Local Variance Driven Self-Organization for Unsupervised Clustering
Matthew J. Kyan, Ling Guan · 2006
We propose a new, novel unsupervised clustering technique based on traditional Kohonen self organization, Competitive Hebbian Learning (CHL), and the Hebbian based Maximum Eigenfilter (HME). This method fits into the family of dynamic selfgenerating, Self-Organizing Map (SOM) algorithms. The approach uses a vigilance based, global parsing strategy as a guide for the hierarchical partitioning of an underlying data distribution into a set of dominant prototypes: each consisting of a dual memory element for the online estimation of both position and maximal local variance. A co-operative scheme exploits the interplay between global vigilance and maximal local variance such that an informed choice may be made regarding insertion sites for new nodes into the Map. The network is related to Self-Organizing Tree Maps (SOTM), Growing Neural Gas (GNG) and their variants. A framework is presented and performance demonstrated against GNG.