Incremental Density Computation for Efficient Programmable Inference
summary
Presents a concrete workflow: compute trace types from the probabilistic program, use type-directed change types to represent insertions/removals/replacements, compile the generative model into a deterministic density program, then automatic incrementalization yields an updater that maps trace changes to density updates while caching intermediates. Prototype implementation compares against nonincremental baselines and existing incremental methods and reports that this approach is the only one that consistently achieves asymptotic scaling in data size for the tested Monte Carlo kernels.
AI-written summary · about
chapters (12)
- 0:00Incremental density computation intro
- 0:24Latent cluster inference example
- 1:52Density function and inference
- 2:58Naive density evaluation challenges
- 4:04Changes in inference programs
- 5:10Automating incremental updates
- 5:51Trace and change types
- 7:38How trace and change types work
- 9:02Two-stage transformation
- 10:07Incremental updaters and caching
- 11:11Prototype and results
- 11:34Q&A
related talks
19mIncremental Computation for Efficient Programmable Inference in Probabilistic Programs
20mProbabilistic Programming with Stochastic Probabilities
26mOptimising Density Computations in Probabilistic Programs via Automatic Loop Vectorisation
21mGuaranteed inference for probabilistic programs: a parallelisable, small-step o...
31mProbabilistic Programming: The What, Why and How
11mExact Inference for Discrete Probabilistic Programs via Generating Functions
47mApproximate Lifted Inference with Probabilistic Databases
01h 03mA Personal Viewpoint on Probabilistic Programming