• Home
• Staff & Students
• Vacancies

Vikas Mansinghka

 

http://web.mit.edu/vkm/www/

 

(MIT)

 

Wednesday 22 February 2012

16:00

 

B10 Seminar Room, Basement,

Alexandra House, 17 Queen Square, London, WC1N 3AR

 

 

How to be stochastic about probabilistic programming, and why it is worth the trouble

 

Over the last 5 years, a host of probabilistic programming systems have been developed, including infer.net, Church, BLOG, and Figaro. These systems all make it possible to build rich, realistic models in terms of code that simulates from a prior stochastic process.
They also all attempt to make Bayesian inference in these models easier and, in some cases, fully automatic.

The stochastic language Church represents an unusual point in this space. Most probabilistic language designers have consciously chosen to restrict expressiveness in pursuit of greater computational tractability, often by reduction to deterministic approximate inference, such as variational message passing or generalizations of variable elimination. In contrast, Church was designed to be maximally expressive and consistently stochastic. It can conditionally execute arbitrary stochastic processes and generate posterior samples, including descriptions of the Church algorithm performing inference over itself, but does not give any direct means of calculating even prior probabilities.

In this talk, I will motivate and illustrate some of these unusual choices. I will argue that they yield an immediate practical payoff, build new bridges between the theory of statistics and computer science, and represent a path towards AI systems that enjoy both greater flexibility and tractability.

I will describe uses of Church that require its unusual features, including model selection over an ODE simulator applied to pharmacokinetic data from a major pharmaceutical company. I will show how embracing stochasticity in the semantics of the language both enables natural support for nonparametric models and builds a bridge between distributed systems, the de Finetti theorem and some of its generalizations, and basic issues in programming language design. I will also sketch the main techniques involved in implementing stochastic inference for Church efficiently, and present some evidence that this approach can also lead to improved inference convergence.

BIO: Vikash Mansinghka is a postdoctoral fellow at MIT, where he runs the Probabilistic Computing Project, and is also a visiting fellow at Harvard. He received an SB in Mathematics, and SB in Computer Science, an MEng in Computer Science, and a PhD in Computation, all from MIT. He held graduate fellowships from the National Science Foundation and MIT's Lincoln Laboratories. His PhD dissertation on natively probabilistic computation won the 2009 MIT George M. Sprowls award for best dissertation in computer science. He previously co-founded and led a venture-backed start-up selling software for Bayesian data analysis. He currently serves on DARPA's Information Science and Technology (ISAT) advisory board.