Research On Neuromorphic Computational Properties Based On Non-ideal Memristor Devices

With the development of modern electronic technology and biotechnology,neuromorphic computing has gradually attracted widespread attention in the academic community.It implements functions such as pattern recognition,automatic control,and information processing by simulating human brain neural networks.However,due to the limitations of network size and synaptic components,the function of neuromorphic computing has been limited.The emergence of memristors provides a new physical basis for improving the level of synaptic simulation,and then provides the possibility for further development of neuromorphic computing.The nanoscale size,low energy consumption and good synaptic properties of the memristor make the neuromorphic computing based on the memristor array further improve the computational power and computational efficiency of the neural network.However,the memristor array fabricated by the current process has problems such as fluctuation,yield,and line resistance.To this end,this paper first introduces a multi-layer perceptron network and a singlelayer spiking neural network based on memristors.Then the volatility model,yield model and line resistance model of the non-ideal memristors are established respectively.Taking MNIST handwritten digital library and Chinese digital library as examples,the two networks are tested in the ideal memristive device to evaluate the calculation performance.The computational performance of these two networks with non-ideal characteristic model are studied by the control variable method.The experimental results show that the resistance fluctuation of the memristor has a certain influence on the multi-layer perceptron network with the memristor cross array used as the synapses,and the fluctuation of the resistance will deteriorate the performance of the network to some extent.As the resistance fluctuation of the memristor increases,the computational performance of the network decreases continuously,and the fluctuation range of the computing performance of the network will increase,too.The yield of the memristor has a large impact on the multi-layer perceptron network whose synapses are based on the memristors in the crossbar array,and its yield will deteriorate the performance of the network to a large extent;compared to the memristor failure to rise to a high-resistance state,the failure reduced to a low-resistance state has a relatively small impact on the network;the line resistance in the memristor cross array is also a nonnegligible factor affecting network performance.As the size of the array increases,the voltage division of the line resistance will seriously deteriorate the performance of the network.In addition,the increase of the line resistance also deteriorates the performance of the network.Unlike multi-layer perceptron networks,single-layer spiking neural networks are binary networks which only have two weights of high-resistance and low-resistance.The high-resistance fluctuation of the memristor has little effect on the computational performance of the single-layer spiking neural network,while the low-resistance fluctuation of the memristor has a certain influence on the network,and with the increase of fluctuation or the low-resistance state of the memristor,the network performance will be reduced;The yield of the memristor array has a greater impact on the network,and the memristor failure reduced to a low-resistance state and the failure rising to a high-impedance state have similar effects on a single-layer spiking neural network;The line resistance in the memristor array also has a large effect on the network.As the line resistance or the size of the array increases,this negative effect will increase significantly.