How Smart Machines Think

Description

Everything you've always wanted to know about self-driving cars, Netflix recommendations, IBM's Watson, and video game-playing computer programs. The future is here: Self-driving cars are on the streets, an algorithm gives you movie and TV recommendations, IBM's Watson triumphed on Jeopardy over puny human brains, computer programs can be trained to play Atari games. But how do all these thingswork? In this book, Sean Gerrish offers an engaging and accessible overview of the breakthroughs in artificial intelligence and machine learning that have made today's machines so smart. Gerrish outlines some of the key ideas that enable intelligent machines to perceive and interact with the world. He describes the software architecture that allows self-driving cars to stay on the road and to navigate crowded urban environments; the million-dollar Netflix competition for a better recommendation engine (which had an unexpected ending); and how programmers trained computers to perform certain behaviors by offering them treats, as if they were training a dog. He explains how artificial neural networks enable computers to perceive the world―and to play Atari video games better than humans. He explains Watson's famous victory on Jeopardy, and he looks at how computers play games, describing AlphaGo and Deep Blue, which beat reigning world champions at the strategy games of Go and chess. Computers have not yet mastered everything, however; Gerrish outlines the difficulties in creating intelligent agents that can successfully play video games like StarCraft that have evaded solution―at least for now. Gerrish weaves the stories behind these breakthroughs into the narrative, introducing readers to many of the researchers involved, and keeping technical details to a minimum. Science and technology buffs will find this book an essential guide to a future in which machines can outsmart people. Contents......Page 8 Foreword......Page 10 Preface......Page 12 Acknowledgments......Page 14 The Flute Player......Page 16 Today’s Automata......Page 18 The Swing of a Pendulum......Page 19 Automata We’ll Discuss in this Book......Page 20 The $1 Million Race in the Desert......Page 24 How to Build a Self-Driving Car......Page 25 Planning a Path......Page 29 Path Search......Page 30 Navigation......Page 33 The Winner of the Grand Challenge......Page 34 A Failed Race......Page 36 The Second Grand Challenge......Page 38 Stanley’s Architecture......Page 40 Avoiding Obstacles......Page 42 Finding the Road’s Edges......Page 44 Seeing the Road......Page 46 Path Planning......Page 47 How Parts of Stanley’s Brain Talked to Each Other......Page 49 The Urban Challenge......Page 52 Perceptual Abstraction......Page 53 The Race......Page 56 Boss’s Higher-Level Reasoning Layer......Page 57 Getting Past Traffic Jams......Page 63 Three-Layer Architectures......Page 65 Classifying the Objects Seen by Self-Driving Cars......Page 69 Self-Driving Cars are Complicated Systems......Page 70 The Trajectory of Self-Driving Cars......Page 71 A Million-Dollar Grand Prize......Page 72 The Contenders......Page 73 How to Train a Classifier......Page 74 The Goals of the Competition......Page 77 A Giant Ratings Matrix......Page 78 Matrix Factorization......Page 82 The First Year Ends......Page 86 Closing the Gap between Contenders......Page 88 The End of the First Year......Page 89 Predictions Over Time......Page 92 Overfitting......Page 94 Model Blending......Page 95 The Second Year......Page 97 The Final Year......Page 98 After the Competition......Page 101 DeepMind Plays Atari......Page 104 Reinforcement Learning......Page 106 Instructions to the Agent......Page 108 Programming the Agent......Page 110 Nuggets of Experience......Page 114 Playing Atari with Reinforcement Learning......Page 118 Approximation, Not Perfection......Page 122 Neural Networks as Mathematical Functions......Page 124 The Architecture of an Atari-Playing Neural Network......Page 129 Digging Deeper into Neural Networks......Page 136 The Mystique of Artificial Intelligence......Page 140 The Automaton Chess Player, or the Turk......Page 141 Misdirection in Neural Networks......Page 143 Recognizing Objects in Images......Page 144 Overfitting......Page 146 ImageNet......Page 148 Convolutional Neural Networks......Page 150 Why “Deep” Networks?......Page 154 Data Bottlenecks......Page 158 Computer-Generated Images......Page 160 Squashing Functions......Page 161 ReLU Activation Functions......Page 163 Android Dreams......Page 166 What It Means for a Machine to “Understand”......Page 172 Deep Speech II......Page 173 Recurrent Neural Networks......Page 174 Generating Captions for Images......Page 179 Long Short-Term Memory......Page 182 Adversarial Data......Page 183 Publicity Stunt or Boon to AI Research?......Page 186 Challenges in Beating Jeopardy......Page 187 Long Lists of Facts......Page 188 The Jeopardy Challenge is Born......Page 190 DeepQA......Page 192 Question Analysis......Page 193 How Watson Interprets a Sentence......Page 195 The Basement Baseline......Page 202 Candidate Generation......Page 204 Searching for Answers......Page 205 Lightweight Filter......Page 208 Evidence Retrieval......Page 209 Scoring......Page 212 Aggregation and Ranking......Page 214 Tuning Watson......Page 217 Was Watson Intelligent?......Page 218 Search for Playing Games......Page 222 Sudoku......Page 223 The Size of the Tree......Page 227 Uncertainty in Games......Page 229 Claude Shannon......Page 233 Evaluation Functions......Page 234 Deep Blue......Page 238 Joining IBM......Page 239 Search and Neural Networks......Page 240 TD-Gammon......Page 241 Limitations of Search......Page 243 Computer Go......Page 244 The Game of Go......Page 246 Sample Moves to Build an Intuition......Page 248 The Hand of God......Page 253 Monte Carlo Tree Search......Page 256 One-Armed Bandits......Page 259 Did AlphaGo Need to Be So Complicated?......Page 261 Limitations of AlphaGo......Page 262 Building Better Gaming Bots......Page 264 StarCraft and AI......Page 265 Simplifying the Game......Page 267 Pragmatic StarCraft Bots......Page 269 OpenAI and DOTA 2......Page 271 The Future of StarCraft Bots......Page 274 The Fits and Starts of AI Development......Page 276 How to Replicate the Successes in this Book......Page 277 Where We Go Next......Page 280 Chapter 2: Self-Driving Cars and the DARPA Grand Challenge......Page 284 Chapter 3: Keeping within the Lanes......Page 286 Chapter 4: Yielding at Intersections......Page 288 Chapter 5: Netflix and the Recommendation-Engine Challenge......Page 290 Chapter 6: Ensembles of Teams......Page 292 Chapter 7: Teaching Computers by Giving Them Treats......Page 294 Chapter 9: Artificial Neural Networks’ View of the World......Page 295 Chapter 10: Looking Under the Hood of Deep Neural Networks......Page 297 Chapter 11: Neural Networks that Can Hear, Speak, and Remember......Page 299 Chapter 12: Understanding Natural Language (and Jeopardy! Questions)......Page 300 Chapter 13: Mining the Best Jeopardy! Answer......Page 302 Chapter 14: Brute-Force Search Your Way to a Good Strategy......Page 304 Chapter 15: Expert-Level Play for the Game of Go......Page 305 Chapter 16: Real-Time AI and StarCraft......Page 307 Chapter 17: Five Decades (or More) from Now......Page 309 Index......Page 310