Gaussian Activation Function Realization with Application to the Neural Network Implementations

Abstract

A CMOS Gaussian function generator circuit suitable for the implementation of analog neural networks is proposed. For this purpose, it is considered the polynomial approximation of the Gaussian function. The proposed circuit realizes the Gaussian function characteristic inherently, that is without requiring any accurate tuning or adjustment of the circuit parameters. In order to show the usefulness of the proposed circuit, simulation results obtained using Spectre Simulation tool in Cadence design environment are provided. These results show the validity of the theoretical analysis and feasibility of the proposed structure.