CS-2022-01
Title Synthesizing Guide Programs for Sound, Effective Deep Amortized Inference
Authors Jianlin Li, Leni Ven, Pengyuan Shi and Yizhou Zhang
Abstract In probabilistic programming languages (PPLs), a critical step in optimization-based inference methods is constructing, for a given model program, a trainable guide program. Soundness and effectiveness of inference rely on constructing good guides, but the expressive power of a universal PPL poses challenges.

This paper introduces an approach to automatically generating guides for deep amortized inference in a universal PPL. Guides are generated using a type-directed translation per a novel behavioral type of system. Guide generation extracts and exploits independence structures using a syntactic approach to conditional independence, with a semantic account left to further work. Despite the control-flow expressiveness allowed by the universal PPL, generated guides are guaranteed to satisfy a critical soundness condition and, moreover, consistently improve training and inference over state-of-the-art baselines for a suite of benchmarks.
Date December 27, 2022
Report CS-2022-01 (PDF)