The general objective of this project is to fully explore the "change-Pumping network (CPN)" concept, and to transform it into a very densely integrable family of microelectronic neural networks. The CPN, in its generic form, is the simplest interconnected array of MOS gated-diodes. It is capable of performing inner product and thresholding operations in one direction and weighted averaging in the opposite in a simultaneous fashion over the same synaptic matrix. Yet, no visible feedback path exists in the array. This creates a very rich bidirectional neural functionality in a very compact network. The research plan includes the development and refinement of network synthesis procedures, a search for self-learning ability and the entire design/fab/test cycle for implementing five different target architectures. The goal is to extend the knowledge base in neural network synthesis by offering a general procedure for non-negative synaptic matrix design, and to help develop a knowledge base in collective multistability Through an analysis of this fundamental concept.
