Study of performances for Restricted Boltzmann Machines

Restricted Boltzmann Machines (RBMs) are one of the most relevant unsupervised learning methods. The aim of this thesis is to study their performances as a function of their parameters. First, we consider binary-valued RBMs and then we introduce the so-called centering trick, which is known to solve the absence of invariance to flip transformations. Moreover, centering also leads to more accurate models. Then, we discuss RBMs with real-valued units. In particular, we focus on rectified linear units, which are able to achieve better generative performances than binary units.