Applications of self-organization networks spatially isomorphic to patterns

Abstract

A new technique based on self-organization is proposed for classifying patterns (which include characters, and two- and three-dimensional objects). A neuronal network, created to be a physical replica of each exemplar, is mapped onto the given test pattern by self-organization, during which the network undergoes deformation in an attempt to match the given test pattern. The extent of deformation is inversely proportional to the correctness of the match: smaller the deformation, better is the match. A deformation measure is proposed, leading to the classification of the test pattern. Also presented are some algorithmic improvements (including the choice of other deformation measures) to speed up computation. Examples illustrate the versatility of the technique.