Efficient training procedures for adaptive kernel classifiers

The authors investigate two training schemes for adapting the locations and receptive field widths of the centroids in radial basis function classifiers. The adaptive kernel classifier is able to adjust the responses of the hidden units during training using an extension of the Delta rule, thus leading to improved performance and reduced network size. The rapid kernel classifier, on the other hand, uses the faster learned vector quantization algorithm to adapt the centroids. This network shows a remarkable reduction in training time with little compromise in accuracy. The performance of these two networks is evaluated using underwater acoustic transient signals