Minimizing free energy of stochastic functions of markov chains

Abstract

Automatic speech recognition has generally been treated as a problem of Bayesian classification. This is suboptimal when the distributions of the training data do not match those of the test data to be recognized. In this paper we propose an alternate analogous classification paradigm, in which classes are modeled by thermodynamic systems, and classification is performed through a minimum energy rule. Bayesian classification is shown to be a specific instance of this paradigm when the temperature of the systems is unity. Classification at elevated temperatures naturally provides a mechanism for dealing with statistical variations between test and training data.