Sequence Clustering by Time Delay Networks

Abstract

This paper outlines the form and structure of an activation passing context rich network ideally suited to tasks such as; recovering from noisy or damaged data, recognition or spelling correction. Such a network has been shown to be in many ways similar to N-Gram analysis or the transition matrix of Markov Models. However it is superior in that the scope of context (N) to be considered is dynamically and locally identified for each node. A pair of algorithms are presented which can be used to produce such networks. The first uses global working memory to deal with the time dependent nature of the input data, the second uses time delay nodes within the network. The latter, however, is considered superior as it reduces algorithmic complexity and increases computational efficiency.