Probabilistic Programming with Stochastic Probabilities
summary
Claims a compiler and runtime that let expressive probabilistic programs use noisy density estimates while preserving sound inference by design. It introduces a stochastic probability interface with estimate_density (unbiased density estimator) and simulate_with_reciprocal_weight (sample plus unbiased reciprocal weight), supports first‑class marginalization, nesting, and unrestricted control flow, and compiles into randomized estimators and modified MCMC/SMC. Implementation GenSP is evaluated on six case studies, yields automated estimators 1–7 orders of magnitude faster than slow exact densities, ~3× microbenchmark overhead, and convergence largely unaffected.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:24Probabilistic programming systems
- 2:09Expressiveness versus tractability
- 3:11Redesigning computational interface
- 4:33Stochastic probability interface
- 7:21Stochastic density estimates suffice
- 9:29Inference algorithms with new interface
- 10:13Expressive modeling language
- 11:56Marginalization and conditioning
- 13:20Empirical evaluation
- 15:25Key takeaways
- 16:13Q&A
related talks
31mProbabilistic Programming: The What, Why and How
19mIncremental Computation for Efficient Programmable Inference in Probabilistic Programs
01h 03mA Personal Viewpoint on Probabilistic Programming
32mProbabilistic and Differentiable Programming in Scientific Simulators
14mSequential Monte Carlo Program Synthesis with Refinement Proposals
36mProbabilistic Programs Which Make (Common) Sense
29mMotivating Probabilistic Programming
01h 03mAn Overview of Probabilistic Programming