On the use of the GTM algorithm for mode detection

Abstract

The problem of detecting the modes of the multivariate continuous distribution generating the data is of central interest in various areas of modern statistical analysis. The popular self-organizing map (SOM) structure provides a rough estimate of that underlying density and can therefore be brought to bear with this problem. In this paper we consider the recently proposed, mixture-based generative topographic mapping (GTM) algorithm for SOM training. Our long-term goal is to develop, from a map appropriately trained via GTM, a fast, integrated and reliable strategy involving just a few key statistics. Preliminary simulations with Gaussian data highlight various interesting aspects of our working strategy. © Springer-Verlag Berlin Heidelberg 2003.