|Organization||University of California|
|Date||January 20, 2012 12:50 PM|
Graphical models provide a powerful framework for reasoning about complex models with uncertainty. "Inference" in a graphical model refers generically to answering probabilistic queries such as computing probabilities or finding optima. Types of inference tasks include max-inference (finding optimal configurations), sum-inference (calculating marginal probabilities or the probability of evidence), and mixed-inference (problems with both max and sum, for example so-called marginal MAP problems). Since all these tasks are NP-hard in general, approximation methods -- especially those that provide bounds -- are of great interest, and variational algorithms such as belief propagation and its variants have emerged as particularly powerful approaches. In this talk I will discuss a general framework that unifies many variational approaches, in which different inference tasks and algorithms are achieved by selecting different "weights" in the algorithm. As one consequence, we will show that this framework leads directly to novel and efficient approximation algorithms for mixed-inference problems, known to be the most difficult of the three tasks.
Alexander Ihler is an Assistant Professor in the Department of Computer Science at the University of California, Irvine. He received his Ph.D. from MIT in 2005 and B.S. from Caltech in 1998. His research focuses on machine learning, graphical models, and algorithms for exact and approximate inference, with applications to areas including sensor networks, computer vision, data mining, and computational biology.