Pattern classification using self-organizing feature maps

The author demonstrates the use of self-organizing feature maps as pattern classifiers. When a set of training patterns is presented to a self-organizing network repeatedly for many iterations, the weight vectors gradually organize themselves to be the cluster centers of these patterns. In the one-dimensional case, they can arrange themselves to satisfy a linear ordering relation. That is, the distance between two weight vectors increases as the physical distance between the two corresponding output nodes increases. However, the latter is true only when an adequate size of neighborhood is used in the network. The author notices a cyclic phenomenon among the distances between weight vectors when the neighborhood size is small