Modular T-mode neural network learning hardware implementations with analog storage capability

Abstract

A modular T-mode (transconductance-mode) design approach is presented for analog hardware implementations of neural networks. This design approach is used to build a BAM network, a Hopfield network, a winner-take-all network, and a simplified ART1 network. The size of these networks can be increased by interconnecting more modular chips together. The approach is extended to include synaptic Hebbian learning as well as an analog scheme to refresh the learned weights. Experimental results of programmable and learning chips from a standard 2- mu m double-metal double-poly CMOS process (MOSIS) are given.< >