Research On The Application Of Spiking Neural Network Based On Titanium Dioxide Memristor

In the context of rapid development of information technology,traditional Von Neumann computing systems,due to the separation of memory and computing units,bring higher power consumption and expensive costs in the existing CMOS circuit systems.To alleviate these problems,neuromorphic computing based on pulse neural networks has become an effective solution for more efficient computing systems.From the perspective of hardware implementation,exploring artificial synapses and neurons based on pulse neural networks is necessary.Memristors are considered the most competitive neuro-morphic devices for implementing artificial synapses and neurons due to their characteristics such as low power consumption,high-density integration,non-volatile storage,and reconfigurability.Among them,oxide memristors have attracted widespread research attention due to their simple fabrication process and compatibility with mainstream semiconductor manufacturing techniques.Currently,oxide memristors are mainly used for constructing synapses in neural networks,while there is relatively less research on utilizing oxide memristors for designing neurons and neural networks.In this paper,by fabricating Ag/TiO₂/Pt memristors as neuromorphic hardware,the modulation between volatile and nonvolatile devices is realized under different current limiting conditions.Based on the non-volatile characteristics of memristors,the synaptic characteristics such as I-V cycle characteristics,data retention ability and long-term plasticity were studied.Based on the volatile characteristics of memristors,synaptic characteristics such as short-term plasticity and modeling of artificial neurons.The volatile and nonvolatile Ag/TiO₂/Pt memristor is modeled as artificial synapse and neuron,and the SNN for supervised learning and unsupervised learning algorithms for handwritten digit recognition is built.The network provides an application solution for future low-power brain-inspired chips.The specific research contents are as follows:（1）The Ag/TiO₂/Pt devices with single structure and crossbar array structure were fabricated by magnetron sputtering.The experimental results show that the device exhibits volatility under I_cc=10μA,and have low power consumption of 1.5μW.The volatile behavior can correspond to the short-term plasticity of biological synapses.The device exhibits good nonvolatility under I_cc>100μA.On this condition,the retention time is longer than 10³ s and the window of R_OFF/R_ON is about 10⁴.Besides,the value of V_SET is as low as 0.25 V.The nonvolatile behavior can correspond to the long-term plasticity of biological synapses.（2）Under the DC voltage sweep,the conductance modulation with good linearity in the switching process of the device was realized.Under the pulse sequence stimulation,the conductance modulation and spike-frequency-dependent plasticity were realized by adjusting the frequency,amplitude and other parameters,and the short-term plasticity,long-term plasticity,transition from short-term to long-term reinforcement,and spike-time-dependent plasticity was simulated（3）Firstly,three commonly used neurons of the spiking neural network were modeled and simulated;then,the pulse sequence test was performed on the device,and the experimental results were fitted with the LIF neuron model;then,the memristive device was used as an artificial Neurons and synapses build a supervised learning network and an unsupervised learning network;finally,the learning ability of the network is tested using MNIST handwritten data sets,and the robustness of the network is verified by changing the network parameters.