Improving Categorization with CALM Maps

Abstract

The Categorizing And Learning Module (CALM) represents different patterns on different nodes through a competitive learning procedure. We study an extension of CALM that enforces a topological structure on the representations. The main difference with Kohonen’s self-organizing feature map is that no external regulating mechanisms are needed to learn a stable map. Simulations show that this CALM Map, in comparison to the standard CALM module, improves categorization because the stretching property of CALM Maps enables a continuous process of separation, whereas CALM will eventually commit itself to a once obtained categorization.