Self-organizing neural nets in signal and image representation

This paper explores the use of piecewise linear approximations to create compact representations of 1-D signals and 2-D thinned edge-extracted images. We use a self-organizing neural net to cluster suitably coded data. The unsupervised training of the net is improved with the use of the reducing spheres of influence. Experimental results indicate that the technique tends to eliminate the problem of neurons capturing incorrect patterns and provides for better clustering. Computationally, the technique is less complex than conscience learning and produces similar results