Supervised and unsupervised learning in radial basis function classifiers

The paper considers a number of strategies for training radial basis function (RBF) classifiers. A benchmark problem is constructed using ten-dimensional input patterns which have to be classified into one of three classes. The RBF networks are trained using a two-phase approach (unsupervised clustering for the first layer followed by supervised learning for the second layer), error backpropagation (supervised learning for both layers) and a hybrid approach. It is shown that RBF classifiers trained with error backpropagation give results almost identical to those obtained with a multilayer perceptron. Although networks trained with the two-phase approach give slightly worse classification results, it is argued that the hidden-layer representation of such networks is much more powerful, especially if it is encoded in the form of a Gaussian mixture model. During training, the number of subclusters present within the training database can be estimated: during testing, the activities in the hidden layer of the classification network can be used to assess the novelty of input patterns and thereby help to validate network outputs.