Probabilistic Programming with Stochastic Probabilities

Explore a groundbreaking approach to probabilistic programming languages (PPLs) in this 20-minute conference talk from PLDI 2023. Delve into the innovative concept of stochastically estimating probability density ratios for probabilistic inference, which eliminates common restrictions in current PPLs. Learn how this method enables a language supporting first-class constructs for marginalization, nested inference, unrestricted stochastic control flow, continuous and discrete sampling, and programmable inference with custom proposals. Discover the novel technique for compiling expressive probabilistic programs into randomized algorithms for unbiased density estimation. Examine the λSP core calculus used to establish compiler correctness and the implementation in GenSP, an open-source extension to Gen. Evaluate the approach through six challenging inference problems, exploring its ability to automate fast density estimators, maintain inference convergence, and compete with hand-coded estimators.

[PLDI'23] Probabilistic Programming with Stochastic Probabilities