Abstract

Probabilistic modeling and Bayesian inference provide a unifying theoretical framework for uncertain reasoning. They have become central tools for engineering machine intelligence, modeling human cognition, and analyzing structured and unstructured data. However, they often seem far less unified, complete and expressive in practice than they are in theory, and can require significant interdisciplinary expertise to apply. Domains such as robotics and statistics involve diverse modeling idioms, speed/accuracy requirements, dataset sizes, and approximation techniques. Inference in simple latent variable models can be computationally challenging, while state-of-the-art models do not fit within standard formalisms and can be cumbersome to specify, let alone use.

In this talk, I will describe probabilistic computing systems that address several of these challenges and that fit together into a mathematically coherent software and hardware stack for Bayesian inference and intelligent computation.

I will focus on Venture, a new, Turing-complete probabilistic programming platform descended from the Church probabilistic programming language. In Venture, models are represented by executable code, with random choices corresponding to latent variables. Inference from data is done via automatic but reprogrammable mechanisms that cover a broad class of approximation strategies, including novel hybrids of Markov chain, sequential Monte Carlo and variational techniques. I will describe applications in text analysis, high-dimensional statistics and computer vision that yield a 100x savings in lines of code versus standard approaches.