Blending Learning and Inference in Conditional Random Fields

摘要

Conditional random fields maximize the log-likelihood oftraining labels given the training data, e.g., objects givenimages. In many cases the training labels are structures thatconsist of a set of variables and the computational complexityfor estimating their likelihood is exponential in the number ofthe variables. Learning algorithms relax this computationalburden using approximate inference that is nested as a sub-procedure. In this paper we describe the objective function fornested learning and inference in conditional random fields. Thedevised objective maximizes the log-beliefs --- probabilitydistributions over subsets of training variables that agree ontheir marginal probabilities. This objective is concave andconsists of two types of variables that are related to thelearning and inference tasks respectively. Importantly, weafterwards show how to blend the learning and inferenceprocedure and effectively get to the identical optimum muchfaster. The proposed algorithm currently achieves the state-of-the-art in various computer vision applications. color="gray">