Joint Training Deep Boltzmann Machines for Classification

Ian Goodfellow, Aaron Courville, Yoshua Bengio

Decision: conferenceOral-iclr2013-workshop

Abstract: We introduce a new method for training deep Boltzmann machines jointly. Prior methods require an initial learning pass that trains the deep Boltzmann machine greedily, one layer at a time, or do not perform well on classification tasks. In our approach, we train all layers of the DBM simultaneously, using a novel inpainting-based objective function that facilitates second order optimization and line searches.