A Funny Thing Happened On The Way to Reimplementing AlphaGo in Go

A talk given at StrangeLoop 2018 on the experiences reimplementing AlphaZero in Go (the language)

Errata: Makoto works for Mercari, not DeNA

1. A Funny Thing Happened On The Way to Reimplementing AlphaGo in Go Xuanyi Chew @chewxy Strange Loop 2018
2. Why AlphaGo in Go? package αޟ @chewxy
3. @chewxy
4. The People Behind This Project @chewxy Darrell Chua @cfgt Data Scientist OnDeck Gareth Seneque @garethseneque Data Engineer ABC Makoto Ito @ynqa Machine Learning Engineer Mercari Xuanyi Chew @chewxy Chief Data Scientist Ordermentum
5. Why Go? • Many re-implementations of AlphaGo. @chewxy
6. Why Go? • Many re-implementations of AlphaGo. • All in Python and with TensorFlow. @chewxy
7. Why Go? • Many re-implementations of AlphaGo. • All in Python and with TensorFlow. • This is the only known implementation outside Python + TF @chewxy
8. Why Go? • Many re-implementations of AlphaGo. • All in Python and with TensorFlow. • If only there’s a library for deep learning in Go out there… ! @chewxy
9. Gorgonia go get gorgonia.org/gorgonia @chewxy
10. Gorgonia The Gorgonia family of libraries for Deep Learning @chewxy
11. Gorgonia The Gorgonia family of libraries for Deep Learning: • gorgonia.org/gorgonia • gorgonia.org/tensor • gorgonia.org/cu • gorgonia.org/dawson • gorgonia.org/randomkit • gorgonia.org/vecf64 • gorgonia.org/vecf32 @chewxy
12. How does Gorgonia Work? 1. Create an expression graph. @chewxy
13. Neural Networks: A Primer Neural networks are mathematical expressions @chewxy
14. Neural Networks: A Primer Neural networks are mathematical expressions @chewxy
15. Neural Networks: A Primer Neural networks are mathematical expressions @chewxy
16. Neural Networks: A Primer Neural networks are mathematical expressions σ(wx + b) @chewxy
17. Neural Networks: A Primer Neural networks are mathematical expressions σ(wx + b) Linear transformations Non-linear transformation @chewxy
18. Neural Networks: A Primer Neural networks are mathematical expressions σ(wx + b) Learnable Input @chewxy
19. Neural Networks: A Primer Neural networks are mathematical expressions σ(wx + b) Learnable @chewxy
20. Neural Networks: Backpropagation ŷ = σ(wx + b) @chewxy
21. Neural Networks: Backpropagation ŷ = σ(wx + b) @chewxy z = y - ŷ
22. Neural Networks: Backpropagation ŷ = σ(wx + b) @chewxy z = y - ŷ Smaller is better
23. Neural Networks: Backpropagation ŷ = σ(wx + b) @chewxy z = y - ŷ Change w and b such that z is minimal
24. Neural Networks: Backpropagation ŷ = σ(wx + b) @chewxy Change w and b such that z is minimal dz d! y =0
25. Neural Networks: Backpropagation Backpropagation finds gradients ŷ = σ(wx + b) @chewxy Partial derivatives – how much to change w and b ∂z ∂w ∂z ∂b
26. Neural Networks: Gradient updates ŷ = σ(wx + b) @chewxy new w = w+ ∂z ∂w new b = b+ ∂z ∂b
27. Neural Networks: A Primer Neural networks are mathematical expressions σ(wx + b) @chewxy
28. Neural Networks: A Primer Neural networks are mathematical expressions σ(add(mul(w, x), b))) @chewxy
29. Neural Networks: A Primer Neural networks are mathematical expressions @chewxy x w mul add σ b
30. How does Gorgonia Work? 1. Create an expression graph. 2. Populate the expression graph with values. @chewxy x = 1 w = 2 mul add σ b = 3
31. How does Gorgonia Work? 1. Create an expression graph. 2. Populate the expression graph with values. 3. Walk towards the root. @chewxy x = 1 w = 2 mul add σ b = 3
32. Convolution @chewxy X X O X O X O Board game positions
33. Convolution @chewxy X X O X O X O 0 0 0 0 1 0 0 0 0 Board game positions
34. Convolution @chewxy 1 1 -1 1 -1 1 -1 0 0 0 0 1 0 0 0 0 Board game positions – Represented as a matrix
35. Convolution @chewxy 0 0 1 0 0 0 1 -1 1 0 0 0 0 0 0 0 -1 0 0 0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 Board game positions – Represented as a matrix
36. Convolution @chewxy 0*0 0*0 1*0 0 0 0*0 1*1 -1*0 1 0 0*0 0*0 0*0 0 0 0 -1 0 0 0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 1
37. Convolution @chewxy 0*0 0*0 1*0 0 0 0*0 1*1 -1*0 1 0 0*0 0*0 0*0 0 0 0 -1 0 0 0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 1 0*0 + 0*0 + 1*0 + 0*0 + 1*1 + -1*0 + 0*0 + 0*0 + 0*0 = 1
38. Convolution @chewxy 0 0*0 1*0 0*0 0 0 1*0 -1*1 1*0 0 0 0*0 0*0 0*0 0 0 -1 0 0 0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 1 -1
39. Convolution @chewxy 0 0 1*0 0*0 0*0 0 1 -1*0 1*1 0*0 0 0 0*0 0*0 0*0 0 -1 0 0 0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 1 -1 1
40. Convolution @chewxy 0 0 1 0 0 0*0 1*0 -1*0 1 0 0*0 0*1 0*0 0 0 0*0 -1*0 0*0 0 0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 1 -1 1 0
41. Convolution @chewxy 0 0 1 0 0 0 1*0 -1*0 1*0 0 0 0*0 0*1 0*0 0 0 -1*0 0*0 0*0 0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 1 -1 1 0 0
42. Convolution @chewxy 0 0 1 0 0 0 1 -1*0 1*0 0*0 0 0 0*0 0*1 0*0 0 -1 0*0 0*0 0*0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 1 -1 1 0 0 0
43. Convolution @chewxy 0 0 1 0 0 0 1 -1 1 0 0*0 0*0 0*0 0 0 0*0 -1*1 0*0 0 0 0*0 1*0 -1*0 0 0 0 0 0 0 1 0 0 0 0 1 -1 1 0 0 0 -1
44. Convolution @chewxy 0 0 1 0 0 0 1 -1 1 0 0 0*0 0*0 0*0 0 0 -1*0 0*1 0*0 0 0 1*0 -1*0 0*0 0 0 0 0 0 1 0 0 0 0 1 -1 1 0 0 0 -1 0
45. Convolution @chewxy 0 0 1 0 0 0 1 -1 1 0 0 0 0*0 0*0 0*0 0 -1 0*0 0*1 0*0 0 1 -1*0 0*0 0*0 0 0 0 0 1 0 0 0 0 1 -1 1 0 0 0 -1 0 0
46. Convolution – Some Nuance @chewxy 0 0 1 0 0 0 1 -1 1 0 0 0 0 0 0 0 -1 0 0 0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 1 -1 1 0 0 0 -1 0 0 Unpadded convolution
47. Convolution – Some Nuance @chewxy 0 0 1 0 0 0 1 -1 1 0 0 0 0 0 0 0 -1 0 0 0 0 1 -1 0 0 0 0 0 0 1 0 0 0 0 1 -1 1 0 0 0 -1 0 0 Identity Kernel
48. Neural Networks: A Primer Neural networks are mathematical expressions σ(wx + b) Linear transformations Non-linear transformation @chewxy
49. Neural Networks: A Primer Neural networks are mathematical expressions σ(x∗w) Linear transformations Non-linear transformation @chewxy
50. Deep Neural Network Architectures Deep neural networks are formed by many layers. @chewxy
51. Deep Neural Network Architectures Deep neural networks are formed by many layers. @chewxy Fully Connected Layer Convolution Layer Prediction Input
52. Deep Neural Network Architectures Deep neural networks are formed by many layers. @chewxy Fully Connected Layer Convolution Layer Prediction Input Many layers in between
53. Residual Network @chewxy Fully Connected Layer Convolution Layer Prediction Input Fully Connected Layer
54. Residual Network @chewxy Fully Connected Layer Convolution Layer Prediction Input Fully Connected Layer +
55. AlphaZero @chewxy
56. How does AlphaZero Work? AlphaZero is comprised of two components and a set of training rules. @chewxy
57. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece @chewxy
58. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece @chewxy Residual Layers Convolution Layers Policy Value Input
59. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece @chewxy Residual Layers Convolution Layers Policy Value Input
60. Two Components of AlphaGo? • Neural network detects patterns on the game board and makes decisions on where to best place a piece @chewxy 0.1 0.1 0.1 0.1 0.2 0.1 0.1 0.1 ... Policy Residual Layers Convolution Layers Policy Value Input
61. How does AlphaGo Work? • Neural network detects patterns on the game board and makes decisions on where to best place a piece @chewxy 0.1 0.1 0.1 0.1 0.2 0.1 0.1 0.1 ... Policy Residual Layers Convolution Layers Policy Value Input 0.8 Value
62. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play @chewxy
63. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play @chewxy
64. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play @chewxy 0 1 2 3 4 5 6 7 8
65. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play 0 1 2 3 4 5 6 7 8 @chewxy X O
66. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play 0 1 2 3 4 5 6 7 8 @chewxy X O
67. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play 0 1 2 3 4 5 6 7 8 @chewxy X O 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.8 Policy Value
68. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play 0 1 2 3 4 5 6 7 8 @chewxy X O 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.8 Policy Value
69. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play 0 1 2 3 4 5 6 7 8 @chewxy X O? O 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.8 Policy Value
70. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play 0 1 2 3 4 5 6 7 8 @chewxy X O O X 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.8 Policy Value
71. Two Components of AlphaGo • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play @chewxy
72. What AlphaGo Does • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play • Take action @chewxy
73. Zero in AlphaZero @chewxy
74. AlphaZero AlphaZero is AlphaGo without training data from humans. @chewxy
75. AlphaZero AlphaZero is AlphaGo without training data from humans. 1. Self-play creates training data. @chewxy
76. AlphaZero AlphaZero is AlphaGo without training data from humans. 1. Self-play creates training data. 2. Train on self-play data. @chewxy
77. AlphaZero AlphaZero is AlphaGo without training data from humans. 1. Self-play creates training data. 2. Train on self-play data. 3. Pit old version of AlphaZero neural network vs new version. @chewxy
78. AlphaZero AlphaZero is AlphaGo without training data from humans. 1. Self-play creates training data. 2. Train on self-play data. 3. Pit old version of AlphaZero neural network vs new version. 4. Goto 1. @chewxy
79. The Implementation @chewxy
80. Neural network is simple The neural network for AlphaZero is simple. @chewxy
81. How Simple Is It? @chewxy
82. Algorithm is simple AlphaZero’s algorithm is also conceptually simple @chewxy
83. @chewxy
84. Running It Is Equally Simple @chewxy
85. Running It Is Equally Simple $ go run *.go @chewxy
86. Running It Is Equally Simple $ go run –tags=cuda *.go @chewxy
87. Live Demo (you had to be there)
88. What’s Hard? @chewxy
89. What’s Hard? • MCTS @chewxy
90. What’s Hard? • MCTS • High performance / pointer free MCTS @chewxy
91. What’s Hard? • MCTS • Go (the game) @chewxy
92. What’s Hard? • MCTS • Go (the game) • It helps if you actually know the game @chewxy
93. What’s Hard? • MCTS • Go (the game) • Training @chewxy
94. What’s Hard? • MCTS • Go (the game) • Training • Uses A LOT of memory (GPU and normal RAM) @chewxy
95. What’s Hard? • MCTS • Go (the game) • Training • Uses A LOT of memory (GPU and normal RAM) • Distributed training = more headache @chewxy
96. Neural Networks: A Primer Neural networks are mathematical expressions σ(wx + b) Learnable Input @chewxy
97. Distributed Training σ(wx + b) @chewxy
98. σ(wx + b) σ(wx + b) σ(wx + b) Distributed Training σ(wx + b) @chewxy
99. Distributed Training σ(wx + b) @chewxy σ(wx + b) σ(wx + b) σ(wx + b)
100. Distributed Training σ(wx + b) @chewxy σ(wx + b) σ(wx + b) σ(wx + b) Gradient Gradient Gradient Gradient
101. Distributed Training σ(wx + b) @chewxy σ(wx + b) σ(wx + b) σ(wx + b) Gradient Gradient Gradient Gradient Parameter Server
102. Distributed Training σ(wx + b) @chewxy σ(wx + b) σ(wx + b) σ(wx + b) Gradient Gradient Gradient Gradient Parameter Server
103. Distributed Training σ(wx + b) @chewxy σ(wx + b) σ(wx + b) σ(wx + b) new W, b new W, b new W, b new W, b Parameter Server
104. Distributed Training σ(wx + b) @chewxy σ(wx + b) σ(wx + b) σ(wx + b) new W, b new W, b new W, b new W, b Parameter Server
105. Distributed Training σ(wx + b) @chewxy σ(wx + b) σ(wx + b) σ(wx + b) new W, b new W, b new W, b new W, b Parameter Server
106. @chewxy
107. And then, something funny happened… @chewxy
108. I Stopped Caring About AlphaGo At least, the Go playing part. @chewxy
109. Interesting Questions and Outcomes • How to improve training speed? @chewxy
110. Interesting Questions and Outcomes • How to improve training speed? • Better training and optimization methodologies. @chewxy
111. Optimization • Distributed Training with Synthetic Gradients @chewxy
112. Distributed Training with Synthetic Gradients • Use between-server communication latency as noise for synthetic gradients. @chewxy
113. Optimization • Distributed Training with Synthetic Gradients • Particle Swarm Optimization @chewxy
114. Optimization • Distributed Training with Synthetic Gradients • Particle Swarm Optimization • Coming Soon to Gorgonia @chewxy
115. Interesting Questions and Outcomes • How to improve training speed? • Better training and optimization methodologies. • What is the goal? @chewxy
116. Interesting Questions and Outcomes • How to improve training speed? • Better training and optimization methodologies. • What is the goal? • Play Go well @chewxy
117. Interesting Questions and Outcomes • How to improve training speed? • Better training and optimization methodologies. • What is the goal? • Play Go well • Take a cue from transfer learning @chewxy
118. Transfer Learning– An Analogy Neural Network : Program :: Transfer Learning : Refactoring @chewxy
119. Transfer Learning @chewxy Some Other Layers Convolution Layers Prediction Input A Task A
120. Transfer Learning @chewxy Residual Layers Convolution Layers Policy Value Input B Task B (AlphaZero)
121. Transfer Learning @chewxy Residual Layers Convolution Layers Policy Value Input B copied over Some Other Layers Convolution Layers Prediction Input A
122. Transfer Learning @chewxy Residual Layers Convolution Layers Policy Value Input B Only train these parts
123. Interesting Questions and Outcomes • How to improve training speed? • Better training and optimization methodologies. • What is the goal? • Play Go well • Take a cue from transfer learning • Multi-task learning @chewxy
124. Multi-task Learning • What if the AlphaGo neural network learned various games all at once? @chewxy
125. Multi-task Learning @chewxy Residual Layers Convolution Layers Policy Value M,N,K Residual Layers Convolution Layers Policy Value Komi Residual Layers Convolution Layers Policy Value Connect4 Shared Avg Shared Avg Shared Avg Shared Avg
126. Multi-task Learning @chewxy Residual Layers Convolution Layers Policy Value Komi M,N,K Connect4
127. Interesting Questions and Outcomes • How to improve training speed? • Better training and optimization methodologies. • What is the goal? • Play Go well • Take a cue from transfer learning • Multi-task learning • What is AlphaGo good for? @chewxy
128. Interesting Questions and Outcomes • How to improve training speed? • Better training and optimization methodologies. • What is the goal? • Play Go well • Take a cue from transfer learning • Multi-task learning • What is AlphaGo good for? • Solving problems with large search spaces @chewxy
129. Interesting Questions and Outcomes • How to improve training speed? • Better training and optimization methodologies. • What is the goal? • Play Go well • Take a cue from transfer learning • Multi-task learning • What is AlphaGo good for? • Solving problems with large search spaces • Drug discovery @chewxy
130. Interesting Questions and Outcomes • How to improve training speed? • Better training and optimization methodologies. • What is the goal? • Play Go well • Take a cue from transfer learning • Multi-task learning • What is AlphaGo good for? • Solving problems with large search spaces • Drug discovery • Neural network weights? @chewxy
131. Feeding AlphaGo Into Itself What if you used AlphaGo as an input to AlphaGo? @chewxy
132. The Big Question Is AlphaGo putting us on the right path to building an AGI? @chewxy
133. How Close Is AlphaGo to The Big Picture Goal? Ability To Humans AlphaGo Understand cause and effect ✓ Compute ✓ Tackle a diverse array of causal computation problems ✓ @chewxy
134. Causal Reasoning A Causal Reasoner Can: • See patterns • Interfere and take actions • Imagine alternative scenarios @chewxy
135. How does AlphaGo Work? • Neural network detects patterns on the game board and makes decisions on where to best place a piece • Monte-carlo tree search for best play • Take action @chewxy
136. Is AlphaGo a Causal Reasoner? Causal Reasoner • See patterns AlphaGo • Convolutional neural network @chewxy
137. Is AlphaGo a Causal Reasoner? Causal Reasoner • See patterns • Imagine alternative scenarios AlphaGo • Convolutional neural network • Monte-carlo tree search @chewxy
138. Is AlphaGo a Causal Reasoner? Causal Reasoner • See patterns • Imagine alternative scenarios • Interfere and take actions AlphaGo • Convolutional neural network • Monte-carlo tree search • Take action @chewxy
139. How Close Is AlphaGo to The Big Picture Goal? Ability To Humans AlphaGo Understand cause and effect ✓ ✓* Compute ✓ Tackle a diverse array of causal computation problems ✓ @chewxy *Contra Judea Pearl
140. Is AlphaGo Recursive? What if you used AlphaGo as an input to AlphaGo? @chewxy
141. Is AlphaGo Recursive? What if you used AlphaGo as an input to AlphaGo? • Is it possible to build a variant that will recurse and never stop? @chewxy
142. @chewxy
143. @chewxy
144. How Close Is AlphaGo to The Big Picture Goal? Ability To Humans AlphaGo Understand cause and effect ✓ ✓* Can compute ✓ ??? Tackle a diverse array of causal computation problems ✓ @chewxy *Contra Judea Pearl
145. Multi-task Learning @chewxy Residual Layers Convolution Layers Policy Value Komi M,N,K Connect4
146. Multi-task Learning – Currently Playing @chewxy Residual Layers Convolution Layers Policy Value Komi M,N,K Connect4 Attn
147. How Close Is AlphaGo to The Big Picture Goal? Ability To Humans AlphaGo Understand cause and effect ✓ ✓* Compute ✓ ??? Tackle a diverse array of causal computation problems ✓ Possible @chewxy *Contra Judea Pearl
148. Closing Thoughts @chewxy
149. https://github.com/gorgonia/agogo @chewxy
150. Certainty of death. Small chance of success. What are we waiting for? @chewxy
151. Thank You Twitter: @chewxy Email: chewxy@gmail.com Play with Gorgonia: go get gorgonia.org/gorgonia @chewxy