Discriminant vector tranformations in neural network classifiers

Abstract

We review the fact that several kinds of neural networks can be trained to approximate other types of discriminant functions, thereby throwing some doubt upon the utility of the No Free Lunch theorem. Using a license plate recognition database with 36 classes, we then demonstrate that multilayer perceptrons estimate posterior probabilities very poorly when the number of classes is large. A method for generating desired posterior probability values is provided. Then an algorithm is developed and demonstrated for warping neural net discriminants into approximate posterior probabilities.