Grokking Machine Learning

inside front cover

The way to descend from the mountain is to take that one small step in the direction that makes us descend the most and to continue doing this for a long time.

Grokking Machine Learning

Luis G. Serrano

To comment go to liveBook

Manning

Shelter Island

For more information on this and other Manning titles go to

www.manning.com

Copyright

For online information and ordering of these  and other Manning books, please visit www.manning.com. The publisher offers discounts on these books when ordered in quantity.

For more information, please contact

Special Sales Department

Manning Publications Co.

20 Baldwin Road

PO Box 761

Shelter Island, NY 11964

Email: [email protected]

©2021 by Manning Publications Co. All rights reserved.

No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by means electronic, mechanical, photocopying, or otherwise, without prior written permission of the publisher.

Many of the designations used by manufacturers and sellers to distinguish their products are claimed as trademarks. Where those designations appear in the book, and Manning Publications was aware of a trademark claim, the designations have been printed in initial caps or all caps.

♾ Recognizing the importance of preserving what has been written, it is Manning’s policy to have the books we publish printed on acid-free paper, and we exert our best efforts to that end. Recognizing also our responsibility to conserve the resources of our planet, Manning books are printed on paper that is at least 15 percent recycled and processed without the use of elemental chlorine.

ISBN: 9781617295911

contents

front matter

foreword

preface

acknowledgments

about this book

about the author

  1  What is machine learning? It is common sense, except done by a computer

Do I need a heavy math and coding background to understand machine learning?

OK, so what exactly is machine learning?

How do we get machines to make decisions with data? The remember-formulate-predict framework

  2  Types of machine learning

What is the difference between labeled and unlabeled data?

Supervised learning: The branch of machine learning that works with labeled data

Unsupervised learning: The branch of machine learning that works with unlabeled data

What is reinforcement learning?

  3  Drawing a line close to our points: Linear regression

The problem: We need to predict the price of a house

The solution: Building a regression model for housing prices

How to get the computer to draw this line: The linear regression algorithm

How do we measure our results? The error function

Real-life application: Using Turi Create to predict housing prices in India

What if the data is not in a line? Polynomial regression

Parameters and hyperparameters

Applications of regression

  4  Optimizing the training process: Underfitting, overfitting, testing, and regularization

An example of underfitting and overfitting using polynomial regression

How do we get the computer to pick the right model? By testing

Where did we break the golden rule, and how do we fix it? The validation set

A numerical way to decide how complex our model should be: The model complexity graph

Another alternative to avoiding overfitting: Regularization

Polynomial regression, testing, and regularization with Turi Create

  5  Using lines to split our points: The perceptron algorithm

The problem: We are on an alien planet, and we don’t know their language!

How do we determine whether a classifier is good or bad? The error function

How to find a good classifier? The perceptron algorithm

Coding the perceptron algorithm

Applications of the perceptron algorithm

  6  A continuous approach to splitting points: Logistic classifiers

Logistic classifiers: A continuous version of perceptron classifiers

How to find a good logistic classifier? The logistic regression algorithm

Coding the logistic regression algorithm

Real-life application: Classifying IMDB reviews with Turi Create

Classifying into multiple classes: The softmax function

  7  How do you measure classification models? Accuracy and its friends

Accuracy: How often is my model correct?

How to fix the accuracy problem? Defining different types of errors and how to measure them

A useful tool to evaluate our model: The receiver operating characteristic (ROC) curve

  8  Using probability to its maximum: The naive Bayes model

Sick or healthy? A story with Bayes’ theorem as the hero

Use case: Spam-detection model

Building a spam-detection model with real data

  9  Splitting data by asking questions: Decision trees

The problem: We need to recommend apps to users according to what they are likely to download

The solution: Building an app-recommendation system

Beyond questions like yes/no

The graphical boundary of decision trees

Real-life application: Modeling student admissions with Scikit-Learn

Decision trees for regression

Applications

10  Combining building blocks to gain more power: Neural networks

Neural networks with an example: A more complicated alien planet

Training neural networks

Coding neural networks in Keras

Neural networks for regression

Other architectures for more complex datasets

11  Finding boundaries with style: Support vector machines and the kernel method

Using a new error function to build better classifiers

Coding support vector machines in Scikit-Learn

Training SVMs with nonlinear boundaries: The kernel method

12  Combining models to maximize results: Ensemble learning

With a little help from our friends

Bagging: Joining some weak learners randomly to build a strong learner

AdaBoost: Joining weak learners in a clever way to build a strong learner

Gradient boosting: Using decision trees to build strong learners

XGBoost: An extreme way to do gradient boosting

Applications of ensemble methods

13  Putting it all in practice: A real-life example of data engineering and machine learning

The Titanic dataset

Cleaning up our dataset: Missing values and how to deal with them

Feature engineering: Transforming the features in our dataset before training the models

Training our models

Tuning the hyperparameters to find the best model: Grid search

Using K-fold cross-validation to reuse our data as training and validation

Appendix A.  Solutions to the exercises

Appendix B.  The math behind gradient descent: Coming down a mountain using derivatives and slopes

Appendix C.  References

index

front matter

foreword

Did you think machine learning is complicated and hard to master? It’s not! Read this book!

Luis Serrano is a wizard when it comes to explaining things in plain English. I met him first when he taught machine learning on Udacity. He made our students feel that all of machine learning is as simple as adding or subtracting numbers. And most of all, he made the material fun. The videos he produced for Udacity were incredibly engaging and remain among the most liked content offered on the platform.

This book is better! Even the most fearful will enjoy the material presented herein, as Serrano demystifies some of the best-held secrets of the machine learning society. He takes you step by step through each of the critical algorithms and techniques in the field. You can become a machine learning aficionado even if you dislike math. Serrano minimizes the mathematical kauderwelsch that so many of us hard-core academics have come to love, and instead relies on intuition and practical explanations.

The true goal of this book is to empower you to master these methods yourself. So the book is full of fun exercises, in which you get to try out those mystical (and now demystified) techniques yourself. Would you rather gorge on the latest Netflix TV show, or spend your time applying machine learning to problems in computer vision and natural language understanding? If the latter, this book is for you. I can’t express how much fun it is to play with the latest in machine learning, and see your computer do magic under your supervision.

And since machine learning is just about the hottest technology to emerge in the past few years, you will now be able to leverage your new-found skills in your job. A few years back, the New York Times proclaimed that there were only 10,000 machine learning experts in the world, with millions of open positions. That is still the case today! Work through this book and become a professional machine learning engineer. You are guaranteed to possess one of the most in-demand skills in the world today.

With this book, Luis Serrano has done an admirable job explaining complex algorithms and making them accessible to almost everyone. But he doesn’t compromise depth. Instead, he focuses on the empowerment of the reader through a sequence of enlightening projects and exercises. In this sense, this is not a passive read. To fully benefit from this book, you have to work. At Udacity, we have a saying: You won’t lose weight by watching someone else exercise. To grok machine learning, you have to learn to apply it to real-world problems. If you are ready to do this, this is your book—whoever you are!

Sebastian Thrun, PhD

Founder, Udacity

Adjunct Professor, Stanford University

preface

The future is here, and that future has a name: machine learning. With applications in pretty much every industry, from medicine to banking, from self-driving cars to ordering our coffee, the interest in machine learning has rapidly grown day after day. But what is machine learning?

Most of the time, when I read a machine learning book or attend a machine learning lecture, I see either a sea of complicated formulas or a sea of lines of code. For a long time, I thought that this was machine learning, and that machine learning was reserved only for those who had a solid knowledge of both math and computer science.

However, I began to compare machine learning with other subjects, such as music. Musical theory and practice are complicated subjects. But when we think of music, we do not think of scores and scales; we think of songs and melodies. And then I wondered, is machine learning the same? Is it really just a bunch of formulas and code, or is there a melody behind it?

Figure FM.1 Music is not only about scales and notes. There is a melody behind all the technicalities. In the same way, machine learning is not only about formulas and code. There is also a melody, and in this book, we sing it.

With this in mind, I embarked on a journey to understand the melody of machine learning. I stared at formulas and code for months. I drew many diagrams. I scribbled drawings on napkins and showed them to my family, friends, and colleagues. I trained models on small and large datasets. I experimented. After a while, I started listening to the melody of machine learning. All of a sudden, some very pretty pictures started forming in my mind. I started writing stories that go along with all the machine learning concepts. Melodies, pictures, stories—that is how I enjoy learning any topic, and it is those melodies, those pictures, and those stories that I share with you in this book. My goal is to make machine learning fully understandable to every human, and this book is a step in that journey—a step that I’m happy you are taking with me!

acknowledgments

First and foremost, I would like to thank my editor, Marina Michaels, without whom this book wouldn’t exist. Her organization, thorough editing, and valuable input helped shape Grokking Machine Learning. I thank Marjan Bace, Bert Bates, and the rest of the Manning team for their support, professionalism, great ideas, and patience. I thank my technical proofers, Shirley Yap and Karsten Strøbæk; my technical development editor, Kris Athi; and the reviewers for giving me great feedback and correcting many of my mistakes. I thank the production editor, Keri Hales, the copy editor, Pamela Hunt, the graphics editor, Jennifer Houle, the proofreader, Jason Everett, and the entire production team for their wonderful work in making this book a reality. I thank Laura Montoya for her help with inclusive language and AI ethics, Diego Hernandez for valuable additions to the code, and Christian Picón for his immense help with the technical aspects of the repository and the packages.

I am grateful to Sebastian Thrun for his excellent work democratizing education. Udacity was the platform that first gave me a voice to teach the world, and I would like to thank the wonderful colleagues and students I met there. Alejandro Perdomo and the Zapata Computing team deserve thanks for introducing me to the world of quantum machine learning. Thanks also to the many wonderful leaders and colleagues I met at Google and Apple who were instrumental in my career. Special thanks to Roberto Cipriani and the team at Paper Inc. for letting me be part of the family and for the wonderful job they do in the education community.

I’d like to thank my many academic mentors who have shaped my career and my way of thinking: Mary Falk de Losada and her team at the Colombian Mathematical Olympiads, where I first started loving mathematics and had the chance to meet great mentors and create friendships that have lasted a lifetime; my PhD advisor, Sergey Fomin, who was instrumental in my mathematical education and my style of teaching; my master’s advisor, Ian Goulden; Nantel and François Bergeron, Bruce Sagan and Federico Ardila, and the many professors and colleagues I had the opportunity to work with, in particular those at the Universities of Waterloo, Michigan, Quebec at Montreal, and York; and finally, Richard Hoshino and the team and students at Quest University, who helped me test and improve the material in this book.

To all the reviewers: Al Pezewski, Albert Nogués Sabater, Amit Lamba, Bill Mitchell, Borko Djurkovic, Daniele Andreis, Erik Sapper, Hao Liu, Jeremy R. Loscheider, Juan Gabriel Bono, Kay Engelhardt, Krzysztof Kamyczek, Matthew Margolis, Matthias Busch, Michael Bright, Millad Dagdoni, Polina Keselman, Tony Holdroyd, and Valerie Parham-Thompson, your suggestions helped make this a better book.

I would like to thank my wife, Carolina Lasso, who supported me at every step of this process with love and kindness; my mom, Cecilia Herrera, who raised me with love and always encouraged me to follow my passions; my grandma, Maruja, for being the angel that looks at me from heaven; my best friend, Alejandro Morales, for always being there for me; and my friends who have enlightened my path and brightened my life, I thank you and love you with all my heart.

YouTube, blogs, podcasts, and social media have given me the chance to connect with thousands of brilliant souls all over the world. Curious minds with an endless passion for learning, fellow educators who generously share their knowledge and insights, form an e-tribe that inspires me every day and gives me the energy to continue teaching and learning. To anyone who shares their knowledge with the world or who strives to learn every day, I thank you.

I thank anyone out there who is striving to make this world a more fair and peaceful place. To anyone who fights for justice, for peace, for the environment, and for equal opportunities for every human on Earth regardless of their race, gender, place of birth, conditions, and choices, I thank you from the bottom of my heart.

And last, but certainly not least, this book is dedicated to you, the reader. You have chosen the path of learning, the path of improving, the path of feeling comfortable in the uncomfortable, and that is admirable. I hope this book is a positive step along your way to following your passions and creating a better world.

about this book

This book teaches you two things: machine learning models and how to use them. Machine learning models come in different types. Some of them return a deterministic answer, such as yes or no, whereas others return the answer as a probability. Some of them use equations; others use if statements. One thing they have in common is that they all return an answer, or a prediction. The branch of machine learning that comprises the models that return a prediction is aptly named predictive machine learning. This is the type of machine learning that we focus on in this book.

How this book is organized: A roadmap

Types of chapters

This book has two types of chapters. The majority of them (chapters 3, 5, 6, 8, 9, 10, 11, and 12) each contain one type of machine learning model. The corresponding model in each chapter is studied in detail, including examples, formulas, code, and exercises for you to solve. Other chapters (chapters 4, 7, and 13) contain useful techniques to use to train, evaluate, and improve machine learning models. In particular, chapter 13 contains an end-to-end example on a real dataset, in which you’ll be able to apply all the knowledge you’ve obtained in the previous chapters.

Recommended learning paths

You can use this book in two ways. The one I recommend is to go through it linearly, chapter by chapter, because you’ll find that the alternation between learning models and learning techniques to train them is rewarding. However, another learning path is to first learn all the models (chapters 3, 5, 6, 8, 9, 10, 11, and 12), and then learn the techniques for training them (chapters 4, 7, and 13). And of course, because we all learn in different ways, you can create your own learning path!

Appendices

This book has three appendices. Appendix A contains the solutions to each chapter’s exercises. Appendix B contains some formal mathematical derivations that are useful but more technical than the rest of the book. Appendix C contains a list of references and resources that I recommend if you’d like to further your understanding.

Requirements and learning goals

This book provides you with a solid framework of predictive machine learning. To get the most out of this book, you should have a visual mind and a good understanding of elementary mathematics, such as graphs of lines, equations, and basic probability. It is helpful (although not mandatory) if you know how to code, especially in Python, because you are given the opportunity to implement and apply several models in real datasets throughout the book. After reading this book, you will be able to do the following:

Describe the most important models in predictive machine learning and how they work, including linear and logistic regression, naive Bayes, decision trees, neural networks, support vector machines, and ensemble methods.

Identify their strengths and weaknesses and what parameters they use.

Identify how these models are used in the real world, and formulate potential ways to apply machine learning to any particular problem you would like to solve.

Learn how to optimize these models, compare them, and improve them, to build the best machine learning models we can.

Code the models, whether by hand or using an existing package, and use them to make predictions on real datasets.

If you have a particular dataset or problem in mind, I invite you to think about how to apply what you learn in this book to it, and to use it as a starting point to implement and experiment with your own models.

I am super excited to start this journey with you, and I hope you are as excited!

Other resources

This book is self-contained. This means that aside from the requirements described earlier, every concept that we need is introduced in the book. However, I include many references, which I recommend you check out if you’d like to understand the concepts at a deeper level or if you’d like to explore further topics. The references are all in appendix C and also at this link: http://serrano.academy/grokking-machine-learning.

In particular, several of my own resources accompany this book’s material. In my page at http://serrano.academy, you can find a lot of materials in the form of videos, posts, and code. The videos are also in my YouTube channel www.youtube.com/c/LuisSerrano, which I recommend you check out. As a matter of fact, most of the chapters in this book have a corresponding video that I recommend you watch as you read the chapter.

We’ll be writing code

In this book, we’ll be writing code in Python. However, if your plan is to learn the concepts without the code, you can still follow the book while ignoring the code. Nevertheless, I recommend you at least take a look at the code, so you get familiarized with it.

This book comes with a code repository, and most chapters will give you the opportunity to code the algorithms from scratch or to use some very popular Python packages to build models that fit given datasets. The GitHub repository is www.github.com/luisguiserrano/manning, and I link the corresponding notebooks throughout the book. In the README of the repository, you will find the instructions for the packages to install to run the code successfully.

The main Python packages we use in this book are the following:

NumPy: for storing arrays and performing complex mathematical calculations

Pandas: for storing, manipulating, and analyzing large datasets

Matplotlib: for plotting data

Turi Create: for storing and manipulating data and training machine learning models

Scikit-Learn: for training machine learning models

Keras (TensorFlow): for training neural networks

About the code

This book contains many examples of source code in line with normal text. In both cases, source code is formatted in a fixed-width font like this to separate it from ordinary text. Sometimes code is also in bold to highlight code that has changed from previous steps in the chapter, such as when a new feature adds to an existing line of code.

In many cases, the original source code has been reformatted; we’ve added line breaks and reworked indentation to accommodate the available page space in the book. Additionally, comments in the source code have often been removed from the listings when the code is described in the text. Code annotations accompany many of the listings, highlighting important concepts.

The code for the examples in this book is available for download on the Manning website (https://www.manning.com/books/grokking-machine-learning), and from GitHub at www.github.com/luisguiserrano/manning.

liveBook discussion forum

Purchase of Grokking Machine Learning includes free access to a private web forum run by Manning Publications where you can make comments about the book, ask technical questions, and receive help from the author and from other users. To access the forum, go to https://livebook.manning.com/#!/book/grokking-machine-learning/discussion. You can also learn more about Manning’s forums and the rules of conduct at https://livebook.manning.com/#!/discussion.

Manning’s commitment to our readers is to provide a venue where a meaningful dialogue between individual readers and between readers and the author can take place. It is not a commitment to any specific amount of participation on the part of the author, whose contribution to the forum remains voluntary (and unpaid). We suggest you try asking the author some challenging questions lest his interest stray! The forum and the archives of previous discussions will be accessible from the publisher’s website as long as the book is in print.

about the author

1 What is machine learning? It is common sense, except done by a computer

In this chapter

what is machine learning

is machine learning hard (spoiler: no)

what do we learn in this book

what is artificial intelligence, and how does it differ from machine learning

how do humans think, and how can we inject those ideas into a machine

some basic machine learning examples in real life

I am super happy to join you in your learning journey!

Welcome to this book! I’m super happy to be joining you in this journey through understanding machine learning. At a high level, machine learning is a process in which the computer solves problems and makes decisions in much the same way as humans.

In this book, I want to bring one message to you: machine learning is easy! You do not need to have a heavy math and programming background to understand it. You do need some basic mathematics, but the main ingredients are common sense, a good visual intuition, and a desire to learn and apply these methods to anything that you are passionate about and where you want to make an improvement in the world. I’ve had an absolute blast writing this book, because I love growing my understanding of this topic, and I hope you have a blast reading it and diving deep into machine learning!

Machine learning is everywhere

Machine learning is everywhere. This statement seems to be truer every day. I have a hard time imagining a single aspect of life that cannot be improved in some way or another by machine learning. For any job that requires repetition or looking at data and gathering conclusions, machine learning can help. During the last few years, machine learning has seen tremendous growth due to the advances in computing power and the ubiquity of data collection. Just to name a few applications of machine learning: recommendation systems, image recognition, text processing, self-driving cars, spam recognition, medical diagnoses . . . the list goes on. Perhaps you have a goal or an area in which you want to make an impact (or maybe you are already making it!). Very likely, machine learning can be applied to that field—perhaps that is what brought you to this book. Let’s find out together!

Do I need a heavy math and coding background to understand machine learning?

No. Machine learning requires imagination, creativity, and a visual mind. Machine learning is about picking up patterns that appear in the world and using those patterns to make predictions in the future. If you enjoy finding patterns and spotting correlations, then you can do machine learning. If I were to tell you that I stopped smoking and am eating more vegetables and exercising, what would you predict will happen to my health in one year? Perhaps that it will improve. If I were to tell you that I’ve switched from wearing red sweaters to green sweaters, what would you predict will happen to my health in one year? Perhaps that it won’t change much (it may, but not based on the information I gave you). Spotting these correlations and patterns is what machine learning is about. The only difference is that in machine learning, we attach formulas and numbers to these patterns to get computers to spot them.

Some mathematics and coding knowledge are needed to do machine learning, but you don’t need to be an expert. If you are an expert in either of them, or both, you will certainly find your skills will be rewarded. But if you are not, you can still learn machine learning and pick up the mathematics and coding as you go. In this book, we introduce all the mathematical concepts we need at the moment we need them. When it comes to coding, how much code you write in machine learning is up to you. Machine learning jobs range from those who code all day long, to those who don’t code at all. Many packages, APIs, and tools help us do machine learning with minimal coding. Every day, machine learning is more available to everyone in the world, and I’m glad you’ve jumped on the bandwagon!

Formulas and code are fun when seen as a language

In most machine learning books, algorithms are explained mathematically using formulas, derivatives, and so on. Although these precise descriptions of the methods work well in practice, a formula sitting by itself can be more confusing than illustrative. However, like a musical score, a formula may hide a beautiful melody behind the confusion. For example, let’s look at this formula: Σi⁴=1i. It looks ugly at first glance, but it represents a very simple sum, namely, 1 + 2 + 3 + 4. And what about Σin=1wi? That is simply the sum of many (n) numbers. But when I think of a sum of many numbers, I’d rather imagine something like 3 + 2 + 4 + 27, rather than 1 Σin=1wi. Whenever I see a formula, I immediately have to imagine a small example of it, and then the picture is clearer in my mind. When I see something like P(A|B), what comes to mind? That is a conditional probability, so I think of some sentence along the lines of The probability that an event A occurs given that another event B already occurs. For example, if A represents rain today and B represents living in the Amazon rain forest, then the formula P(A|B) = 0.8 simply means The probability that it rains today given that we live in the Amazon rain forest is 80%.

If you do love formulas, don’t worry—this book still has them. But they will appear right after the example that illustrates them.

The same phenomenon happens with code. If we look at code from far away, it may look complicated, and we might find it hard to imagine that someone could fit all of that in their head. However, code is simply a sequence of steps, and normally each of these steps is simple. In this book, we’ll write code, but it will be broken down into simple steps, and each step will be carefully explained with examples or illustrations. During the first few chapters, we will be coding our models from scratch to understand how they work. In the later chapters, however, the models get more complicated. For these, we will use packages such as Scikit-Learn, Turi Create, or Keras, which have implemented most machine learning algorithms with great clarity and power.

OK, so what exactly is machine learning?

To define machine learning, first let’s define a more general term: artificial intelligence.

What is artificial intelligence?

Artificial intelligence (AI) is a general term, which we define as follows:

artificial intelligence The set of all tasks in which a computer can make decisions

In many cases, a computer makes these decisions by mimicking the ways a human makes decisions. In other cases, they may mimic evolutionary processes, genetic processes, or physical processes. But in general, any time we see a computer solving a problem by itself, be it driving a car, finding a route between two points, diagnosing a patient, or recommending a movie, we are looking at artificial intelligence.

What is machine learning?

Machine learning is similar to artificial intelligence, and often their definitions are confused. Machine learning (ML) is a part of artificial intelligence, and we define it as follows:

machine learning The set of all tasks in which a computer can make decisions based on data

What does this mean? Allow me to illustrate with the diagram in figure 1.1.

Figure 1.1 Machine learning is a part of artificial intelligence.

Let’s go back to looking at how humans make decisions. In general terms, we make decisions in the following two ways:

By using logic and reasoning

By using our experience

For example, imagine that we are trying to decide what car to buy. We can look carefully at the features of the car, such as price, fuel consumption, and navigation, and try to figure out the best combination of them that adjusts to our budget. That is using logic and reasoning. If instead we ask all our friends what cars they own, and what they like and dislike about them, we form a list of information and use that list to decide, then we are using experience (in this case, our friends’ experiences).

Machine learning represents the second method: making decisions using our experience. In computer lingo, the term for experience is data. Therefore, in machine learning, computers make decisions based on data. Thus, any time we get a computer to solve a problem or make a decision using only data, we are doing machine learning. Colloquially, we could describe machine learning in the following way:

Machine learning is common sense, except done by a computer.

Going from solving problems using any means necessary to solving problems using only data may feel like a small step for a computer, but it has been a huge step for humanity (figure 1.2). Once upon a time, if we wanted to get a computer to perform a task, we had to write a program, namely, a whole set of instructions for the computer to follow. This process is good for simple tasks, but some tasks are too complicated for this framework. For example, consider the task of identifying if an image contains an apple. If we start writing a computer program to develop this task, we quickly find out that it is hard.

Figure 1.2 Machine learning encompasses all the tasks in which computers make decisions based on data. In the same way that humans make decisions based on previous experiences, computers can make decisions based on previous data.

Let’s take a step back and ask the following question. How did we, as humans, learn how an apple looks? The way we learned most words was not by someone explaining to us what they mean; we learned them by repetition. We saw many objects during our childhood, and adults would tell us what these objects were. To learn what an apple was, we saw many apples throughout the years while hearing the word apple, until one day it clicked, and we knew what an apple was. In machine learning, that is what we get the computer to do. We show the computer many images, and we tell it which ones contain an apple (that constitutes our data). We repeat this process until the computer catches the right patterns and attributes that constitute an apple. At the end of the process, when we feed the computer a new image, it can use these patterns to determine whether the image contains an apple. Of course, we still need to program the computer so that it catches these patterns. For that, we have several techniques, which we will learn in this book.

And now that we’re at it, what is deep learning?

In the same way that machine learning is part of artificial intelligence, deep learning is a part of machine learning. In the previous section, we learned we have several techniques we use to get the computer to learn from data. One of these techniques has been performing tremendously well, so it has its own field of study called deep learning (DL), which we define as follows and as shown in figure 1.3:

deep learning The field of machine learning that uses certain objects called neural networks

What are neural networks? We’ll learn about them in chapter 10. Deep learning is arguably the most used type of machine learning because it works really well. If we are looking at any of the cutting-edge applications, such as image recognition, text generation, playing Go, or self-driving cars, very likely we are looking at deep learning in some way or another.

Figure 1.3 Deep learning is a part of machine learning.

In other words, deep learning is part of machine learning, which in turn is part of artificial intelligence. If this book were about transportation, then AI would be vehicles, ML would be cars, and DL would be Ferraris.

How do we get machines to make decisions with data? The remember-formulate-predict framework

In the previous section, we discussed that machine learning consists of a set of techniques that we use to get the computer to make decisions based on data. In this section, we learn what is meant by making decisions based on data and how some of these techniques work. For this, let’s again analyze the process humans use to make decisions based on experience. This is what is called the remember-formulate-predict framework, shown in figure 1.4. The goal of machine learning is to teach computers how to think in the same way, following the same framework.

How do humans think?

When we, as humans, need to make a decision based on our experience, we normally use the following framework:

We remember past situations that were similar.

We formulate a general rule.

We use this rule to predict what may happen in the future.

For example, if the question is, Will it rain today?, the process to make a guess is the following:

We remember that last week it rained most of the time.

We formulate that in this place, it rains most of the time.

We predict that today it will rain.

We may be right or wrong, but at least we are trying to make the most accurate prediction we can based on the information we have.

Figure 1.4 The remember-formulate-predict framework is the main framework we use in this book. It consists of three steps: (1) We remember previous data; (2) we formulate a general rule; and (3) we use that rule to make predictions about the future.

Some machine learning lingo—models and algorithms

Before we delve into more examples that illustrate the techniques used in machine learning, let’s define some useful terms that we use throughout this book. We know that in machine learning, we get the computer to learn how to solve a problem using data. The way the computer solves the problem is by using the data to build a model. What is a model? We define a model as follows:

model A set of rules that represent our data and can be used to make predictions

We can think of a model as a representation of reality using a set of rules that mimic the existing data as closely as possible. In the rain example in the previous section, the model was our representation of reality, which is a world in which it rains most of the time. This is a simple world with one rule: it rains most of the time. This representation may or may not be accurate, but according to our data, it is the most accurate representation of reality that we can formulate. We later use this rule to make predictions on unseen data.

An algorithm is the process that we used to build the model. In the current example, the process is simple: we looked at how many days it rained and realized it was the majority. Of course, machine learning algorithms can get much more complicated than that, but at the end of the day, they are always composed of a set of steps. Our definition of algorithm follows:

algorithm A procedure, or a set of steps, used to solve a problem or perform a computation. In this book, the goal of an algorithm is to build a model.

In short, a model is what we use to make predictions, and an algorithm is what we use to build the model. Those two definitions are easy to confuse and are often interchanged, but to keep them clear, let’s look at a few examples.

Some examples of models that humans use

In this section we focus on a common application of machine learning: spam detection. In the following examples, we will detect spam and non-spam emails. Non-spam emails are also referred to as ham.

spam and ham spam is the common term used for junk or unwanted email, such as chain letters, promotions, and so on. The term comes from a 1972 Monty Python sketch in which every item in the menu of a restaurant contained Spam as an ingredient. Among software developers, the term ham is used to refer to non-spam emails.

Example 1: An annoying email friend

In this example, our friend Bob likes to send us email. A lot of his emails are spam, in the form of chain letters. We are starting to get a bit annoyed with him. It is Saturday, and we just got a notification of an email from Bob. Can we guess if this email is spam or ham without looking at it?

To figure this out, we use the remember-formulate-predict method. First, let us remember, say, the last 10 emails that we got from Bob. That is our data. We remember that six of them were spam, and the other four were ham. From this information, we can formulate the following model:

Model 1: Six out of every 10 emails that Bob sends us are spam.

This rule will be our model. Note, this rule does not need to be true. It could be outrageously wrong. But given our data, it is the best that we can come up with, so we’ll live with it. Later in this book, we learn how to evaluate models and improve them when needed.

Now that we have our rule, we can use it to predict whether the email is spam. If six out of 10 of Bob’s emails are spam, then we can assume that this new email is 60% likely to be spam and 40% likely to be ham. Judging by this rule, it’s a little safer to think that the email is spam. Therefore, we predict that the email is spam (figure 1.5).

Again, our prediction may be wrong. We may open the email and realize it is ham. But we have made the prediction to the best of our knowledge. This is what machine learning is all about.

You may be thinking, can we do better? We seem to be judging every email from Bob in the same way, but there may be more information that can help us tell the spam and ham emails apart. Let’s try to analyze the emails a little more. For example, let’s see when Bob sent the emails to see if we find a pattern.

Figure 1.5 A very simple machine learning model

Example 2: A seasonal annoying email friend

Let’s look more carefully at the emails that Bob sent us in the previous month. More specifically, we’ll look at what day he sent them. Here are the emails with dates and information about being spam or ham:

Monday: Ham

Tuesday: Ham

Saturday: Spam

Sunday: Spam

Sunday: Spam

Wednesday: Ham

Friday: Ham

Saturday: Spam

Tuesday: Ham

Thursday: Ham

Now things are different. Can you see a pattern? It seems that every email Bob sent during the week is ham, and every email he sent during the weekend is spam. This makes sense—maybe during the week he sends us work email, whereas during the weekend, he has time to send spam and decides to roam free. So, we can formulate a more educated rule, or model, as follows:

Model 2: Every email that Bob sends during the week is ham, and those he sends during the weekend are spam.

Now let’s look at what day it is today. If it is Sunday and we just got an email from Bob, then we can predict with great confidence that the email he sent is spam (figure 1.6). We make this prediction, and without looking, we send the email to the trash and carry on with our day.

Figure 1.6 A slightly more complex machine learning model

Example 3: Things are getting complicated!

Now, let’s say we continue with this rule, and one day we see Bob in the street, and he asks, Why didn’t you come to my birthday party? We have no idea what he is talking about. It turns out last Sunday he sent us an invitation to his birthday party, and we missed it! Why did we miss it? Because he sent it on the weekend, and we assumed that it would be spam. It seems that we need a better model. Let’s go back to look at Bob’s emails—this is our remember step. Let’s see if we can find a pattern.

1 KB: Ham

2 KB: Ham

16 KB: Spam

20 KB: Spam

18 KB: Spam

3 KB: Ham

5 KB: Ham

25 KB: Spam

1 KB: Ham

3 KB: Ham

What do we see? It seems that the large emails tend to be spam, whereas the smaller ones tend to be ham. This makes sense, because the spam emails frequently have large attachments.

So, we can formulate the following rule:

Model 3: Any email of size 10 KB or larger is spam, and any email of size less than 10 KB is ham.

Now that we have formulated our rule, we can make a prediction. We look at the email we received today from Bob, and the size is 19 KB. So, we conclude that it is spam (figure 1.7).

Figure 1.7 Another slightly more complex machine learning model

Is this the end of the story? Not even close.

But before we keep going, notice that to make our predictions, we used the day of the week and the size of the email. These are examples of features. A feature is one of the most important concepts in this book.

feature Any property or characteristic of the data that the model can use to make predictions

You can imagine that there are many more features that could indicate if an email is spam or ham. Can you think of some more? In the next paragraphs, we’ll see a few more features.

Example 4: More?

Our two classifiers were good, because they rule out large emails and emails sent on