Research On Synaptic Performance Of LiAlO_x Memristor And Its Neural Network Application

With the advent of the big data era,computers based on the Von Neumann structure are facing increasingly serious challenges in power consumption and latency.At the same time,memristors exhibit the same electrical characteristics as human brain synapses,and have the advantages of high speed and low power consumption.However,when constructing a memristor-baesd neural network,the inherent non-ideal characteristics of memristors often cause the network performance to decrease.Therefore,manufacturing memristors with excellent synaptic performance and exploring the application of devices in big data tasks such as neural networks have become the focus of current research.In this paper,a memristor based on LiAlO_X was prepared,and its synaptic properties and physical mechanism were systematically studied.Through the optimization of both device and network level,the performance of the memristive synapse in neural network tasks has been improved.The main research work is as follows:(1)On the basis of TiN/LiAlO_X/Pt structure memristor,characterization methods such as TEM,XPS,and FTIR were used to analyze the actual structure and composition of the device.Meanwhile,electrical test experiments such as DC scanning were carried out to demonstrate the resistance switching characteristics of the device.Through comparative experiments of different through-hole sizes,a conductive filament type memristor model is constructed.(2)Synaptic plasticity and memory characteristics were simulated in LiAlO_X devices,including LTP/LTD behavior under pulse modulation,STDP performance and retention characteristics.The current transmission mechanism of the memristor was explored based on variable temperature experiments.In addition,the resistance switching mechanism of synaptic devices is analyzed in detail.Furthermore,the principle of increasing the initial conductance to optimize the modulation nonlinearity is clarified.(3)A scheme to realize convolutional neural network based on LiAlO_X memristive array is proposed.Synaptic performance of the LiAlO_X memristor is evaluated by using a shallow convolutional neural network for handwritten digit recognition.Moreover,an update strategy with a gradually increasing threshold is proposed to optimize the recognition and convergence performance of memristive neural network.A deep memristive network for action recognition is constructed,and the simulation of the network illustrates the necessity of using optimization strategies at both device and network levels.