CMOS analogue neurone circuit with programmable activation functions utilising MOS transistors with optimised process/device parameters

Abstract

In on-chip learning feedforward and feedback neural networks, different nonlinear activation functions are required. To date, these different functions are realised by physically changing the activation circuits. A novel CMOS analogue circuit suitable for modelling neurons with programmable activation functions is introduced. The proposed circuit is realised by optimising the second-order components of MOS transistors. Programmability of the activation function is achieved using an external controlling signal. Intensive simulations based on SPICE3 indicate that the proposed circuit can realise different activation functions including step, linear threshold, and sigmoid functions.