Research On Design Of Non-Ideal Memristor-based Convolutional Neural Network

The integration of memory and computation based on memristor synaptic crossbar can effectively solve the problem of Von Neumann bottleneck in the hardware implementation of traditional artificial neural network.Memristor crossbar naturally realizes vector matrix multiplication,which provides huge parallel computing for neural network and significantly improves data throughput,thus achieving efficient hardware acceleration.By utilizing the spatial structure of the input image,the Convolutional Neural Network(CNN)is more suitable for visual tasks than other neural network structures(such as the fully connected neural networks).In addition,the number of synapses required by the CNN is much less and requires much smaller expense of hardware for pattern recognition tasks due to the network structure of incomplete connection and weight sharing.At present,CNNs often use ideal synaptic behavior to process information,but the actual memristor synaptic behavior shows inevitable non ideal factors,such as the number of conductance states(N_state),asymmetric write nonlinearities,device variations,etc.All of these may degrade the network learning and inference performance.In view of the above key problems,firstly,based on the experimentally measured electrical characteristics of the two typical TiN/LiSiO_x/Pt and TiN/HfO₂/Ti memristor devices,the code of the memristor-based CNN is written using the MATLAB platform.According to the characteristics of the memristor crossbar,the corresponding hardware deployment design of the CNN is implemented to achieve high parallelism and hardware acceleration.A quantization method is proposed to map the simulation weights of the ideal CNN to the discrete and limited conductance states of the memristor synaptic device,and different quantization training methods are proposed for the same quantization method.Based on MNIST handwritten font task,the simulation results show that the N_state needed by on-line quantitative training method is more than that of off-line quantitative training method,but it can process information in real time.Then,several non-ideal characteristics mentioned above are quantified,and their effects on the performance of CNNs are comprehensively studied.The simulation results show that the available conductance states,asymmetric write nonlinearities and cycle-to-cycle variation are critical factors to the learning accuracy,while symmetric write nonlinearities and device-to-device variation go trivial.Finally,according to the above key factors affecting the recognition accuracy,three mitigation strategies are proposed:1)limiting the weight range to improve the utilization of N_state;2)adopting the proposed"with-read"update scheme to mitigate the effects of asymmetric write nonlinearities;3)employing multiple memristors for each kernel element to alleviate the impact of cycle-to-cycle variation.When all the measures are applied to the written memristor-based CNN code,the recognition accuracy is greatly improved.When all the measures are applied to the program,the recognition accuracy of the network based on TiN/LiSiO_x/Pt and TiN/HfO₂/Ti devices has been greatly improved,which is from?86.69%to?95.25%and?93.54%to?96.81%,respectively.This work would provide guidance for the hardware implementation and optimization of CNN in memristor crossbar.