Effect Handlers for Programmable Inference
summary
Claims a Haskell library that uses algebraic effects and effect handlers to represent models as trees of sample/observe operations and to express inference patterns as their own abstract operations. Implements Metropolis–Hastings variants (independence MH with very low acceptance, single-site Metropolis with much higher acceptance) and particle filters (multinomial PF with 100 particles that collapses to one high-weight particle, and resample-move particle filter that applies MH per particle and yields richer, more diverse traces). Handlers operate by interpreting traces, weights, suspend/resume and resampling.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:42Inferring Missing Data
- 4:48Algebraic Effects
- 8:08Framework
- 10:34Inference
- 12:57Independence Metropolis
- 15:56Particle Filters
- 18:13Multinomial Particle Filter
- 20:22Resample Move Particle Filter
- 22:29Recap
- 22:48Questions
related talks
15mModular Probabilistic Models via Algebraic Effects
01h 17mSmart Handlers: Handling the selection monad
17mMachine Learning Primitives as Algebraic Effects
33mRethinking Effects with Effect Systems
39mDesigning software with effects systems
25mHandling Higher-Order Effectful Operations with Judgemental Monadic Laws
27mWith or Without You: Programming with Effect Exclusion
50mFunctional Programming with Effects