Fully Memristive Spiking Neural Network Based On V/VO_x/HfWO_x/Pt Multimode Memristors

Neuromorphic computing based on emerging memory technologies offers a viable solution to overcome the von Neumann bottleneck.The energy and area efficiency of neuromorphic computing architectures can be effectively enhanced by exploiting the analogue storage and dynamic switching properties of memristor devices to emulate synaptic and neuronal functions.In this work,to address the key issues regarding process,performance and parameter mismatching among neuromorphic devices in fully memristive neural networks,the reliability and neural functions of memristors are improved by means of material doping modification,preparation and measurement method optimizations.Moreover,the hardware-friendly training schemes are developed based on the fast and bidirectional switching of as-fabricated memristors.In this way,a homogeneous fully memristive spiking neural network（SNN）is proposed for remedying the lack and inefficiency of algorithms in SNN.The main results are as follows.At the device level,the multimode V/VO_x/HfWO_x/Pt memristors exhibiting both memorable and switching capability is intentially designed by exploiting the non-volatile resistive switching（RS）property of HfWO_x and the volatile threshold switching（TS）of VO_xfilms.Specifically,the storage performance of the WO_x-based resistive layer is improved by Hf doping,while the oxidation of the V electrode interface during deposition is optimised to prepare VO_x films of less than 5 nm thickness.The device reconfigurably delivers resistive and threshold switching without electroforming and high temperature annealing processes.More than 10³ on/off ratio,retention for 10⁴ seconds,and a pulse endurance for 10¹⁰ cycles are fonund in the RS mode.And the TS mode shows fast switching（less than 30 ns）,high endurance(more than 10¹² cycles)and excellent threshold stability.At the function level,the same V/VO_x/HfWO_x/Pt memristor device is used for the first time to emulate synaptic and nonpolar neuronal functions as well as neural signal interconnections.Specifically,the oxygen vacancy migrations in the HfWO_x-based RS layer is used to mimic synaptic plasticity,including long-term potentiation/depression and spike timing dependent plasticity.And the neuronal circuit is demonstrated by exploiting the metal insulator transition（MIT）in the VO_x-based TS layer,in which the symmetrical integration and firing behaviours of neurons are successgully emulated under positive and negative voltage excitations.Furthermore,the synapse-neuron integration scheme is proposed based on V/VO_x/HfWO_x/Pt memristors.By delicately designing the circuits,the neuron and the synaptic devices match well in operating parameters,and thus the effects of different synaptic weights on the spiking time and frequency of neurons can be demonstrated.At the neuromorphic application level,the homogeneous full memristive SNN based on V/VO_x/HfWO_x/Pt memristors shows excellent classification capability on both MNIST and CIFAR10 datasets.Specifically,a temporally encoded single-layer SNN is constructed towards the MNIST handwritten digits recognition task.The fast spiking of neurons and the introduction of temporal coding lead to a significant reduction in network latency（⁰.5 μs）and power consumption（¹0 f J per synaptic operation）.Furthermore,for the CIFAR10 classification task,a multi-layer convolutional SNN with nonpolar frequency coding is constructed,in which the nonpolar neurons are used for the first time to extract phase-opposite informations of inputs.Compared to unipolar networks,the synaptic and neuronal hardware overheads in the nonpolar network are respectively reduced by 75% and 50% without loss of learning rate,which provides an effective solution for the training of the deep fully memristive SNN.