On the Model Selection of Bernoulli Restricted Boltzmann Machines Through Harmony Search

Abstract

Restricted Boltzmann Machines (RBMs) are amongst the most widely pursued techniques in deep learning-based environments. However, the problem of selecting a suitable set of parameters still remains an open question, since it is not straightforward to choose them without prior knowledge. In this paper, we introduce the Harmony Search (HS) optimization algorithm to find out a suitable set of parameters that minimize the reconstruction error of Bernoulli RBMs, which address binary-valued visible and hidden units. The results have shown the suitability of using HS for such task when compared to other optimization techniques.