Unsupervised sequence classification

Abstract

The authors first introduce a novel approach for unsupervised sequence classification, the competitive sequence learning (CSL) system. The CSL system consists of several extended Kohonen feature maps which are ordered in a hierarchy. The CSL maps develop a representation for subsequences during the training procedure, with an increasing abstraction on the higher maps. The authors apply their approach to real speech data and report preliminary results on a word recognition task. A generalization rate of 70% is achieved. The CSL system performs learning by listening: it divides the continuous sequence of input patterns into statistically relevant subsequences. This representation can be used to find appropriate subword models by means of a self-organizing neural network.<>