Synthesis of cascade recurrent neural networks using feedforward generalization properties

This paper presents a new analysis and synthesis technique for a class of recurrent networks known as a Cascade Recurrent Network (CRN). In this technique, a feedforward (FF) sub-network is used with synchronous feedback to implement associative memory (AM). FF network mapping properties are shown to determine CRN stability, and, on the basis of this stability-mapping relation, a new synthesis technique is given. This technique utilizes the optimization of the FF mapping sub-network generalization as a synthesis procedure for CRN. Sample results are shown.