Local Perturb-and-MAP for Structured Prediction

Two fundamental problems in the context of probabilistic graphical models are learning and inference. Many traditional probabilistic methods resort to approximations in either learning, inference, or even both steps due to their large complexity cost. This leads to algorithms where the learning and inference steps are disjoint, which often degrades algorithm performance. In this work, we present a Local Perturb-and-MAP (locPMAP) method, a novel framework for structured prediction based on local optimization over randomly perturbed potential functions. Unlike most prior methods, our proposed scheme does not rely on approximations and also unifies the learning and inference steps. This allows our approach to outperform other methods on several vision tasks, where structured models are commonly used. Additionally, we demonstrate how to apply our proposed scheme for an end-to-end training of a deep structured network. Finally, our framework gives a novel interpretation for pseudolikelihood, enabling various powerful extensions, such as the use of pseudolikelihood under the Bayesian framework.