A tunable Gaussian/square function computation circuit for analogneural networks

摘要

A Gaussian/square function computation circuit suitable for analognneural networks is proposed. It can realize Gaussian and squarenfunctions when operating in weak and strong inversion region,nrespectively. It is shown that the center, width, and peak amplitude ofnthe dc transfer curve can be controlled separably. Measurement resultsnon 3-Μm CMOS fabricated chips confirm theoretical and simulationnfindings