CrossNets: Neuromorphic networks for nanoelectronic implementation

Abstract

Hybrid "CMOL" integrated circuits, incorporating advanced CMOS devices for neural cell bodies, nanowires as axons and dendrites, and single-molecule latching switches as synapses, may be used for the hardware implementation of extremely dense (∼107 cells and ∼1012 synapses per cm2) neuromorphic networks, operating up to 106 times faster than their biological prototypes. We are exploring several "CrossNet" architectures that accommodate the limitations imposed by CMOL hardware and should allow effective training of the networks without a direct external access to individual synapses. CrossNet training in the Hopfield mode have been confirmed on a software model of the network. © Springer-Verlag Berlin Heidelberg 2003.