Table of Contents

Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv

Part I. The Fundamentals of Machine Learning

1. The Machine Learning Landscape. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
What Is Machine Learning? 4
Why Use Machine Learning? 5
Examples of Applications 8
Types of Machine Learning Systems 9
Training Supervision 10
Batch Versus Online Learning 17
Instance-Based Versus Model-Based Learning 21
Main Challenges of Machine Learning 27
Insufficient Quantity of Training Data 27
Nonrepresentative Training Data 28
Poor-Quality Data 30
Irrelevant Features 30
Overfitting the Training Data 30
Underfitting the Training Data 33
Stepping Back 33
Testing and Validating 34
Hyperparameter Tuning and Model Selection 34
Data Mismatch 35
Exercises 37

iii
 2. End-to-End Machine Learning Project. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Working with Real Data 39
Look at the Big Picture 41
Frame the Problem 41
Select a Performance Measure 43
Check the Assumptions 46
Get the Data 46
Running the Code Examples Using Google Colab 46
Saving Your Code Changes and Your Data 48
The Power and Danger of Interactivity 49
Book Code Versus Notebook Code 50
Download the Data 50
Take a Quick Look at the Data Structure 51
Create a Test Set 55
Explore and Visualize the Data to Gain Insights 60
Visualizing Geographical Data 61
Look for Correlations 63
Experiment with Attribute Combinations 66
Prepare the Data for Machine Learning Algorithms 67
Clean the Data 68
Handling Text and Categorical Attributes 71
Feature Scaling and Transformation 75
Custom Transformers 79
Transformation Pipelines 83
Select and Train a Model 88
Train and Evaluate on the Training Set 88
Better Evaluation Using Cross-Validation 89
Fine-Tune Your Model 91
Grid Search 91
Randomized Search 93
Ensemble Methods 95
Analyzing the Best Models and Their Errors 95
Evaluate Your System on the Test Set 96
Launch, Monitor, and Maintain Your System 97
Try It Out! 100
Exercises 101

3. Classification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
MNIST 103
Training a Binary Classifier 106
Performance Measures 107

iv | Table of Contents
 Measuring Accuracy Using Cross-Validation 107
Confusion Matrices 108
Precision and Recall 110
The Precision/Recall Trade-off 111
The ROC Curve 115
Multiclass Classification 119
Error Analysis 122
Multilabel Classification 125
Multioutput Classification 127
Exercises 129

4. Training Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

Linear Regression 132
The Normal Equation 134
Computational Complexity 137
Gradient Descent 138
Batch Gradient Descent 142
Stochastic Gradient Descent 145
Mini-Batch Gradient Descent 148
Polynomial Regression 149
Learning Curves 151
Regularized Linear Models 155
Ridge Regression 156
Lasso Regression 158
Elastic Net Regression 161
Early Stopping 162
Logistic Regression 164
Estimating Probabilities 164
Training and Cost Function 165
Decision Boundaries 167
Softmax Regression 170
Exercises 173

5. Support Vector Machines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

Linear SVM Classification 175
Soft Margin Classification 176
Nonlinear SVM Classification 178
Polynomial Kernel 180
Similarity Features 181
Gaussian RBF Kernel 181
SVM Classes and Computational Complexity 183

Table of Contents | v
 SVM Regression 184
Under the Hood of Linear SVM Classifiers 186
The Dual Problem 189
Kernelized SVMs 190
Exercises 193

6. Decision Trees. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

Training and Visualizing a Decision Tree 195
Making Predictions 197
Estimating Class Probabilities 199
The CART Training Algorithm 199
Computational Complexity 200
Gini Impurity or Entropy? 201
Regularization Hyperparameters 201
Regression 204
Sensitivity to Axis Orientation 206
Decision Trees Have a High Variance 207
Exercises 208

7. Ensemble Learning and Random Forests. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

Voting Classifiers 212
Bagging and Pasting 215
Bagging and Pasting in Scikit-Learn 217
Out-of-Bag Evaluation 218
Random Patches and Random Subspaces 219
Random Forests 220
Extra-Trees 220
Feature Importance 221
Boosting 222
AdaBoost 222
Gradient Boosting 226
Histogram-Based Gradient Boosting 230
Stacking 232
Exercises 235

8. Dimensionality Reduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

The Curse of Dimensionality 238
Main Approaches for Dimensionality Reduction 239
Projection 239
Manifold Learning 241
PCA 243

vi | Table of Contents
 Preserving the Variance 243
Principal Components 244
Projecting Down to d Dimensions 245
Using Scikit-Learn 246
Explained Variance Ratio 246
Choosing the Right Number of Dimensions 247
PCA for Compression 249
Randomized PCA 250
Incremental PCA 250
Random Projection 252
LLE 254
Other Dimensionality Reduction Techniques 256
Exercises 257

9. Unsupervised Learning Techniques. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259

Clustering Algorithms: k-means and DBSCAN 260
k-means 263
Limits of k-means 272
Using Clustering for Image Segmentation 273
Using Clustering for Semi-Supervised Learning 275
DBSCAN 279
Other Clustering Algorithms 282
Gaussian Mixtures 283
Using Gaussian Mixtures for Anomaly Detection 288
Selecting the Number of Clusters 289
Bayesian Gaussian Mixture Models 292
Other Algorithms for Anomaly and Novelty Detection 293
Exercises 294

Part II. Neural Networks and Deep Learning

10. Introduction to Artificial Neural Networks with Keras. . . . . . . . . . . . . . . . . . . . . . . . . . 299
From Biological to Artificial Neurons 300
Biological Neurons 301
Logical Computations with Neurons 303
The Perceptron 304
The Multilayer Perceptron and Backpropagation 309
Regression MLPs 313
Classification MLPs 315
Implementing MLPs with Keras 317

Table of Contents | vii

 Building an Image Classifier Using the Sequential API 318
Building a Regression MLP Using the Sequential API 328
Building Complex Models Using the Functional API 329
Using the Subclassing API to Build Dynamic Models 336
Saving and Restoring a Model 337
Using Callbacks 338
Using TensorBoard for Visualization 340
Fine-Tuning Neural Network Hyperparameters 344
Number of Hidden Layers 349
Number of Neurons per Hidden Layer 350
Learning Rate, Batch Size, and Other Hyperparameters 351
Exercises 353

11. Training Deep Neural Networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357

The Vanishing/Exploding Gradients Problems 358
Glorot and He Initialization 359
Better Activation Functions 361
Batch Normalization 367
Gradient Clipping 372
Reusing Pretrained Layers 373
Transfer Learning with Keras 375
Unsupervised Pretraining 377
Pretraining on an Auxiliary Task 378
Faster Optimizers 379
Momentum 379
Nesterov Accelerated Gradient 381
AdaGrad 382
RMSProp 383
Adam 384
AdaMax 385
Nadam 386
AdamW 386
Learning Rate Scheduling 388
Avoiding Overfitting Through Regularization 392
ℓ1 and ℓ2 Regularization 393
Dropout 394
Monte Carlo (MC) Dropout 397
Max-Norm Regularization 399
Summary and Practical Guidelines 400
Exercises 402

viii | Table of Contents

12. Custom Models and Training with TensorFlow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
A Quick Tour of TensorFlow 403
Using TensorFlow like NumPy 407
Tensors and Operations 407
Tensors and NumPy 409
Type Conversions 409
Variables 410
Other Data Structures 410
Customizing Models and Training Algorithms 412
Custom Loss Functions 412
Saving and Loading Models That Contain Custom Components 413
Custom Activation Functions, Initializers, Regularizers, and Constraints 415
Custom Metrics 416
Custom Layers 419
Custom Models 422
Losses and Metrics Based on Model Internals 424
Computing Gradients Using Autodiff 426
Custom Training Loops 430
TensorFlow Functions and Graphs 433
AutoGraph and Tracing 435
TF Function Rules 437
Exercises 438

13. Loading and Preprocessing Data with TensorFlow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441

The tf.data API 442
Chaining Transformations 443
Shuffling the Data 445
Interleaving Lines from Multiple Files 446
Preprocessing the Data 448
Putting Everything Together 449
Prefetching 450
Using the Dataset with Keras 452
The TFRecord Format 453
Compressed TFRecord Files 454
A Brief Introduction to Protocol Buffers 454
TensorFlow Protobufs 456
Loading and Parsing Examples 457
Handling Lists of Lists Using the SequenceExample Protobuf 459
Keras Preprocessing Layers 459
The Normalization Layer 460
The Discretization Layer 463

Table of Contents | ix
 The CategoryEncoding Layer 463
The StringLookup Layer 465
The Hashing Layer 466
Encoding Categorical Features Using Embeddings 466
Text Preprocessing 471
Using Pretrained Language Model Components 473
Image Preprocessing Layers 474
The TensorFlow Datasets Project 475
Exercises 477

14. Deep Computer Vision Using Convolutional Neural Networks. . . . . . . . . . . . . . . . . . . 479

The Architecture of the Visual Cortex 480
Convolutional Layers 481
Filters 484
Stacking Multiple Feature Maps 485
Implementing Convolutional Layers with Keras 487
Memory Requirements 490
Pooling Layers 491
Implementing Pooling Layers with Keras 493
CNN Architectures 495
LeNet-5 498
AlexNet 499
GoogLeNet 502
VGGNet 505
ResNet 505
Xception 509
SENet 510
Other Noteworthy Architectures 512
Choosing the Right CNN Architecture 514
Implementing a ResNet-34 CNN Using Keras 515
Using Pretrained Models from Keras 516
Pretrained Models for Transfer Learning 518
Classification and Localization 521
Object Detection 523
Fully Convolutional Networks 525
You Only Look Once 527
Object Tracking 530
Semantic Segmentation 531
Exercises 535

x | Table of Contents
15. Processing Sequences Using RNNs and CNNs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
Recurrent Neurons and Layers 538
Memory Cells 540
Input and Output Sequences 541
Training RNNs 542
Forecasting a Time Series 543
The ARMA Model Family 549
Preparing the Data for Machine Learning Models 552
Forecasting Using a Linear Model 555
Forecasting Using a Simple RNN 556
Forecasting Using a Deep RNN 557
Forecasting Multivariate Time Series 559
Forecasting Several Time Steps Ahead 560
Forecasting Using a Sequence-to-Sequence Model 562
Handling Long Sequences 565
Fighting the Unstable Gradients Problem 565
Tackling the Short-Term Memory Problem 568
Exercises 576

16. Natural Language Processing with RNNs and Attention. . . . . . . . . . . . . . . . . . . . . . . . 577

Generating Shakespearean Text Using a Character RNN 578
Creating the Training Dataset 579
Building and Training the Char-RNN Model 581
Generating Fake Shakespearean Text 582
Stateful RNN 584
Sentiment Analysis 587
Masking 590
Reusing Pretrained Embeddings and Language Models 593
An Encoder–Decoder Network for Neural Machine Translation 595
Bidirectional RNNs 601
Beam Search 603
Attention Mechanisms 604
Attention Is All You Need: The Original Transformer Architecture 609
An Avalanche of Transformer Models 620
Vision Transformers 624
Hugging Face’s Transformers Library 629
Exercises 633

17. Autoencoders, GANs, and Diffusion Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635

Efficient Data Representations 637
Performing PCA with an Undercomplete Linear Autoencoder 639

Table of Contents | xi
 Stacked Autoencoders 640
Implementing a Stacked Autoencoder Using Keras 641
Visualizing the Reconstructions 642
Visualizing the Fashion MNIST Dataset 643
Unsupervised Pretraining Using Stacked Autoencoders 644
Tying Weights 645
Training One Autoencoder at a Time 646
Convolutional Autoencoders 648
Denoising Autoencoders 649
Sparse Autoencoders 651
Variational Autoencoders 654
Generating Fashion MNIST Images 658
Generative Adversarial Networks 659
The Difficulties of Training GANs 663
Deep Convolutional GANs 665
Progressive Growing of GANs 668
StyleGANs 671
Diffusion Models 673
Exercises 681

18. Reinforcement Learning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683

Learning to Optimize Rewards 684
Policy Search 685
Introduction to OpenAI Gym 687
Neural Network Policies 691
Evaluating Actions: The Credit Assignment Problem 693
Policy Gradients 694
Markov Decision Processes 699
Temporal Difference Learning 703
Q-Learning 704
Exploration Policies 706
Approximate Q-Learning and Deep Q-Learning 707
Implementing Deep Q-Learning 708
Deep Q-Learning Variants 713
Fixed Q-value Targets 713
Double DQN 714
Prioritized Experience Replay 714
Dueling DQN 715
Overview of Some Popular RL Algorithms 716
Exercises 720

xii | Table of Contents

19. Training and Deploying TensorFlow Models at Scale. . . . . . . . . . . . . . . . . . . . . . . . . . . 721
Serving a TensorFlow Model 722
Using TensorFlow Serving 722
Creating a Prediction Service on Vertex AI 732
Running Batch Prediction Jobs on Vertex AI 739
Deploying a Model to a Mobile or Embedded Device 741
Running a Model in a Web Page 744
Using GPUs to Speed Up Computations 746
Getting Your Own GPU 747
Managing the GPU RAM 749
Placing Operations and Variables on Devices 752
Parallel Execution Across Multiple Devices 753
Training Models Across Multiple Devices 756
Model Parallelism 756
Data Parallelism 759
Training at Scale Using the Distribution Strategies API 765
Training a Model on a TensorFlow Cluster 766
Running Large Training Jobs on Vertex AI 770
Hyperparameter Tuning on Vertex AI 772
Exercises 776
Thank You! 777

A. Machine Learning Project Checklist. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 779

B. Autodiff. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 785

C. Special Data Structures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793

D. TensorFlow Graphs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 801

Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 811

Table of Contents | xiii