Probabilistic Programming: Generative Probabilistic Graphics Programming and New Research Directions

Опубликовано 8 августа 2016, 19:36

Probabilistic programming has recently attracted much attention in Computer Science and Machine Learning communities. I will demonstrate two generative probabilistic graphics programs (models), which I contributed to develop. The first can read text from simple CAPTCHAs, and the second can find roads from real-world images. Both work by performing approximate inference over the executions of simple renderers, using the general-purpose Metropolis-Hastings inference engine built into a probabilistic programming system. I will briefly touch on two other research directions I am interested in pursuing: a path to scaling up general-purpose approximate inference in probabilistic programs using parallelism, based on preliminary work on a multithreaded approximate MCMC scheme for Church-like languages, and a much longer-term path to automatic programming via general-purpose approximate inference. This is based on joint work with Vikash Mansinghka, Tejas Kulkarni, and Joshua Tenenbaum, especially "Approximate Bayesian Image Interpretation using Generative Probabilistic Graphics Programs", arxiv.org/abs/1307.0060