Graphical models are used to organize and structure probability distributions over large systems, and enable efficient approximate or exact reasoning. My group balances developing theoretical and algorithmic advances with applications to the real-world systems of our collaborators.
Algorithms. One of our main focuses is on finding maxima or computing probabilities using variational methods, including the family of belief propagation (BP) message-passing algorithms. Our contributions include analyzing the convergence and accuracy properties of BP, developing new BP-like bounds, extending BP techniques to continuous valued systems, improving the efficiency of "adaptive" or incremental inference, and extending variational algorithms to ``mixed'' inference tasks such as marginal MAP and decision making problems, including influence diagrams (or decision networks) and distributed team decision problems.
Applications. We have applied our algorithms to a wide variety of problems, including tracking and understanding data from sensor networks, efficient representations for large text corpora, computer vision and image processing, and gene expression data in biology.
I am co-program chair (with Dominik Janzing) of this year's Uncertainty in Artificial Intelligence (UAI) conference, to be held June 25-29 in New Jersey.
Our solver ("ai") won first place in five categories of UAI's 2014 Approximate Inference Challenge. Congratulations also to Rina Dechter's group ("daoopt"), which won several other categories.
We co-organized the NIPS'13 workshop, "Crowdsourcing: Theory, Algorithms and Applications".
I received an NSF CAREER award, "Estimation and Decisions in Graphical Models" (IIS-1254071)