Abstract

Probabilistic modeling and machine learning provide powerful tools for processing ambiguous and noisy information. They play crucial roles in commercial systems ranging from search to autonomous vehicles to user interfaces. Unfortunately, state-of-the-art probabilistic modeling and machine learning can be difficult, involving custom mathematics, custom approximation algorithms, and custom software that all require expert tuning.

I will describe new, stochastic software and hardware abstractions, along with implemented prototypes. These are all based on an emerging marriage of ideas from probability and statistics, specifically stochastic processes, random variables and Bayesian inference, with ideas from computer science, specifically circuits, processors and programs. Throughout, we will abandon the reliance on rock-solid determinism that is central to computer science and engineering.

I will give an overview of three probabilistic computing systems:

1. The Venture probabilistic programming environment. Venture makes it easy to solve state-of-the-art machine learning problems using ~100x less code, and also to build richer, more realistic models. I will show examples in text document modeling and preliminary work in computer vision.
2. Predictive databases, the first commercial, "big data" application of probabilistic computing. These enable users without statistics training to reliably solve data analysis problems using a simple, SQL-like interface. The machine learning backend is based on a 40-line probabilistic program.
3. Stochastic digital circuits and special-purpose inference processors, prototyped on FPGAs. These deliver 1000x speed improvements and 10-100x power savings on stereo vision, motion perception and clustering.

Speaker Bio

Dr. Vikash Mansinghka is an Intelligence Initiative Fellow at MIT's Computer Science and AI Laboratory and the Department of Brain & Cognitive Sciences, where he leads the Probabilistic Computing Project. Dr. Mansinghka 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 Laboratory. 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 startup selling predictive database software that was ultimately acquired by Salesforce.com in 2012. He served on DARPA's Information Science and Technology advisory board from 2010-2012.