Accelerated Message Passing for Entropy-Regularized MAP Inference
This work provides incremental improvements to computational efficiency for a fundamental machine learning problem, benefiting researchers and practitioners in fields like computer vision and natural language processing.
The paper tackles the problem of accelerating entropy-regularized message passing algorithms for MAP inference in Markov random fields, showing that the proposed randomized methods achieve faster convergence rates and recover the true MAP solution in fewer iterations when the linear programming relaxation is tight.
Maximum a posteriori (MAP) inference in discrete-valued Markov random fields is a fundamental problem in machine learning that involves identifying the most likely configuration of random variables given a distribution. Due to the difficulty of this combinatorial problem, linear programming (LP) relaxations are commonly used to derive specialized message passing algorithms that are often interpreted as coordinate descent on the dual LP. To achieve more desirable computational properties, a number of methods regularize the LP with an entropy term, leading to a class of smooth message passing algorithms with convergence guarantees. In this paper, we present randomized methods for accelerating these algorithms by leveraging techniques that underlie classical accelerated gradient methods. The proposed algorithms incorporate the familiar steps of standard smooth message passing algorithms, which can be viewed as coordinate minimization steps. We show that these accelerated variants achieve faster rates for finding $ε$-optimal points of the unregularized problem, and, when the LP is tight, we prove that the proposed algorithms recover the true MAP solution in fewer iterations than standard message passing algorithms.