CMOS-Memristive Sigmoid Activation Function

Abstract

Sigmoid is a commonly used activation function in Artificial Neural Network (ANN). There are several digital, mixed signal and analog implementations of a sigmoid function; however the existing sigmoid circuits limit the scalability of ANN due to large on-chip area and high power consumption. In this paper, we address this issue and propose analog memristor-based sigmoid activation function implementation. The CMOS components of conventional gain adjustable voltage-mode sigmoid circuit are replaced with memristive element. The behavior of the modified sigmoid is simulated in SPICE and compared with the existing CMOS design. The simulation results show that memristor based sigmoid allow to reduce on-chip area and power dissipation by 20% and 5%, respectively. The gain of the sigmoid is increased by 20%, while the processing time remains unchanged.