Learning Deep Learning

Magnus Ekman, Ph.D., is a director of architecture at NVIDIA Corporation. His doctorate is in computer engineering, and he is the inventor of multiple patents. He was first exposed to artificial neural networks in the late nineties in his native country, Sweden. After some dabbling in evolutionary computation, he ended up focusing on computer architecture and relocated to Silicon Valley, where he lives with his wife Jennifer, children Sebastian and Sofia, and dog Babette. He previously worked with processor design and R&D at Sun Microsystems and Samsung Research America, and has been involved in starting two companies, one of which (Skout) was later acquired by The Meet Group, Inc. In his current role at NVIDIA, he leads an engineering team working on CPU performance and power efficiency for system on chips targeting the autonomous vehicle market. As the Deep Learning (DL) field exploded the past few years, fueled by NVIDIA's GPU technology and CUDA, Dr. Ekman found himself in the middle of a company expanding beyond computer graphics into becoming a deep learning (DL) powerhouse. As a part of that journey, he challenged himself to stay up-to-date with the most recent developments in the field. He considers himself to be an educator, and in the process of writing Learning Deep Learning ( LDL), he partnered with the NVIDIA Deep Learning Institute (DLI), which offers hands-on training in AI, accelerated computing, and accelerated data science. He is thrilled about DLI's plans to add LDL to its existing portfolio of self-paced online courses, live instructor-led workshops, educator programs, and teaching kits.

Foreword by Dr. Anima Anandkumar xxi
Foreword by Dr. Craig Clawson xxiii
Preface xxv
Acknowledgments li
About the Author liii

Chapter 1: The Rosenblatt Perceptron 1
Example of a Two-Input Perceptron 4
The Perceptron Learning Algorithm 7
Limitations of the Perceptron 15
Combining Multiple Perceptrons 17
Implementing Perceptrons with Linear Algebra 20
Geometric Interpretation of the Perceptron 30
Understanding the Bias Term 33
Concluding Remarks on the Perceptron 34

Chapter 2: Gradient-Based Learning 37
Intuitive Explanation of the Perceptron Learning Algorithm 37
Derivatives and Optimization Problems 41
Solving a Learning Problem with Gradient Descent 44
Constants and Variables in a Network 48
Analytic Explanation of the Perceptron Learning Algorithm 49
Geometric Description of the Perceptron Learning Algorithm 51
Revisiting Different Types of Perceptron Plots 52
Using a Perceptron to Identify Patterns 54
Concluding Remarks on Gradient-Based Learning 57

Chapter 3: Sigmoid Neurons and Backpropagation 59
Modified Neurons to Enable Gradient Descent for Multilevel Networks 60
Which Activation Function Should We Use? 66
Function Composition and the Chain Rule 67
Using Backpropagation to Compute the Gradient 69
Backpropagation with Multiple Neurons per Layer 81
Programming Example: Learning the XOR Function 82
Network Architectures 87
Concluding Remarks on Backpropagation 89

Chapter 4: Fully Connected Networks Applied to Multiclass Classification 91
Introduction to Datasets Used When Training Networks 92
Training and Inference 100
Extending the Network and Learning Algorithm to Do Multiclass Classification 101
Network for Digit Classification 102
Loss Function for Multiclass Classification 103
Programming Example: Classifying Handwritten Digits 104
Mini-Batch Gradient Descent 114
Concluding Remarks on Multiclass Classification 115

Chapter 5: Toward DL: Frameworks and Network Tweaks 117
Programming Example: Moving to a DL Framework 118
The Problem of Saturated Neurons and Vanishing Gradients 124
Initialization and Normalization Techniques to Avoid Saturated Neurons 126
Cross-Entropy Loss Function to Mitigate Effect of Saturated Output Neurons 130
Different Activation Functions to Avoid Vanishing Gradient in Hidden Layers 136
Variations on Gradient Descent to Improve Learning 141
Experiment: Tweaking Network and Learning Parameters 143
Hyperparameter Tuning and Cross-Validation 146
Concluding Remarks on the Path Toward Deep Learning 150

Chapter 6: Fully Connected Networks Applied to Regression 153
Output Units 154
The Boston Housing Dataset 160
Programming Example: Predicting House Prices with a DNN 161
Improving Generalization with Regularization 166
Experiment: Deeper and Regularized Models for House Price Prediction 169
Concluding Remarks on Output Units and Regression Problems 170

Chapter 7: Convolutional Neural Networks Applied to Image Classification 171
The CIFAR-10 Dataset 173
Characteristics and Building Blocks for Convolutional Layers 175
Combining Feature Maps into a Convolutional Layer 180
Combining Convolutional and Fully Connected Layers into a Network 181
Effects of Sparse Connections and Weight Sharing 185
Programming Example: Image Classification with a Convolutional Network 190
Concluding Remarks on Convolutional Networks 201

Chapter 8: Deeper CNNs and Pretrained Models 205
VGGNet 206
GoogLeNet 210
ResNet 215
Programming Example: Use a Pretrained ResNet Implementation 223
Transfer Learning 226
Backpropagation for CNN and Pooling 228
Data Augmentation as a Regularization Technique 229
Mistakes Made by CNNs 231
Reducing Parameters with Depthwise Separable Convolutions 232
Striking the Right Network Design Balance with EfficientNet 234
Concluding Remarks on Deeper CNNs 235

Chapter 9: Predicting Time Sequences with Recurrent Neural Networks 237
Limitations of Feedforward Networks 241
Recurrent Neural Networks 242
Mathematical Representation of a Recurrent Layer 243
Combining Layers into an RNN 245
Alternative View of RNN and Unrolling in Time 246
Backpropagation Through Time 248
Programming Example: Forecasting Book Sales 250
Dataset Considerations for RNNs 264
Concluding Remarks on RNNs 265

Chapter 10: Long Short-Term Memory 267
Keeping Gradients Healthy 267
Introduction to LSTM 272
LSTM Activation Functions 277
Creating a Network of LSTM Cells 278
Alternative View of LSTM 280
Related Topics: Highway Networks and Skip Connections 282
Concluding Remarks on LSTM 282

Chapter 11: Text Autocompletion with LSTM and Beam Search 285
Encoding Text 285
Longer-Term Prediction and Autoregressive Models 287
Beam Search 289
Programming Example: Using LSTM for Text Autocompletion 291
Bidirectional RNNs 298
Different Combinations of Input and Output Sequences 300
Concluding Remarks on Text Autocompletion with LSTM 302

Chapter 12: Neural Language Models and Word Embeddings 303
Introduction to Language Models and Their Use Cases 304
Examples of Different Language Models 307
Benefit of Word Embeddings and Insight into How They Work 313
Word Embeddings Created by Neural Language Models 315
Programming Example: Neural Language Model and Resulting Embeddings 319
King − Man + Woman! = Queen 329
King − Man + Woman ! = Queen 331
Language Models, Word Embeddings, and Human Biases 332
Related Topic: Sentiment Analysis of Text 334
Concluding Remarks on Language Models and Word Embeddings 342

Chapter 13: Word Embeddings from word2vec and GloVe 343
Using word2vec to Create Word Embeddings Without a Language Model 344
Additional Thoughts on word2vec 352
word2vec in Matrix Form 353
Wrapping Up word2vec 354
Programming Example: Exploring Properties of GloVe Embeddings 356
Concluding Remarks on word2vec and GloVe 361

Chapter 14: Sequence-to-Sequence Networks and Natural Language Translation 363
Encoder-Decoder Model for Sequence-to-Sequence Learning 366
Introduction to the Keras Functional API 368
Programming Example: Neural Machine Translation 371
Experimental Results 387
Properties of the Intermediate Representation 389
Concluding Remarks on Language Translation 391

Chapter 15: Attention and the Transformer 393
Rationale Behind Attention 394
Attention in Sequence-to-Sequence Networks 395
Alternatives to Recurrent Networks 406
Self-Attention 407
Multi-head Attention 410
The Transformer 411
Concluding Remarks on the Transformer 415

Chapter 16: One-to-Many Network for Image Captioning 417
Extending the Image Captioning Network with Attention 420
Programming Example: Attention-Based Image Captioning 421
Concluding Remarks on Image Captioning 443

Chapter 17: Medley of Additional Topics 447
Autoencoders 448
Multimodal Learning 459
Multitask Learning 469
Process for Tuning a Network 477
Neural Architecture Search 482
Concluding Remarks 502

Chapter 18: Summary and Next Steps 503
Things You Should Know by Now 503
Ethical AI and Data Ethics 505
Things You Do Not Yet Know 512
Next Steps 516

Appendix A: Linear Regression and Linear Classifiers 519
Linear Regression as a Machine Learning Algorithm 519
Computing Linear Regression Coefficients 523
Classification with Logistic Regression 525
Classifying XOR with a Linear Classifier 528
Classification with Support Vector Machines 531
Evaluation Metrics for a Binary Classifier 533

Appendix B: Object Detection and Segmentation 539
Object Detection 540
Semantic Segmentation 549
Instance Segmentation with Mask R-CNN 559

Appendix C: Word Embeddings Beyond word2vec and GloVe 563
Wordpieces 564
FastText 566
Character-Based Method 567
ELMo 572
Related Work 575

Appendix D: GPT, BERT, AND RoBERTa 577
GPT 578
BERT 582
RoBERTa 586
Historical Work Leading Up to GPT and BERT 588
Other Models Based on the Transformer 590

Appendix E: Newton-Raphson versus Gradient Descent 593
Newton-Raphson Root-Finding Method 594
Relationship Between Newton-Raphson and Gradient Descent 597

Appendix F: Matrix Implementation of Digit Classification Network 599
Single Matrix 599
Mini-Batch Implementation 602

Appendix G: Relating Convolutional Layers to Mathematical Convolution 607

Appendix H: Gated Recurrent Units 613
Alternative GRU Implementation 616
Network Based on the GRU 616

Appendix I: Setting up a Development Environment 621
Python 622
Programming Environment 623
Programming Examples 624
Datasets 625
Installing a DL Framework 628
TensorFlow Specific Considerations 630
Key Differences Between PyTorch and TensorFlow 631

Appendix J: Cheat Sheets 637

Works Cited 647
Index 667