Discriminant Processing In Multi-class Pattern Recognition Systems

Abstract

Multi-Layer Perceptron neural network classifiers face problems when applications have numerous output classes. A major problem is the fact that the MLP discriminant values given by the MLP differ considerably from the posterior probabilities of the Bayes decision rule. A non-linear mapping technique is developed in this thesis, which warps the neural network outputs into posterior probabilities. A second problem is that when the neural network is given inputs for classes it is not trained to handle, the output discriminant values become very noisy, as compared to the values seen for correct inputs. Variance based methods are investigated for detecting unanticipated classes. A method is developed for detecting cases where a class is confused with another. In this case, a follow on chapter helps clear up the confusion.