Discriminative Learning for Minimum Error Classification

Recently, due to the advent of artificial neural networks and learning vector quantizers, there is a resurgent interest in reexamining the classical techniques of discriminant analysis to suit the new classifier structures. One of the particular problems of interest is minimum error classification in which the misclassification probability is to be minimized based on a given set of training samples. In this paper, we propose a new formulation for the minimum error classification problem, together with a fundamental technique for designing a classifier that approaches the objective of minimum classification error in a more direct manner than traditional methods. We contrast the new method to several traditional classifier designs in typical experiments to demonstrate the superiority of the new learning formulation. The method can be applied to other classifier structures as well. Experimental results pertaining to a speech recognition task are also provided to show the effectiveness of the new technique.