Probabilistic models, particularly those with causal dependencies, can be succinctly written as probabilistic programs. Recent years have seen a proliferation of languages for writing such probabilistic programs, as well as tools and techniques for performing inference over these programs In this talk, we show that static and dynamic program analysis techniques can be used to infer quantities of interest to the machine learning community, thereby providing a new and interesting domain of application for program analysis. In particular, we show that static analysis techniques inspired by dataflow analysis and iterative refinement techniques can be used for Bayesian inference. We also show that dynamic analysis techniques inspired by directed testing and symbolic execution can be used for sampling probabilistic programs (joint work with Aditya Nori, Microsoft Research India).
