Self-organizing map with centroidal Voronoi tessellation and generation of neurons

Abstract

Kohonen's Self-Organizing feature Map (SOM) is a method to obtain a topology-preserving mapping from high-dimensional invisible feature space to two or less dimensional visible space. For obtaining a good mapping, the user has to set up a good combination of parameters, type of lattice, number of neurons and neighboring radii varied with iteration etc.. This paper proposes a new method to obtain a mapping without any setting of parameters except for the termination condition. The method starts with six neurons. And the method moves positions of neurons in the visible space and generates new neurons in accordance with the data distribution in the feature space.