Contrastive divergence learning for the Restricted Boltzmann Machine

The Deep Belief Network (DBN) recently introduced by Hinton is a kind of deep architectures which have been applied with success in many machine learning tasks. The DBN is based on Restricted Boltzmann Machine (RBM), which is a particular energy-based model. In this paper, we lay more emphasis on the modeling process and learning algorithm of the RBM. Furthermore, we design two kinds of experiments to prove the efficiency of the algorithm based on synthetic dataset and real dataset. The reconstruction data experiments are aimed at proving the convergence of the learning algorithm. The classification experiments are designed to testify the efficiency of the trained models. The result shows that contrastive divergence learning is an effective training algorithm for the RBM model.