CAREER: Undirected Bipartite Graphical Models

University Of California-Irvine, Irvine CA

Investigators

Abstract

Modern society increasingly relies on processing, storing and communicating large amounts of information. The exponential growth of databases necessitates the development of algorithms that structure, compress and query them efficiently. The goal of this research is to develop new tools based on probabilistic models to achieve these objectives. A new class of "undirected bipartite graphical models" is studied that embeds documents into a low dimensional "topic-space". These representations are efficient and capture semantic relationships. Training of the underlying probabilistic model is achieved through a technique called "contrastive divergence learning" which is particularly well adapted to the UBG model. The main contributions of this research are the development of a new class of probabilistic graphical model, the development of improved learning algorithms that scale up to large data-sets and the application of these novel techniques to two real world applications: image restoration and information retrieval. Research is integrated with teaching through the development of new classes in machine learning both at the undergraduate and the graduate level, where students will be engaged in research in the above application areas. The proposed research makes important contributions that can have a broad impact on security, web-technologies, commerce, multi-media, medical expert systems etc. In particular, the proposed projects in image restoration and information retrieval have the potential to make an impact on tomorrow's technologies. An open-source, web-based repository with freely available software will be developed to help achieve that goal. http://www.ics.uci.edu/~welling/NSFcareer/NSFcareer.html