A LIF Neuron With Adaptive Firing Frequency Based on the GaSe Memristor
Haowei Wang, Yichun Xu, Rui Yang, Xiangshui Miao
Abstract
The frequency-adaptive function of neurons is crucial to neural signal transmission in the biological neural system. And, artificial neurons with a frequency-adaptive function can effectively improve the energy efficiency of the neural networks. However, the reported frequency-adaptive neurons based on CMOS transistors generally require complex circuits. Here, a compact leaky integrate-and-fire (LIF) neuron with a frequency-adaptive function was demonstrated in a simple circuit by taking advantage of the dynamic turn-on delay process of the Ag/Ti/GaSe/Pt/Ti threshold switching memristor. And, high learning rate and low power consumption were realized in the spiking neural network (SNN) based on this frequency-adaptive neuron for digital recognition. The underlying physical mechanisms for the dynamic turn-on delay process were investigated in the Ag/Ti/GaSe/Pt/Ti device. And, it is found that the dynamic turn-on delay process is highly relative to the evolution process of Ag filaments during continuous switching in the device. This evolution process of Ag filaments can be modulated by introducing Ga vacancies in the GaSe layer.