Composing Inference Algorithms
Rob Zinkov
Abstract
Probabilistic programs while, allowing many models to be easily expressed are notorious difficult to write inference algorithms for them. Worse still, it is difficult to extend existing a Probabilistic Programming System with a new inference algorithm. In this talk, I will show inference algorithms can be expressed as program transformations from probabilistic programs to probabilistic programs. This enables defining inference procedures in a more modular and simpler way. The resulting programs, after these transformations, have a running time comparable to that of a hand-written inference procedure.