Links between self-organizing feature maps and weighted vector quantization

Abstract

A novel learning algorithm for self-organizing feature maps (SOFMs) is presented. The learning algorithm is based on an extension of vector quantization called weighted vector quantization (WVQ). WVQ distortion is a weighted sum of the distortion between an input vector and each of the codevectors in the codebook. A formulation of WVQ is given, as well as two optimality conditions which are analogous to the nearest neighbor and centroid conditions of vector quantization. The authors then incorporate the SOFM neighborhood mechanism into WVQ, and use the WVQ optimality conditions to derive the algorithm.< >