Reconfigurable Synaptic and Neuronal Functions in a V/VOx/HfWOx/Pt Memristor for Nonpolar Spiking Convolutional Neural Network

The fully memristive neural network consisting of the threshold switching (TS) material-based electronic neurons and resistive switching (RS) one-based synapses shows the potential for revolutionizing the energy and area efficiency in neuromorphic computing while being confronted with challenges such as reliability and process compatibility between memristive synaptic and neuronal devices. Here, a spiking convolutional neural network (SCNN) is constructed with the forming-and-annealing-free V/VOx/HfWOx/Pt memristive devices. Specifically, both highly reliable RS (endurance >10(10), on-off ratio >10(3)) and TS (endurance >10(12)) are found in the same device by setting it at RRAM or selector mode with either the HfWOx or naturally oxidized VOx layers dominating the conductance tuning. Such reconfigurability enables the emulation of both synaptic and nonpolar neuronal behaviors within the same device. A V/VOx/HfWOx/Pt-based hardware system is thus experimentally demonstrated at much simplified process complexity and higher reliability, in which typical neural dynamics including synaptic plasticity and nonpolar neuronal spiking response are imitated. At the network level, a fully memristive SCNN incorporating nonpolar neurons is proposed for the first time. The system level simulation shows competency in pattern recognition with a dramatically reduced hardware consumption, paving the way for implementing fully memristive intelligent systems.

Automated summary

This study constructs a fully memristive neural network using reliable synaptic and neuronal devices and reconfigurable V/VOx/HfWOx/Pt memristive devices. The network successfully emulates typical neural dynamics, showing competency in pattern recognition with reduced hardware consumption. This work provides a direction for implementing fully memristive intelligent systems.