Implementation of A Neuron Using Sigmoid Activation Function with CMOS

A multi-input neuron circuit with high-precision Sigmoid activation function (AF) is presented in this paper. The proposed circuit composed of input signal weighting circuit, current-voltage conversion circuit, and Sigmoid AF fitting circuit. Designed circuit can fit Sigmoid function based on the current-voltage relationship of differential pairs. The circuit is simulated in TSMC 0.18 um CMOS technology. The post-simulation shows that the error between the output of Sigmoid AF fitting circuit and the ideal Sigmoid function is 1.76%. The area of the layout is $375\ \mu m \times 238\ \mu m$. The maximum error caused by noise in the output of the circuit is $80\ pA/\sqrt{Hz}$.