Variations on the Boltzmann machine

Abstract

The performance of several neural-network-like models for pattern recognition tasks is analysed. A comparison based on recognition of random points in a multidimensional space is made between backpropagation, learning vector quantisation and three variations of Boltzmann machine models. The ordinary Boltzmann machine is found to perform well although at the expense of being time consuming. Replacement of the Monte Carlo method in the Boltzmann machine by a direct formula promises faster evaluation without significant loss of precision.