A tunable Gaussian/square function computation circuit for analog neural networks

Abstract
A Gaussian/square function computation circuit suitable for analog neural networks is proposed. It can realize Gaussian and square functions when operating in weak and strong inversion region, respectively. It is shown that the center, width, and peak amplitude of the dc transfer curve can be controlled separably. Measurement results on 3-μm CMOS fabricated chips confirm theoretical and simulation findings

This publication has 14 references indexed in Scilit: