A parallel recurrent cascade-correlation neural network with natural connectionist glue

Abstract

Some problems of "unlearning" were encountered when using Fahlman's recurrent cascade correlation learning architecture (RCC) for phoneme recognition. In this paper the authors present a parallel-modular RCC. The original RCC is transformed into a modular RCC, trained with natural connectionist glue. This is done in order to concentrate the "knowledge" about a group of patterns in a module, instead of distributing it across the whole network. The modules are connected in parallel, in contrast to the completely cascaded structure of the original RCC. This new approach provides an improvement in the recognition rates for tasks involving large numbers of features to be learned. The modularity, besides providing a better learning, makes training of large sample-sets easier and faster.