Stochastic Systems Group  

Jason Johnson
SSG Doctoral Student
This research is aimed at developing nearoptimal, scalable inference algorithms for random fields which admit compact description as graphical models. Estimation of GaussMarkov random fields (GMRF) is the current focus of this work. The approach being developed, referred to as recursive cavity modeling (RCM), combines ideas from the multiscale modeling method, developed by researchers in our group before, with extensions of existing approximate inference techniques which operate by frontier propagation (FP). FP involves the construction and propagation of a frontier model which provides a faithful yet compact graphical model for the surface of one subfield for the purpose of inferring another. While FP performs inference with respect to a chain model of the field, RCM uses a tree model. The RCM approach has the advantage of decomposing the field in a more natural manner and of being ideally suited for distributed processing. Possible future extensions of this work might include the consideration of dynamic random fields as well as more general classes of static random fields. In this talk I will discuss a basic implementation of the RCM algorithm for estimating GMRFs and provide MonteCarlo simulation results examining the scalability and performance of this algorithm.
Problems with this site should be emailed to jonesb@mit.edu