Haskell for choice-based learning
summary
Claims that choice-based learning is implementable by combining algebraic effects and handlers with the selection monad, so optimizers become handlers that receive a loss-continuation. Provides a Haskell prototype library that encodes choices (choose, loss) and shows concrete implementations: SGD as a handler differentiating the provided loss, nested hyperparameter search, GANs as opposing handlers, and Q‑learning producing a learned policy (edge‑hugging vs safer paths). Works as a proof‑of‑concept; plan to target JAX for real workloads.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:24Machine learning challenges
- 5:41Choice-based learning
- 7:06MDP learning example
- 13:10Design with algebraic effects
- 14:33Algebraic effects basics
- 23:01Training with effect ops
- 29:23Selection monad role
- 33:20Hyperparameter optimization
- 35:27Generative adversarial networks implementation
- 36:50Reinforcement learning
- 42:48Language design summary
- 43:52Q&A
related talks
14mEffect Handlers for Choice-Based Learning
17mMachine Learning Primitives as Algebraic Effects
01h 09mLatent Effects and Handlers + Automatic Differentiation via Effects and Handlers
53mOff-policy Policy Optimization
47mIntroduction to Algebraic Effects
59mA Welcome to Causal Probabilistic Programming
13mTyped Abstractions for Causal Probabilistic Programming
39mA PAC Approach to Application-Specific Algorithm Selection