Enhancing performance of restricted Boltzmann machines via log-sum regularization

摘要

Restricted Boltzmann machines (RBMs) are often used as building blocks to construct a deep belief network. By optimizing several RBMs, the deep networks can be trained quickly to achieve good performance on the tasks of interest. To further improve the performance of data representation, many researches focus on incorporating sparsity into RBMs. In this paper, we propose a novel sparse RBM model, referred to as LogSumRBM. Instead of constraining the expected activation of every hidden unit to the same low level of sparsity as done in [27], we explicitly encourage the hidden units to be sparse through adding a log-sum norm constraint on the totality of the hidden units' activation probabilities. In this approach, we do not need to keep the "firing rate" of each hidden unit at a certain level that is set beforehand, and therefore the level of sparsity corresponding to each hidden unit can be automatically learnt based on the task at hand. Some experiments conducted on several image data sets of different scales show that LogSumRBM learns sparser and more discriminative representations compared with the related state-of-the-art models, and stacking two LogSumRBMs learns more significant features which mimic computations in the cortical hierarchy. Meanwhile, LogSumRBM can also be used to pre-train deep networks, and achieve better classification performance.