Research On The Application Of Memristor With Both Volatile And Non-volatile In Neuromorphic Computing

The development of artificial intelligence depends on the innovation of algorithms,the accumulation of effective data and the enhancement of computing power.Neuromorphic computing based on memristors is expected to break the Von-Neumann bottleneck and enhance the computing power of computers.In recent years,volatile memristors and non-volatile memristors have often been reported to simulate neurons and synapses in neuromorphic computing,but there are few reports on the application of combining the volatile and non-volatile properties of memristors to simulate synapses.The non-volatility and volatility of memristors can well simulate the memory and forgetting characteristics of synapses.Combining these two characteristics of the memristor to simulate synapses and conduct neuromorphic computing application research is expected to improve the biological rationality and energy efficiency of artificial synapses and is of great significance to the development of artificial intelligence.This research first designed a spiking neural network algorithm based on the hybrid synapse of volatile memristor and non-volatile memristor.This algorithm uses the spontaneous attenuation characteristic of the volatile memristor to train the network,reduces the number of operations on the memristor.Afterwards,the non-volatile memristor array is used for weight transfer and network inference.A spiking neural network model is built through Python,and the feasibility of the algorithm is verified according to the classification accuracy of the MNIST data set in the network.Subsequently,the robustness of the algorithm to various influencing factors was studied,including the forgetting rate of the volatile memristor,the initial conductance state of the device,the nonlinearity of the non-volatile memristor,the error of the device,the number of device conductance states and the image noise.The simulation results show that device errors have a greater impact on the classification performance of the algorithm,but the algorithm has a strong tolerance for device forgetting rate,device initial conductance state,device nonlinearity,number of device conductance states,and noise.Finally,from the perspective of artificial intelligence data accumulation,a hardware implementation algorithm for feature selection is designed based on the conversion of memristors from volatile to non-volatile.Use the previously designed spiking neural network algorithm and the newly built three-layer perceptron backpropagation algorithm to evaluate the feature subset,and compare it with the classic feature selection method based on search strategy.The results show that the proposed algorithm reduces the number of features used in training by about half at the expense of small network classification accuracy,reduces the complexity of network training.This research is based on the combination of volatile and non-volatile memristors,which opens up new ideas for the efficient application of memristors.At the same time,it has enlightening significance for the hardware implementation of neuromorphic computing and feature selection.