Artificial Neurons Based on a Threshold Switching Memristor with Ultralow Threshold Voltage

Abstract

Brain-inspired neuromorphic systems have recently garnered significant interest owing to their ability to effectively overcome the von Neumann bottleneck to increase computing and energy efficiency in the era of the rapid development of artificial intelligence. A hardware artificial neuron with a rectified linear unit (ReLU) activation function is highly desired for introducing a nonlinear activation function and resolving the vanishing gradient problem. In this work, we developed a ReLU artificial neuron based on a threshold switching memristor (TSM) device of Pt/Ag/Al₂O₃/HfO₂/Ag-NIs/Pt structure with an ultralow threshold voltage. This artificial neuron realizes the ReLU activation function by correlating the amplitude of the output spike with the amplitude of the input voltage, which is reported for the first time. To mitigate the potential “dying ReLU” problem that can arise when the ReLU activation function is applied to deep spiking neural networks (SNNs), we developed a LeakyReLU artificial neuron. Experimental results showed that we successfully developed a high-integration and low-power ReLU artificial neuron and its variant, the LeakyReLU artificial neuron, and realized a digital recognition function in a simulated single-layer fully connected SNN, which is of great significance for the construction of large-scale SNNs in the future.