Variants of SOM

Abstract

In order to create spatially ordered and organized representations of input occurrences in a neural network, the most essential principle seems to be to confine the learning corrections to subsets of network units that lie in the topological neighborhood of the best-matching unit. There seems to exist an indefinite number of ways to define the matching of an input occurrence with the internal representations, and even the neighborhood of a unit can be defined in many ways. It is neither necessary to define the corrections as gradient steps in the parameter space: improvements in matching may be achieved by batch computation or evolutionary changes. Consequently, all such cases will henceforth be regarded to belong to the broader category of the Self-Organizing Map (SOM) algorithms. This category may also be thought to contain both supervised and unsupervised learning methods.