Simple analog nonlinear circuits for neural networks

Abstract

Simple analog nonlinear circuits are proposed for implementing neural networks using standard CMOS technology. Three analog nonlinear circuits are proposed: a nonlinear difference circuit, a nonlinear synapse circuit and a nonlinear multiplier (which is composed of the proposed nonlinear difference circuit and the proposed nonlinear synapse circuit.) The proposed multiplier takes less silicon area than the conventional linear multipliers do. The proposed nonlinear circuits are fully simulated using HPSICE. The proposed nonlinear circuits are applied for implementation of multi-layered feedforward circuits and MEBP (modified error backpropagation) learning circuitry. The implemented neural networks have been simulated using HSPICE circuit simulator and produce an output voltage, which is uniquely determined by any pair of learning input patterns. The proposed nonlinear circuits are very suitable for future implementation of the large-scale neural networks with learning.