VQ assistance for training perceptron networks

Vector quantization algorithms are used to find finite sets of exemplars which represent a data set to within an a priori error tolerance. Such representation is of the essence in codebook based data compression and transmission. We first develop modifications of the basic algorithm and then explore the use of vector quantization as a tool to speed the training of perceptron networks. We show that the vector quantization provides an efficient initialization for the backprop algorithm. We also explore the use of vector quantization to decompose large scale computational problems into more computable parts. Classification problems are considered