Learning Clique Potentials for High-Order Graphical Models

Probabilistic models that incorporate high-order interactions between variables have demonstrated promise in solving many difficult problems in computer vision, natural language processing, computational biology, and other fields. Recent work shows how to perform efficient inference in such models. Unfortunately, there is currently no good method for learning the model parameters, limiting their wide-spread use. This project addresses that problem. We will develop both theory and practical algorithms for learning models with high-order interactions. Our methods will be motivated by the computer vision application of holistic scene understanding, but will also be applicable to many other application domains.