Abstract

In this paper, we propose an extended self-organising learning scheme, in which both distance measure and neighbourhood function have been replaced by the neuron's posterior probabilities. Updating of weights is within a limited but fixed sized neighbourhood of the winner. Each unit will converge to one component of a mixture distribution of input samples, so that an optimal pattern classifier can be formed. The proposed learning scheme can be used to train other forms of unsupervised networks, such as radial-basis-function networks. An application example on textured image segmentation is presented.