Spoken Letter Recognition with Neural Networks

摘要

A strategy for optimizing a neural network known as the Radial Basis Functionclassifier (RBF) on a large spoken letter recognition problem is developed. The strategy can be viewed as a compromise between a fully adaptive approach involving prohibitively large amounts of computation, and a heuristic approach resulting in poor generalization. A value for the optimal number of kernel functions is suggested, and methods for determining the positions of the centers and the values of kernel function widths are provided. During the evolution of the optimization strategy it was demonstrated that spatial organization of the centers does not adversely affect the ability of the classifier to generalize. An RBF employing the optimization strategy achieved a lower error rate than Woodland's multilayer perceptron and two traditional static pattern classifiers in the same problem. The error rate of the RBF was very close to the predicted minimum error rate obtainable with an optimal Bayesian classifier. An examination of the computational requirements of the classifiers illustrated a significant tradeoff between the computational investment in training and level of generalization achieved.