Design And Implementation Of Memristor-based Convolutional Neural Network

With the continuous development of technology,deep learning has been widely used in image processing,natural language processing,speech recognition and other fields and has achieved great success.Among them,one of the most frequently used deep learning models is Convolutional Neural Network(CNN).Because it can learn more abstract,higher-order,and closer to actual features from data,it is widely used in image classification and recognition tasks.However,because the traditional von Neumann computing architecture has great limitations,that is,it is unable to self-learn and improve according to actual needs,resulting in the bottleneck of the performance improvement of the deep learning model in the traditional von Neumann architecture.Therefore,this paper uses a memristor with variable resistance and storage and calculation to construct a convolutional neural network accelerator to solve the shortcomings of the traditional architecture.1.The mathematical model of the memristor is described,and its circuit characteristics are analyzed.First,introduce the four types of memristor models commonly used in the market,namely HP memristor model,spin memristor model,threshold switch memristor model,and voltage threshold adaptive memristor model,and analyze each Mathematical characteristics and circuit characteristics of the model.Then the basic mathematical derivation of the memristor complex circuit and analysis of its electrical characteristics are carried out.Namely series,parallel,and hybrid.Finally,the basic knowledge of neural networks is explained,focusing on the analysis of convolutional neural network models,and then a simple analysis of memristive neural networks.2.Based on the basic characteristics of the convolutional neural network and the memristor cross-array,an accelerating module that is suitable for the convolutional neural network model is designed.The accelerating module is composed of many computing arrays.Each computing array can support basic operations in convolutional neural networks,such as convolution,pooling,and activation.In addition,the computing array can be configured into a computing mode and a storage mode according to actual needs.According to the relationship between the storage capacity and calculation capacity of each layer of the convolutional neural network,the calculation array part is reasonably allocated as data storage,and the remaining calculation array is used for the convolutional neural network calculation.3.Based on the memristor cross-array acceleration module,hybrid mapping and pipeline optimization are used to improve the performance of the acceleration module.For the convolution operation,a hybrid mapping method is used to improve the space and time utilization of the acceleration module.Parallel convolution kernels are used in space,and the parallelism of mapping is improved through input and output replication.In terms of time,according to the calculation amount of each layer of convolution,more weights are mapped to the idle acceleration array,and finally a pipeline balance is achieved.Experiments show that for the classic convolutional neural network Le Net,compared to the conventional mapping,the hybrid mapping and pipeline in this paper will increase the performance of the acceleration module by 23.2 times.Compared with existing work,the energy efficiency of the acceleration module has increased by 20.1%.