Structures of the covariance matrices in the classifier design

Abstract

Structurization of the covariance matrices helps to reduce a number of parameters to he estimated. When assumptions on the structure of the matrix are correct the structurization of the covariance matrix helps to reduce the generalization error in small learning-set cases. Efficacy of the matrix structurization increases if one decorrelates and scales the data, and uses the optimally stopped single layer perceptron classifier afterwards.