Performance analysis and improvements on a hybrid cascade architecture for multi-layer neural networks

A series of improvements in a hybrid architecture for multilayer networks is presented. This architecture incorporates the incoming connection strengths and the neurons of each layer into one stage by a multiplexing scheme, hence reducing the complexity of interstage wiring. An analysis of the performance of this architecture is performed and, based on its results, the authors propose a number of improvements. Also, a three-layer network has been implemented in a double metal, single polysilicon p-well CMOS technology based on the proposed improvements. The performance of the improved version is analyzed and compared to the original structure. Bounds on the operating speed of the system are also presented