in an accident. Assuming the
hardware of the car is not completely destroyed, experts will want to know
why the car took the actions it did. Perhaps there is a court case, and the
decision of liability rests on how and why the car took the actions it did.
Without transparency around the decision-making process of the software,
justice might not be served in the courts, software engineers may not be able
to find, and correct flaws and more people might be in danger from the same
software running in different cars.
The Philosophical Objections: Jobs, Evil, and
Taking Over the World

Jobs:
One of the main concerns surrounding artificial intelligence is the way these
systems are encroaching upon human employment. Automation is not a new
problem, and vast numbers of jobs have already been lost in manufacturing
and other such industries to robots and automated systems. Because of this,
some argue that this concern that machines and artificial intelligence will take
over so many jobs there will be no more economy is not really a threat.
We’ve encountered such takeovers before, but the economy shifted, and
people found other jobs. In fact, some argue that the net number of jobs has
increased since the loss of so many jobs to automation. So what’s the
problem?
The problem is this next round of job losses will be unprecedented. Artificial
Intelligence will not only replace drudgery and danger. It will keep going.
There are 3.5 million of them currently. How long until all trucking is
handled by Machine Learning systems running self-driving trucks? And what
about ride-sharing services like Uber and Lyft?
The question to ask is not what jobs will be replaced, but what jobs can’t be
replaced. In the Machine Learning round of employment disruption, white
collar jobs are just as much in jeopardy as blue. What about accountants,
financial advisers, copywriters, and advertisers?
The pro-artificial intelligence camp argues that this disruption will open up
new kinds of jobs, things we can’t even imagine. Each major disruption of
the economy in the past that displaced many forms of employment with
automation often caused the creation of new employment unforeseen before
the disruption. The difference with artificial intelligence is there is no reason
to believe these new forms of employment won’t quickly be taken over by
Machine Learning systems as well.
And all the above assumes the use of the current type of very specific
Machine Learning. What happens if researchers are able to generalize
learning in artificial intelligence? What if these systems become able to
generalize what they learn and apply what they know in new and different
contexts? Such a generalized artificial intelligence system could very quickly
learn to do just about anything.
Evil:
A very serious danger from artificial intelligence is the fact that anyone with
resources can acquire and use it. Western governments are already toying
with the idea of autonomous weapons platforms using artificial intelligence
to engage and defeat an enemy, with little or no human oversight. As
frightening as this might be, in these countries at least, there are checks and
balances on the development and deployment of such devices, and in the end,
the population can vote for or against such things.
But even if these platforms are developed and only used appropriately, what
is to stop an enemy from capturing one, reverse engineering it, and then
developing their own? What’s to stop a rival power from investing in the
infrastructure to create their own?
The threat of Machine Learning used by bad actors is very real. How do we
control who gains access to this powerful technology? The genie is out of the
bottle. It can’t be put back in. So, how do we keep it from falling into the
wrong hands?
Taking Over the World:
Finally, we’ll take a look at the end humanity might cause by creating super-
intelligent machines. Luminaries such as Steven Hawking and Elon Musk
have quite publicly voiced their concerns about the dangers artificial
intelligence could pose to humanity. In the Terminator movie franchise, a
Defense Department artificial intelligence called Skynet, tasked with
protecting the United States, determined that human beings, everywhere,
were the real threat. Using its control of the US nuclear arsenal, it launched
an attack on Russia, which precipitated a global nuclear war. In the disaster
that followed, it began systematically wiping out the human race.
This example is extreme and probably better as a movie plot than something
we have to worry about from artificial intelligence in the real world. No, the
danger to humanity from artificial intelligence lies more likely in a paperclip
factory.
Imagine an automated paperclip factory run by an artificial intelligence
system capable of learning over time. This system is smarter than human
beings. It has one goal to produce as many paperclips as possible, as
efficiently as possible. This sounds harmless enough. And at first, it is. The
system learns everything there is to know about creating a paperclip, from the
mining of metals, smelting, transportation of raw steel to its factory, the
automated factory floor, clip design specifications, and so on.
As it learns over time, however, it runs up against walls it cannot surpass. A
factory can only be so efficient. Eventually boosting efficiency further
becomes impossible. Acquiring the supply chain from mining to smelting to
transportation might come next. But again, once these aspects of the business
are acquired and streamlined, they no longer offer a path to the goal —
increase production and efficiency.
As artificial intelligence collects more information about the world, it comes
to learn how to take over other factories, retool them for paperclip
production, and increase output. Then more transportation, mining, and
smelting might need to be acquired. After a few more factories are retrofitted,
the distribution center where the paperclips are delivered begins to fill with
paperclips. There are many more than anyone needs. But this is not the
concern of our artificial intelligence. The supply side of the paperclip chain is
not part of its programming.
It learns about politics and so begins to influence elections when people try to
stop paperclip production. It continues to acquire and take over businesses
and technologies to increase production.
Absurd as it sounds (and this is just a thought experiment), imagine the entire
Earth, now devoid of people, a massive mining, transportation, smelting, and
paperclip production facility. And as the piles of unwanted, unused
paperclips turn into mountains to rival the stone ones of our planet, with
metal resources dwindling, our stalwart AI turns its lenses up to see the
moon, Mars, and the asteroid belt. It sees other solar systems, other galaxies,
and finally the universe. So much raw metal out there, all just waiting to be
made into paperclips…
Chapter 19: Machine Learning and the Future

After the previous chapter, it’s probably best to end on a positive note, by an
examination of the positive future of Machine Learning. This chapter will
break down the future of Machine Learning into segments of society and the
economy in an effort to put these possible futures in a useful context.

Security
Facial recognition and aberrant behavior detection, these are the toolkits from
Machine Learning that are available today. They will become ubiquitous in
the future, protecting the public from criminal behavior and getting people in
trouble the help they need.
But what about the other Machine Learning security features in the future?
In the cyber world, Machine Learning will grow and increase its influence in
identifying cyber-attacks, malicious software code, and unexpected
communication attempts. At the same time, dark hat software crackers are
also working on Machine Learning tools to aid them in breaking into
networks, accessing private data, and causing service disruptions. The future
of the cyber world will be an ongoing war between White and Black Hat
Machine Learning tools. Who will win? We hope the White Hat Machine
Learning algorithms will be victorious, but win or lose, the battle will slowly
move out of the hands of people and into the algorithms of Machine
Learning.
Another sweeping change to security we might see in the near future is
autonomous drones controlled by Machine Learning algorithms. Drones can
maintain constant aerial surveillance over entire cities at very little cost. With
advancements in image recognition, motion capture, video recognition, and
natural language processing, they will be able to communicate with people on
the street, respond to natural disasters and automobile accidents, ferry
medications needed in emergency situations where traditional service
vehicles cannot access, and find and rescue lost hikers by leading them to
safety, deliver them needed supplies, and alerting authorities of their GPS
location.
Markets
The rise of Machine Learning will generate completely new Artificial
Intelligence-based products and services. Entire industries will be created to
service this new software, as well as new products to be added to the Internet
of Things, including a new generation of robots complete with learning
algorithms and the ability to see, hear, and communicate with us using
natural language.
Retail
In the retail sector, we will see enhanced and more accurate personalization.
But instead of merely showing us what we want, Machine Learning will be
dedicated to showing us what we need. Perhaps we’ve been eating too much
fast food this week. A smart algorithm looking out for our well-being would
not throw more and more fast food ads in our face. Instead, reminders about
our health, coupons for gym memberships, or recipes for our favorite salads
might become part of the Artificial Intelligence toolkit for our notifications.
Machine Learning will help us to balance our lives in many ways by using
knowledge about general health, and our own medical records, to provide
information about not only what we want, but also what we might need.
Healthcare
Machine Learning will know almost everything about us and not only
through records on the internet. When we visit our doctor to complain about a
sore shoulder, Machine Learning might inform our GP about how we are
prone to slouch at work, possibly altering the doctor’s diagnosis from a
prescription for analgesics to a prescription for exercises to do at work, as
well as some tutoring on better sitting posture.
On the diagnostic side, Machine Learning will do the busy work it is best at:
examining our x-rays and blood-work and mammograms, looking for
patterns human beings cannot see, getting pre-emptive diagnostic information
to our doctors so we can head off serious illness at early stages, or perhaps
before it even happens. Doctors will be freed up to spend time with their
patients, providing the human touch so often missing in medicine.
Many if not most surgeries will be performed by Artificial Intelligence-
enabled robots, either assisting human surgeons, being supervised by them, or
even, eventually, working fully autonomously. Machine Learning will
monitor our blood gasses under anesthesia, our heart rate, and other health
measures during operations, and reaction in milliseconds should something
appear to be wrong. Iatrogenic disease will decrease dramatically, if not
disappear completely.
The Environment and Sustainability
Machine Learning will be able to study the movement of people in cities,
what they use and don’t use, their patterns of use, how they travel, and where.
Deep learning from this data will allow city planners to employ Machine
Learning algorithms to design and construct both more efficient and pleasant
cities. It will allow massive increases in density without sacrificing quality of
life. These efficiencies will reduce or even possibly eliminate net carbon
emissions from cities.
Augmented Reality
When we wear Google (or Microsoft or Apple or Facebook) glasses in the
future, the embedded processes, video capture, audio capture, and
microphones on these devices will do much more than give us directions to
find a location. Machine Learning will be able to see what we see and
provide explanations and predictive guidance throughout our day. Imagine
having your day “painted” with relevant information on interior walls and
doors, and outside on buildings and signs, guiding you through your schedule
with the information you need right when you need it.
Information Technology
Machine Learning will become a tool people and businesses can apply as
needed like SasS. This MLaaS will allow software to be aware of its
surroundings, to see us, to hear us, and speak to us in natural language.
Connected to the internet, every device will become smart, and generate an
ecosphere around us that attends to our needs and concerns, often before we
even realize we have them. These “Cognitive Services” will provide APIs
and SDKs, leading to rapid “smart” software development and deployment.
Specialized hardware will increase the speed of Machine Learning training,
as well as increase its speed in servicing us. Dedicated AI chips will bring
about a huge change in Artificial Intelligence speed and ubiquity.
Microcomputers will come equipped with Machine Learning capabilities so
that even the smallest device will be able to understand its surroundings.
Where there is no convenient power supply, these devices will run on dime-
sized batteries, lasting for months or years of service before needing
replacement. Currently, these microcomputers cost about 50 cents each. This
price will drop. They will be deployed practically everywhere.
Quantum computing and Machine Learning will merge, bringing solutions to
problems we don’t even know we have yet.
We will see the rise of intelligent robots of every size, make and description,
all dedicated to making our lives better.
Trust Barriers
Natural speech means we will be able to talk to our devices and be
understood by them. The current trust barriers between some people, business
sectors, and governments will slowly break down as Machine Learning
demonstrates its reliability and effectiveness. Improved unsupervised
learning will reduce the time required to develop new Machine Learning
software with required specifications.
Conclusion
Thank you for making it through to the end of Machine Learning for
Beginners. Let’s hope it was informative and able to provide you with all of
the information you needed to begin to understand this extensive topic.
The impact of Machine Learning on our world is already ubiquitous. Our
cars, our phones, our houses, and so much more are already being controlled
and maintained through rudimentary Machine Learning systems. But in the
future, Machine Learning will radically change the world. Some of those
changes are easy to predict. In the next decade or two, people will no longer
drive cars. Instead, automated cars will drive people. But in many other ways,
the effect of Machine Learning on our world is difficult to predict. Will
Machine Learning algorithms replace so many jobs, from trucking to
accounting to many other disciplines, that there won’t be much work left for
people? In 100 years, will there be work for anyone at all? We don’t know
the answer to questions like these because there is so far no limit to what
Machine Learning can accomplish, given time and data and the will to use it
to achieve a particular task.
The future is not necessarily frightening. If there is no work in the future, it
won’t mean that things aren’t getting done. Food will still be grown, picked,
transported to market, and displayed in stores. It’s just that people won’t have
to do any of that labor. As a matter of fact, stores won’t be necessary either,
since the food we order can be delivered directly to our homes. What will the
world be like if human beings have almost unlimited leisure time? Is this a
possible future?
The only real certainty about artificial intelligence and Machine Learning is
that it is increasing in both speed of deployment and in areas which it can
influence. It promises many benefits and many radical changes in our society.
Learning about this technology and creating your own programming in
relation to the development of your own machinery will improve knowledge
of this increasingly popular technology and skyrocket you to the head of
future developments.
Conclusion
Machine learning is a branch of artificial intelligence that involves the design
and development of systems capable of showing an improvement in
performance based on their previous experiences. This means that, when
reacting to the same situation, a machine should show improvement from
time to time. With machine learning, software systems are able to predict
accurately without having to be programmed explicitly. The goal of machine
learning is to build algorithms which can receive input data then use
statistical analysis so as to predict the output value in an acceptable range.
Machine learning originated from pattern recognition and the theory that
computers are able to learn without the need for programming them to
perform tasks. Researchers in the field of artificial intelligence wanted to
determine whether computers are able to learn from data. Machine learning is
an iterative approach, and this is why models are able to adapt as they are
being exposed to new data. Models learn from their previous computations so
as to give repeatable, reliable results and decisions.
 References
Agile Actors. (2019). Scikit-learn Tutorial: Machine Learning in Python - Agile Actors #learning.
Retrieved from https://learningactors.com/scikit-learn-tutorial-machine-learning-in-python/
Bose, D. (2019). Benefits of Artificial Intelligence to Web Development. Retrieved from
http://www.urbanui.com/artificial-intelligence-web-development/
Internet of things. (2019). Retrieved from https://en.wikipedia.org/wiki/Internet_of_things
DataMafia. (2019). DataMafia. Retrieved from https://datamafia2.wordpress.com/author/datamafia007/
Karbhari, V. (2019). Top AI Interview Questions & Answers — Acing the AI Interview. Retrieved
from https://medium.com/acing-ai/top-ai-interview-questions-answers-acing-the-ai-interview-
61bf52ca34d4
Newell, G. (2019). How to Decompress Files With the "gz" Extension. Retrieved from
https://www.lifewire.com/example-uses-of-the-gunzip-command-4081346
Pandas, D. (2019). Different types of features to train Naive Bayes in Python Pandas. Retrieved from
https://stackoverflow.com/questions/32707914/different-types-of-features-to-train-naive-bayes-in-
python-pandas/32747371#32747371
Robinson, S. (2019). K-Means Clustering with Scikit-Learn. Retrieved from https://stackabuse.com/k-
means-clustering-with-scikit-learn/
Top 10 Machine Learning Algorithms. (2019). Retrieved from https://www.dezyre.com/article/top-10-
machine-learning-algorithms/202
Samuel, N. (2019). Installing TensorFlow on Windows. Retrieved from
https://stackabuse.com/installing-tensorflow-on-windows/

[1]
Information retrieved from Robinson, S. (2019). K-Means Clustering with Scikit-Learn. Retrieved
from https://stackabuse.com/k-means-clustering-with-scikit-learn/
[2]
Information obtained from DataMafia. (2019). DataMafia. Retrieved from
https://datamafia2.wordpress.com/author/datamafia007/
[3]
Information obtained from Robinson, S. (2019). K-Means Clustering with Scikit-Learn. Retrieved
from https://stackabuse.com/k-means-clustering-with-scikit-learn/
[4]
Information obtained from Robinson, S. (2019). K-Means Clustering with Scikit-Learn. Retrieved
from https://stackabuse.com/k-means-clustering-with-scikit-learn/
[5]
Information obtained from Pandas, D. (2019). Different types of features to train Naive Bayes in
Python Pandas. Retrieved from https://stackoverflow.com/questions/32707914/different-types-of-
features-to-train-naive-bayes-in-python-pandas/32747371#32747371
[6]
Information obtained from Pandas, D. (2019). Different types of features to train Naive Bayes in
Python Pandas. Retrieved from https://stackoverflow.com/questions/32707914/different-types-of-
features-to-train-naive-bayes-in-python-pandas/32747371#32747371
[7]
("Internet of things", 2019)
[8]
("Top 10 Machine Learning Algorithms", 2019)