A programmable CMOS gaussian synapse for analogue VLSI neural networks

Abstract

Feedforward neural networks with a single hidden layer using a gaussian function have been proven to have the capability of universal approximation in a satisfactory sense. Back-propagation neural networks with gaussian function synapses have a better convergence property over those with linear multiplying synapses. A programmable gaussian synapse for analogue VLSI neural networks with hardware implementation and the programming techniques to program the cell are presented. The standard deviation and the magnitude of the gaussian synapse can be programmed externally. The proposed gaussian synapse was designed with single-ended inputs. To verify the programmability of the proposed gaussian synapse, a prototype chip was fabricated using a 1.2 μm CMOS process and the experimental results obtained are presented.