CS Faculty Candidate Colloquium
Friday **Special Time & Location** February 22 10:45 - 11:50 AM Kelley 1005 Brian Potetz EECS Colloquium: Computer Science Faculty Candidate Carnegie Mellon University Efficient Statistical Inference and its Applications for Computer Vision Statistical inference in high-dimensional continuous probability distributions is a central issue in artificial intelligence and computation in general. One method that has proven highly successful in the past is belief propagation, which estimates the marginal of each variable in a distribution. Unfortunately, the computational complexity of belief propagation is exponential in the size of the largest clique of the graphical model underlying a probability distribution. Thus, this powerful inference technique has only been found practical for a very simple subclass of continuous-valued inference problems known as pairwise-connected Markov Random Fields. In this talk, I introduce a new technique that allows belief propagation to run in time linear with respect to clique size, allowing belief propagation to be performed efficiently on a much wider subclass of inference problems. This discovery has important applications in several fields, and I demonstrate the technique for se! veral computer vision problems relating to the inference of 3D shape. For example, in the inference of shape from shading, efficient belief propagation results in a substantial improvement over previous methods, both in quality and in flexibility. Finally, in addition to performing efficient inference, the efficient belief propagation techniques I introduce also make it possible to efficiently learn the parameters of graphical models, which is a fundamental issue in machine learning. I demonstrate these learning techniques by learning statistical priors for natural images, and illustrate these spatial priors by applying them to image denoising. Biography: Brian Potetz is a graduate student at Carnegie Mellon University, where he works with his advisor, Tai Sing Lee. His interests include developing methods for statistical inference and machine learning, exploring the statistics of natural scenes and images, and applying these advances to develop new algorithms for computer vision and to further understand how visual information is processed in the brain. He received Bachelor degrees in Mathematics and Computer Science from Case Western Reserve University in 1999, after which he spent two years developing software for the analysis of satellite imagery. He is the recipient of an NSF Graduate Research Fellowship and an NSF IGERT fellowship.
_______________________________________________ Colloquium mailing list [email protected] https://secure.engr.oregonstate.edu/mailman/listinfo/colloquium
