A supervised approach for training Gaussian Mixture Model classifiers
Abstract
A new method for training Gaussian Mixture Model (GMM) classifiers is presented. First, an objective function is defined in terms of the number of clusters, K, per class, the mean vectors, the inverse covariance matrices for each class, and the prior probabilities for each class. For each increment in K, gradients of the objective function improve upon the prior probabilities, mean vectors, and inverse covariance matrices. Improvement in accuracy for different data-sets are shown and results are compared with the EM algorithm.