Probabilistic computing for Bayesian inference

Although probabilistic modeling and Bayesian inference provide a unifying theoretical framework for uncertain reasoning, they can be difficult to apply in practice. Inference in simple models can seem intractable, while more realistic, flexible models can be difficult to specify, let alone implement correctly. My talk will describe three prototype probabilistic computing systems --- including probabilistic programming languages, a Bayesian database system, and intentionally stochastic hardware --- designed to mitigate these challenges. I will focus on Venture, a new, open-source, Turing-complete probabilistic programming platform that aims to be sufficiently expressive, extensible and efficient for general-purpose use. Venture programmers specify models via executable code, where random choices correspond to latent variables; this approach can yield a 100x reduction in code size for state-of-the-art models. Multiple scalable, computationally universal inference algorithms are provided, based on MCMC, conditional SMC and mean field variational techniques. Unlike probabilistic programming tools like BUGS or Church, users can also reprogram the inference strategy for each application and easily implement novel approximation schemes. I will review applications of Venture to text modeling with millions of observations; 2D and 3D computer vision problems; and structured inverse problems in geophysics. I will also touch on the ways the ideas behind Venture fit together into a mathematically coherent software and hardware stack for Bayesian inference. This talk includes joint work with Daniel Roy, Eric Jonas, Daniel Selsam and Yura Perov.Non UBCUnreviewedAuthor affiliation: Massachusetts Institute of TechnologyOthe