Fitness Functions for the Optimization of Self-Organizing Maps.
Daniel Polani · 1997
Kohonen's Self-Organizing Maps do not possess a canonical criterion that characterizes their organization state. Therefore, to be able to optimize e.g. their topology, adequate fitness criteria have to be found. The paper introduces, analyzes and discusses three criteria, a criterion derived from the quantization error, a Hebbian and a hybrid measure incorporating favourable properties of both mentioned criteria. 1 Introduction The study of self-organizing maps and similar unsupervised neural network models related or derived from the Kohonen Feature Map has recently attained a lot of attention [3]. Not only do these models provide a framework for the study of biologically plausible mechanisms of information processing in the cortex of mammals, but due to their concise mathematical structure and geometric interpretability they also offer a geometry oriented tool for the analysis of data [10, 11]. Remarkably enough, there remain still a lot of open questions regarding to the mechanis...