| With the explosive growth of the amount of data and the increasing depth of neural networks,it is increasingly difficult for traditional chips to meet the computing power requirements of deep learning systems.Brain-inspired computing directly constructs hardware neural networks by designing efficient artificial synapses and artificial neurons,thereby overcoming the bottleneck of separation of storage and computing in traditional Von-Neumann architecture.However,the current research of brain-inspired computing mainly focuses on the traditional deterministic computing,which lacks quantitative evaluation of the uncertainty of neural network output results.By fabricating stochastic neurons or synapses,probabilistic brain-inspired computing based on Bayes’ theorem can accurately quantify the uncertainty of neural network predictions.The memristor reflects the inherent randomness in the resistive switching process,while it has the advantages of low power consumption,high speed,high integration,and the ability to realize in-memory computing,which perfectly meet the requirements of probabilistic brain-inspired computing for devices.This paper aims at the uncertainty quantification of the output results of the memristive neural network,by fabricating the threshold switching memristor(TSM)with different electrical characteristics,and utilizing the inherent randomness of the TSM devices to construct stochastic neurons and stochastic synapses.The main research results are as follows:Firstly,three different types of TSM devices,including ovonic threshold switching memristor,conductive bridge threshold switching memristor,and metal-insulator transition memristor,have been fabricated based on three functional layers of Ge Tex,Cu S/Ge Se,and VOx.The TSM devices have been investigated through basic electrical performance tests to get their basic switching characteristics and reliability.Secondly,inspired by the characteristics of the neurons in the probabilistic spiking neural network(SNN),a stochastic neuron circuit based on the randomness of the threshold voltage of the TSM devices has been designed.In the three kinds of TSM devices,the variation range of the threshold voltage Vth distribution of the Ti W/Cu S/Ge Se/Pt memristor reaches 0.34 V,the standard deviation reaches 0.076 V,and the variation of Vth is not related to the number of cycles,which best meets the requirements of stochastic neurons.Therefore,Ti W/Cu/Ge Se/Pt is selected to fabricate stochastic neurons.Physical modeling proves that thermal noise caused by temperature in the process of device conductance modulation is the physical source of stochastic threshold switching behavior.Then a stochastic neuron circuit that satisfies the relationship between spiking probability and membrane potential in probabilistic SNN has been built.Thirdly,focusing on utilizing the dynamic excitation characteristics of stochastic neurons to quantify the uncertainty of neural networks,a probabilistic SNN based on Ti W/Cu S/Ge Se/Pt stochastic neurons has been designed and simulated for the breast cancer diagnosis task,achieving a high diagnostic accuracy of 97.0%,and obtaining the classification probability of tumor data.Using predictive entropy to quantify the uncertainty of the network prediction results and setting a threshold of 0.2 for uncertainty,the critical error of malignancy being diagnosed as benign can be completely avoided.Fourthly,inspired by the characteristics of synapses in the bayesian neural network using Dropconnect,a stochastic synaptic circuit based on the stochastic delay time during switch-on of TSM devices has been designed.By connecting TSM devices in series with the gates of transistors in 1T1 R unit,the synapses in memristive neural networks can realize the function of stochastic valid/invalid.W/Ge Tex/W devices are selected to construct random synapse because of their high switching speed and high reliability.Sufficient randomness during switch-on of the device has been obtained through subthreshold pulse operation.The random invalid characteristics of the 1T1 R unit based on the W/Ge Tex/W device have been verified by circuit simulation,and a stochastic synaptic unit that meets the requirements of Dropconnect synapses has been fabricated.Lastly,focusing on quantifying the uncertainty of neural networks by using the stochastic valid/invalid characteristics of synapses,a neural network with feature extraction and Bayesian classification,has been designed and simulated for the task of diagnosing COVID-19.A 1T2 R cell based on TSM device is designed to accurately map the positive and negative weights of the stochastic synapse.Using the actual parameters of the Ti N/Li Al Ox/Pt nonvolatile memristor and the W/Ge Tex/W TSM,the memristive Bayesian neural network achieves a diagnostic accuracy of more than 95%.Using the prediction entropy to quantify the uncertainty of the network prediction results,with the prediction entropy 0.3 as the threshold,the classification accuracy of the memristive Bayesian neural network with a Dropconnect probability of 0.46 has been improved to 99.02%,which verify the effectiveness of the uncertainty quantification ability of the memristive Bayesian neural network. |
