SOM-based pattern generator: Pattern generation based on the backward projection in a self-organizing map and its applications

Abstract

A major feature of the self-organizing map (SOM) is a topology-preserving projection from the input layer to the competitive layer, and it has been used mainly as an analytical tool for discovering underlying rules in the given data set. Even though recent splendid progress in this area, there are few novel ideas to break such a conventional style. On the contrary, based on its distinctive nature, a new method for generating patterns through backward projection from the competitive layer to the input layer is proposed recently. Moreover, a promising technology for producing animation as a series of backward-projected patterns along with any pathways on the competitive layer is presented. Then, in order to carry out further considerations, some computer simulations with a variety of posed stick figures are tried in this paper. After training, four kinds of pathways, which correspond to different movements such as dancing, exercising and walking, are prepared. Though some of them does not contain any training samples, all of them worked well as we have intended in advance. As a result, it is found that the proposed method shows good performance and it is also confirmed its effectiveness. © 2010 The Institute of Electrical Engineers of Japan.