Abstract
Mean-field variational inference is one of the most popu-lar approaches to inference in discrete random fields. Stan-dard mean-field optimization is based on coordinate descentand in many situations can be impractical. Thus, in prac-tice, various parallel techniques are used, which either relyon ad hoc smoothing with heuristically set parameters, orput strong constraints on the type of models. In this paper, we propose a novel proximal gradientbased approach to optimizing the variational objective. Itis naturally parallelizable and easy to implement. We prove its convergence, and demonstrate that, in practice, it yields faster convergence and often finds better optima than more traditional mean-field optimization techniques. Moreover, our method is less sensitive to the choiceof parameters.