Machine Learning Primitives as Algebraic Effects
summary
Claims that Haskell encodes ML stages and primitives as composable effects, so automatic differentiation as an algebraic effect and parameter initialization, regularization, noise, hyper‑parameter nudging become tracked in row types and handled differently for training vs evaluation. Demonstrates a working toy library (CPU backend, no benchmarks yet) that composes models by merging effect rows and uses writer-like logs for differentiable regularization; paper and polished library are still in progress.
AI-written summary · about
chapters (9)
- 0:00ML with algebraic effects
- 3:25Naive ML model composition
- 5:54Traditional ML composition limits
- 7:19Extensible effects for ML
- 9:24Effects in training and inference
- 10:28Regularization and Hamilton Bridge
- 12:15ML terms to Haskell effects
- 14:03Future work and release
- 15:05Q&A
related talks
47mIntroduction to Algebraic Effects
01h 09mLatent Effects and Handlers + Automatic Differentiation via Effects and Handlers
15mModular Probabilistic Models via Algebraic Effects
28mA Series of Unfortunate Effects
29mAutomatic Differentiation via Effects and Handlers in OCaml
44mEffects, Algebraically yours ..
01h 06mThe Evolution of Effects
50mFunctional Programming with Effects