Neurons with continuous varying activation in self-organizing maps

Abstract

A new training and recall method for self-organizing maps (SOM) is developed by comparison of SOM to the human information processing system. As neurons and cortical columns do not change their activity instantly, it is increased or decreased in a smooth way. This fact is introduced in SOM-neurons. In a same way, recogni2ion of objects is supposed to be a task of analysing complete sets of feature vectors and finding the region in the SOM which represents the current inputs best. This method especially allows the evalutation of ambiguous feature vectors and of objects which are decomposed in sets of basic feature vectors or which are aquired in a continuous temporal flow.