On modifying the weights in a modular recurrent connectionist system

A modular recurrent connectionist architecture is proposed to classify binary and continuous patterns. This system consists of three networks: one feedforward backpropagation (BP) network and two self-organization map (SOM) networks. The feedforward (basic) network is trained until a saturation error level occurs. Simultaneously, the first SOM (input control) network and the last SOM (output control) define the mapping features for the given input/output patterns. The resultant features are used by a Gaussian potential function to adjust the weights of the basic network and to classify the given patterns.<>