Neuromorphic Application And Performance Research Based On Memristor Arrays

The emerging nonvolatile memory represented by memristor has significant advantages in memory computing integration,and has the possibility to break through the Von Neumann architecture of computers.Compared with traditional devices,the cross array composed of memristor can efficiently realize matrix vector multiplication in neural network.However,the non-ideal characteristics of actual devices due to the randomness of ion motion will cause the cumulative errors of neural network simulation calculation,thus reducing the calculation accuracy of neural network.Therefore,it is of great significance to quantitatively test and analyze the influence of non-ideal characteristics of memristor on the correlation performance of memristor neural network applications.This paper takes memristor array as the research object,and gives its simulation circuit design and test results in logic calculation,image processing and neural network applications.Combined with Neurosim simulation platform,a method to extract nonideal characteristic parameters of real memristor is presented.A large number of simulation tests are carried out on the handwritten digit recognition performance of the multi-layer perceptron network based on different memristors.According to the test results,the influence of the non-ideal characteristics of the memristor on the performance of the multi-layer perceptron is quantitatively analyzed,and the device design optimization suggestions are given.It mainly includes the following work:(1)Aiming at the problems such as boundary locking and inability to describe the real device resistance in HP memristor model,by comparing the characteristics of different window functions and analyzing the influence of migration current parameters in AIST memristor model on the resistive behavior of memristor,a new memristor model based on improved window function and adding migration current parameters was proposed.The simulation results show that the model can better simulate the resistance phenomenon of real devices,solve the boundary locking problem and enhance the generalization.(2)The ead/write circuit based on memristor array and the method of weight coding and mapping of memory unit are designed.The synaptic circuits are designed and the functions of AND logic,OR logic,IMP logic,NIMP logic,XOR logic,NXOR logic are realized respectively.A binary image edge detection simulation circuit based on memristor array is designed and implemented.(3)Based on the Neurosim simulation platform,the recognition performance of the memristor multilayer perceptron network constructed by different memristors in handwritten numeral set(MNIST)was tested and analyzed under offline learning and online learning modes.The influence of the non-ideal characteristics of the memristor on the learning performance of the neural network is discussed,and the related factors affecting the recognition performance of the memristor multilayer perceptron are analyzed.The method of improving the recognition performance of the memristor multilayer perceptron based on the device level is proposed,including adjusting the nonlinear value and window ratio of the memristor.In this paper,the mathematical model of memristor gradient curve in Pspice is studied,and the detailed steps to get the non-ideal characteristic parameters of the real memristor by fitting the pulse-conductance normalized curve of the memristor with Matlab are given.(4)The improved memristor proposed in this paper and three kinds of memristor devices successfully prepared from existing literature were selected to construct the memristor multilayer perceptron,and the performance indexes such as recognition accuracy,system integration area,read and write power consumption and read and write delay in the MNIST data set were quantitatively evaluated.By comparing and analyzing the experimental data,the influence of the non-ideal characteristics of the memristor on the related performance of the memristor multilayer perceptron and the design optimization direction of the memristor device are discussed,which provides a feasible test and analysis for the researchers facing the complex memristor neural network application,and select the method suitable for the memristor device according to its specific application scenario and system performance index.