An automatic setting for training restricted boltzmann machine

Recently, deep learning techniques create much of a stir in machine learning. Its prominent performances can found in large number of applications in pattern recognition, such as image recognition, speech recognition and video processing. Most of the existing deep architectures are based on or related to restricted Boltzmann machine(RBM). However, the model selection problem in RBM and its deep architecture is very intractable since both their learning and inference are highly time-consuming. In this paper, we introduce an automatic setting for training a RBM, including structure design, log-likelihood approximation and batch learning design. We give a practical setting guide for training a RBM, which is deduced from the experiments on Bar-and-Stripe benchmark inpainting.