Gaussian Mixture Models Reduction by Variational Maximum Mutual Information

Abstract

Gaussian mixture models (GMMs) are widely used in a variety of classification tasks where it is often important to approximate high order models by models with fewer components. The paper proposes a novel approach to this problem based on a parametric realization of the maximum mutual information (MMI) criterion and its approximation by a closed-form expression named variational-MMI (VMMI). The maximization of the VMMI can be carried out in an analytically tractable manner and it aims at improving the discrimination ability of the reduced set of models, a goal that was not targeted in previous approaches that simplify each class-related GMM independently. Two effective algorithms are proposed and studied for the optimization of the VMMI criterion. One is a steepest descent type algorithm, and the other, called line search A-functions (LSAF), uses concave associated functions. Experiments held in two speech related tasks, phone recognition and language recognition, demonstrate that the VMMI-based parametric model reduction algorithms significantly outperform previous non-discriminative methods. According to these experiments, the EM-like LSAF-based algorithm requires less iterations and converges to a better value of the objective function compared to the steepest descent algorithm.