Topic: Nonparametric Inference for Graphical Models
By: Erik Sudderth (MIT), Alexander Ihler (MIT), William Freeman (MIT), Alan Willsky (MIT)
In many applications of graphical models, the hidden variables of interest are most naturally specified by continuous, non-Gaussian distributions. There exist inference algorithms for discrete approximations to these continuous distributions, but for the high-dimensional variables arising in domains such as visual tracking, discrete inference becomes infeasible. Drawing on ideas from regularized particle filters and belief propagation (BP), we have developed a nonparametric belief propagation (NBP) algorithm applicable to general graphs. Each NBP iteration uses an efficient sampling procedure to update kernel-based approximations to the true, continuous likelihoods. We have applied NBP to a pair of challenging applications: distributed self-calibration of an ad hoc sensor network from range measurements, and visual tracking of three-dimensional hand motion using a kinematic model. Current research focuses on the continuing development of efficient sampling and model identification algorithms, and exploration of additional application domains.
Questions and comments to the webmaster. This page is copyrighted by the Massachusetts Institute of Technology.