Multi-Language Probabilistic Programming
summary
Introduces MultiPPL, a multi-language that stitches an exact-discrete engine (DICE / BDD knowledge compilation) to an approximate sampler (K with importance weights) via explicit boundary expressions, proves soundness by refining a high-level denotation to low-level inference models, implements a reference compiler, and evaluates trade-offs: reports L1 sample-quality gains and wall-clock cost per 1,000 samples, shows boundary placement matters (heuristic balances accuracy vs time), and publishes reproducible artifacts.
AI-written summary · about
chapters (10)
- 0:00Multi-language probabilistic programming
- 1:01Limitations of DICE
- 2:50Motivation for DICE and Pyro
- 3:32Multi-language probabilistic programming
- 4:54Interoperation between DICE and KA
- 6:22Sound interoperation invariants
- 7:52Importance weighting in multi-language
- 9:17Soundness proof and semantics
- 11:03Compiler and evaluation results
- 12:13Q&A
related talks
15mScaling Exact Inference for Discrete Probabilistic Programs
35mScalable structure learning and inference for domain-specific probabilistic prog
25mStatically and Dynamically Delayed Sampling for Typed Probabilistic Programming Languages
22mStochastic Lazy Knowledge Compilation for Inference in Discrete Probabilistic Programs
09mMultiSPPL: extending SPPL with multivariate leaf nodes
01h 03mAn Overview of Probabilistic Programming
32mProbabilistic and Differentiable Programming in Scientific Simulators
31mProbabilistic Programming: The What, Why and How