Configurable activation function realized by non-linear memristor for neural network

Abstract

The activation function is a crucial part for memristive neural networks. For the first time, we propose a memristor-based activation function by using the natural non-linear characteristics of the memristor itself. Compared to the virtual ground circuit in traditional memristive neural networks, the feedback resistance was replaced by the W/TaOx/Ru memristor with no additional expense. Simulation results demonstrate that the proposed memristor-based activation function can achieve a performance similar to that of traditional activation functions on the Mixed National Institute of Standards and Technology database. In addition, an improvement in the recognition rate of up to 2% can be obtained in different neuromorphic networks by modulating the non-linearity of the memristor. Furthermore, the memristor-based activation function can also receive a 94% recognition rate even considering the non-ideal factors of the device.