[PLDI'23] Probabilistic Programming with Stochastic Probabilities

Probabilistic Programming with Stochastic Probabilities (Video, PLDI 2023) Alexander K. Lew, Matin Ghavamizadeh, Martin C. Rinard, and Vikash K. Mansinghka (Massachusetts Institute of Technology, USA; Massachusetts Institute of Technology, USA; Massachusetts Institute of Technology, USA; Massachusetts Institute of Technology, USA) Abstract: We present a new approach to the design and implementation of probabilistic programming languages (PPLs), based on the idea of stochastically estimating the probability density ratios necessary for probabilistic inference. By relaxing the usual PPL design constraint that these densities be computed exactly, we are able to eliminate many common restrictions in current PPLs, to deliver a language that, for the first time, simultaneously supports first-class constructs for marginalization and nested inference, unrestricted stochastic control flow, continuous and discrete sampling, and programmable inference with custom proposals. At the heart of our approach is a new technique for compiling these expressive probabilistic programs into randomized algorithms for unbiasedly estimating their densities and density reciprocals. We employ these stochastic probability estimators within modified Monte Carlo inference algorithms that are guaranteed to be sound despite their reliance on inexact estimates of density ratios. We establish the correctness of our compiler using logical relations over the semantics of λSP, a new core calculus for modeling and inference with stochastic probabilities. We also implement our approach in an open-source extension to Gen, called GenSP, and evaluate it on six challenging inference problems adapted from the modeling and inference literature. We find that: (1)  ‍can automate fast density estimators for programs with very expensive exact densities; (2) convergence of inference is mostly unaffected by the noise from these estimators; and (3) our sound-by-construction estimators are competitive with hand-coded density estimators, incurring only a small constant-factor overhead. Article: https://doi.org/10.1145/3591290 ORCID: https://orcid.org/0000-0002-9262-4392, https://orcid.org/0000-0003-3052-7412, https://orcid.org/0000-0001-8095-8523, https://orcid.org/0000-0003-2507-0833 Video Tags: probabilistic programming, semantics, approximate computing, pldi23main-p520-p, doi:10.1145/3591290, orcid:0000-0002-9262-4392, orcid:0000-0003-3052-7412, orcid:0000-0001-8095-8523, orcid:0000-0003-2507-0833 Presentation at the PLDI 2023 conference, June 17–21, 2023, https://pldi23.sigplan.org/ Sponsored by ACM SIGPLAN,