Circuit Design of Piecewise Linear Activation Function for Memristive Neural Networks

Currently, in memristor-based neural network circuits, a three-piecewise linear activation function circuit is mainly used to approximate the sigmoid and tanh functions. This study proposes a five-piecewise linear activation function circuit to more accurately approximate the sigmoid and tanh functions. The circuit is divided into signal generation module, signal processing module, and output module. Firstly, the voltage signal is input into the signal processing module to generate line segments with different slopes. Then, the signal processing module solves the signal discontinuity problem at the inflection points. Finally, the corresponding output voltage is obtained through the output module. The correctness of the above circuit is verified through PSPICE simulation. Subsequently, based on the five-piecewise linear activation function circuit proposed in this study, an iris flower classification network is constructed. By testing the accuracy of iris flower classification, the effectiveness of the circuit is validated.