Learning Stochastic Inverses for Adaptive Inference in Probabilistic Programs

We describe an algorithm for adaptive inference in probabilistic programs. Dur-ing sampling, the algorithm accumulates information about the local probability distributions that compose the program’s overall distribution. We use this informa-tion to construct targeted samples: given a value for an intermediate expression, we stochastically invert each of the steps giving rise to this value, sampling back-wards in order to assign values to random choices such that we get a likely parse of the intermediate value. We propose this algorithm for importance sampling and as a means of constructing blocked proposals for a Metropolis-Hastings sampler.