Inference of Probabilistic Programs with Moment-Matching Gaussian Mixtures

Inference of Probabilistic Programs with Moment-Matching Gaussian Mixtures

Explore a 20-minute video presentation from POPL 2024 conference introducing Gaussian Semantics, a novel approach for approximating probabilistic program semantics using Gaussian mixtures. Learn about the universal approximation theorem and a second-order Gaussian approximation (SOGA) method for matching moments analytically. Discover how this technique provides accurate estimates for complex models, outperforming alternative methods in collaborative filtering and programs with mixed continuous and discrete distributions. Examine case studies demonstrating SOGA's improved accuracy and computational efficiency compared to existing techniques.