CIF: Small: Message Passing Networks

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

The goal of this project is to develop foundations for message-passing algorithms, in the realm of probabilistic graphical models. The project will focus on three classes of problems in graphical models: (a) learning graphical models from observed data, (b) computing the mode of a graphical model, and (c) computing marginals of variables in a graphical model. The project will identify the strengths and limitations of the popular belief propagation algorithm for all three problems. In addition, the project will develop new class of efficient message-passing algorithms with provable performance guarantees by means of exploiting the geometry of the graphical model. Probabilistic graphical models have become a standard way to represent uncertainty succinctly in a wide variety of applications: communication and signal processing, computation, vision and image processing, bioinformatics, natural language processing, and more. The problems faced in these applications are largely The research outcome of this project will be folded in the course titled "Algorithms for Inference (6.438)" taught by the PI. The course is a popular entry level graduate course which also disseminates material online through MIT Open CourseWare and going forward, it may explore possibility of wider dissemination through MITx/EdX.