Fundamentals of Deep Learning: Designing Next-Generation Machine Intelligence Algorithms

Cover......Page 1
Copyright......Page 4
Table of Contents......Page 5
 Conventions Used in This Book......Page 11
 Safari® Books Online......Page 12
 Acknowledgements......Page 13
 Building Intelligent Machines......Page 15
 The Limits of Traditional Computer Programs......Page 16
 The Mechanics of Machine Learning......Page 17
 The Neuron......Page 21
 Expressing Linear Perceptrons as Neurons......Page 22
 Feed-Forward Neural Networks......Page 23
 Linear Neurons and Their Limitations......Page 26
 Sigmoid, Tanh, and ReLU Neurons......Page 27
 Looking Forward......Page 29
 The Fast-Food Problem......Page 31
 Gradient Descent......Page 33
 The Delta Rule and Learning Rates......Page 35
 Gradient Descent with Sigmoidal Neurons......Page 36
 The Backpropagation Algorithm......Page 37
 Stochastic and Minibatch Gradient Descent......Page 39
 Test Sets, Validation Sets, and Overfitting......Page 41
 Preventing Overfitting in Deep Neural Networks......Page 48
 Summary......Page 51
 What Is TensorFlow?......Page 53
 How Does TensorFlow Compare to Alternatives?......Page 54
 Installing TensorFlow......Page 55
 Creating and Manipulating TensorFlow Variables......Page 57
 Placeholder Tensors......Page 59
 Sessions in TensorFlow......Page 60
 Navigating Variable Scopes and Sharing Variables......Page 62
 Managing Models over the CPU and GPU......Page 65
 Specifying the Logistic Regression Model in TensorFlow......Page 66
 Logging and Training the Logistic Regression Model......Page 69
 Leveraging TensorBoard to Visualize Computation Graphs and Learning......Page 72
 Building a Multilayer Model for MNIST in TensorFlow......Page 73
 Summary......Page 76
 The Challenges with Gradient Descent......Page 77
 Local Minima in the Error Surfaces of Deep Networks......Page 78
 Model Identifiability......Page 79
 How Pesky Are Spurious Local Minima in Deep Networks?......Page 80
 Flat Regions in the Error Surface......Page 83
 When the Gradient Points in the Wrong Direction......Page 85
 Momentum-Based Optimization......Page 88
 A Brief View of Second-Order Methods......Page 91
 Learning Rate Adaptation......Page 92
  AdaGrad—Accumulating Historical Gradients......Page 93
  RMSProp—Exponentially Weighted Moving Average of Gradients......Page 94
  Adam—Combining Momentum and RMSProp......Page 95
 Summary......Page 97
 Neurons in Human Vision......Page 99
 The Shortcomings of Feature Selection......Page 100
 Vanilla Deep Neural Networks Don’t Scale......Page 103
 Filters and Feature Maps......Page 104
 Full Description of the Convolutional Layer......Page 109
 Max Pooling......Page 112
 Full Architectural Description of Convolution Networks......Page 113
 Closing the Loop on MNIST with Convolutional Networks......Page 115
 Image Preprocessing Pipelines Enable More Robust Models......Page 117
 Accelerating Training with Batch Normalization......Page 118
 Building a Convolutional Network for CIFAR-10......Page 121
 Visualizing Learning in Convolutional Networks......Page 123
 Leveraging Convolutional Filters to Replicate Artistic Styles......Page 127
 Learning Convolutional Filters for Other Problem Domains......Page 128
 Summary......Page 129
 Learning Lower-Dimensional Representations......Page 131
 Principal Component Analysis......Page 132
 Motivating the Autoencoder Architecture......Page 134
 Implementing an Autoencoder in TensorFlow......Page 135
 Denoising to Force Robust Representations......Page 148
 Sparsity in Autoencoders......Page 151
 When Context Is More Informative than the Input Vector......Page 154
 The Word2Vec Framework......Page 157
 Implementing the Skip-Gram Architecture......Page 160
 Summary......Page 166
 Analyzing Variable-Length Inputs......Page 167
 Tackling seq2seq with Neural N-Grams......Page 169
 Implementing a Part-of-Speech Tagger......Page 170
 Dependency Parsing and SyntaxNet......Page 178
 Beam Search and Global Normalization......Page 182
 A Case for Stateful Deep Learning Models......Page 186
 Recurrent Neural Networks......Page 187
 The Challenges with Vanishing Gradients......Page 190
 Long Short-Term Memory (LSTM) Units......Page 192
 TensorFlow Primitives for RNN Models......Page 197
 Implementing a Sentiment Analysis Model......Page 199
 Solving seq2seq Tasks with Recurrent Neural Networks......Page 203
 Augmenting Recurrent Networks with Attention......Page 205
 Dissecting a Neural Translation Network......Page 208
 Summary......Page 231
 Neural Turing Machines......Page 233
 Attention-Based Memory Access......Page 235
 NTM Memory Addressing Mechanisms......Page 237
 Differentiable Neural Computers......Page 240
 Interference-Free Writing in DNCs......Page 243
 DNC Memory Reuse......Page 244
 Temporal Linking of DNC Writes......Page 245
 The DNC Controller Network......Page 246
 Visualizing the DNC in Action......Page 248
 Implementing the DNC in TensorFlow......Page 251
 Teaching a DNC to Read and Comprehend......Page 256
 Summary......Page 258
 Deep Reinforcement Learning Masters Atari Games......Page 259
 What Is Reinforcement Learning?......Page 261
 Markov Decision Processes (MDP)......Page 262
  Policy......Page 263
  Future Return......Page 264
 Explore Versus Exploit......Page 265
 Policy Versus Value Learning......Page 267
  OpenAI Gym......Page 268
  Creating an Agent......Page 269
  Keeping Track of History......Page 271
  Policy Gradient Main Function......Page 272
  PGAgent Performance on Pole-Cart......Page 274
  The Bellman Equation......Page 275
  Approximating the Q-Function......Page 276
  Learning Stability......Page 277
  From Q-Function to Policy......Page 278
  Playing Breakout wth DQN......Page 279
  Setting Up Training Operations......Page 282
  Implementing Experience Replay......Page 283
  DQN Main Loop......Page 284
  DQNAgent Results on Breakout......Page 286
  Deep Recurrent Q-Networks (DRQN)......Page 287
  Asynchronous Advantage Actor-Critic Agent (A3C)......Page 288
  UNsupervised REinforcement and Auxiliary Learning (UNREAL)......Page 289
 Summary......Page 290
Index......Page 291
Colophon......Page 298