A Memristor-Based Leaky Integrate-and-Fire Artificial Neuron With Tunable Performance

Artificial neurons have received extensive attention as an important part of neuromorphic computing. Recently, tremendous efforts have been made on the memristor-based neurons, while the regulation of performance of such neurons and its underlining mechanism has been rarely studied. In this work, we propose an artificial neuron device based on Ag/TaO_x/Si, which exhibits good threshold switching characteristics (on-off ratio above 10⁵) along with good device stability and cycling stability. The Leaky Integrate-and-Fire (LIF) neuron model is successfully simulated without additional circuitry, including leaky integrated firing and refractory periods. In addition, the effect of oxygen vacancy concentration on the performance of artificial neurons is investigated, and the results showed that an increase of oxygen vacancies can significantly reduce the threshold voltage of neuron activation, the holding voltage and the probability of refractory period. This work provides a simple and effective strategy for the development of artificial neurons with tunable properties.

I. Introduction

Currently, the rapid development of information technology has put forward urgent requirements for high efficiency and ultra-low power consumption [1]. Neuromorphic computing has attracted extensive attention due to its high parallelism, extremely low power consumption and integration of storage and computing [2]–[4]. As an important cornerstone of neuromorphic computing, artificial neurons play an important role in neuromorphic computing. From a computational point of view, artificial neurons integrate multiple inputs and transmit signals to the next level of neurons in the form of spikes when a threshold is reached [5]–[7]. So far, there have been many reports on the realization of artificial neurons by analog circuits, but complementary metal-oxide-semiconductor (CMOS) technology faces many difficulties in further reducing power consumption, reducing area, and simulating the inherent characteristics of biology [8]–[12]. Therefore, artificial neuron devices with unique physical mechanisms are of great significance to the development of neuromorphic computing, the research and development of new non-von Neumann architecture chips and the ultimate realization of brain-like intelligence [13]–[20].