Multi-SOMs for classification

Abstract

We propose a method to use the classification capabilities of self organising neural networks to extract symbolic information from raw data. The Multi-SOM (M-SOM) approach is a variant of Self Organising Maps (SOM). Multi-SOMS consist of a set of partner SOMs, trained simultaneously and in concurrence to each other, to adapt to different classes. The trained M-SOM transforms the non-linear time series of a strange attractor into a stream of symbols, adequate for further classification or for control tasks. We are convinced, that using the Multi-SOM approch for classification, gives a variety of new applications.