Kohonen network

Abstract

A Kohonen network as a self-organizing mechanism supplies an important contribution in the development of neural networks. The learning aspect is mainly aimed at the quantification of vectors, which can be accompanied by a reduction of the dimension. Further, the property that “shapes” remain kept with self-organizing feature maps, makes the Kohonen network a very strong instrument.

A striking fact is that both anatomically and functionally certain cortex areas can be discerned that have similar properties as Kohonen networks. Examples are the processing of sound and light stimuli.

With respect to the theoretical properties of Kohonen networks little is known. Especially the rate of convergence related to the time-dependent learning factor is an uncertain aspect. For the application to the TSP no theoretical foundation is present indicating the quality of the heuristic. In conclusion, a Kohonen network appears to be a promising technique for which coming mathematical foundations have to show the quality as well as future directions of development.

In the appendix a further application of the self-organizing property of a Kohonen network is worked out. This example is given by Henseler and Postma (1990).