Stochastic Lazy Knowledge Compilation for Inference in Discrete Probabilistic Programs
summary
Implements lazy knowledge compilation in a Julia tool called Pluck and combines it with sequential Monte Carlo that uses locally‑optimal proposals computed from partial compilations; adds a guarded caching scheme to avoid needless recompilation. Claims exact inference when possible, an approximate SMC fallback, and reproducible artifacts. On structured‑data benchmarks (perturb, PCFGs, sorted lists) it cuts runtime versus strict compilation by up to orders of magnitude (reported 1,000–10,000×), and handles many infinite‑execution cases without manual recursion bounds.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:21Pluck programs
- 3:11Laziness improves inference
- 8:32Knowledge compilation basics
- 9:58Combining laziness with compilation
- 12:32Guarded caching
- 13:15Structured data models
- 15:01Performance vs Aliia
- 16:54Scaling with observations
- 18:01Infinite distributions
- 19:07Implementation details
- 20:10Q&A
related talks
19mIncremental Computation for Efficient Programmable Inference in Probabilistic Programs
14mSequential Monte Carlo Program Synthesis with Refinement Proposals
35mScalable structure learning and inference for domain-specific probabilistic prog
14mMulti-Language Probabilistic Programming
15mScaling Optimization Over Uncertainty via Compilation
36mProbabilistic scripts for automating common-sense tasks
15mScaling Exact Inference for Discrete Probabilistic Programs
47mApproximate Lifted Inference with Probabilistic Databases