A Funny Thing Happened On The Way To Reimplementing AlphaGo
summary
A reimplementation of AlphaZero in Go uses the gorgonia family (a Tensor-like library) and ships a playable engine with MCTS and CUDA flags; distributed training experiments include a synthetic‑gradient inspired scheme that multiplies gradients by network latency predictions (linear regression) which proves unreliable and a particle swarm optimization attempt that works but is slower. Transfer learning and multi‑game training speed up bootstrapping. Feeding the system into itself exposes an infinite pass loop (a halting‑problem–like failure) and training costs on AWS balloon unexpectedly.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:38What is AlphaGo
- 2:03What is Gorgonian
- 10:52Neural Network Architectures
- 12:01The Core
- 12:51The AI
- 14:26Zero
- 15:05Implementation
- 18:00Why Google
- 22:37Synthetic Gradients
- 25:27Transfer Learning
- 28:58Are We Almost There
related talks
34mBuilding a Go AI with Kubernetes and TensorFlow
40mWhat did AlphaGo do to beat the strongest human Go player?
31mMachine Learning for Gophers
48mMachine learning by an imp-op-steur
29mThe Go-to Language for AI: Opportunities and Challenges
50mDeep Learning with Go
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
31mDecoding AI: A Go Programmer's Perspective