Research On Key Technology Of Memristor Based Binary Convolutional Neural Network

Convolutional neural networks have powerful data processing capabilities,and research on them is a frontier topic in the field of brain-like computing.At present,hardware convolutional neural network circuits are usually implemented based on transistors.However,as the device size is gradually approaching the physical limit,it is more and more difficult to improve the performance of convolutional neural network based on transistor devices.Compared with transistors,memristors have obvious advantages in integration density,power consumption,and non-volatile storage.In addition to that,they also have neurosynaptic characteristics and are considered to be the most likely new physical foundation to support the disruptive development of brain-like computing.Therefore,it is of great significance to carry out research on memristor based convolutional neural networks.Nevertheless,limited by the existing device manufacturing process and other problem,multi-value memristive devices are still unstable,and binary memristive devices are basically available,but there are also non-ideal characteristics such as resistance fluctuations and array yield which greatly affect the recognition performance of convolutional neural network to a large extend.This paper focus on the memristor based convolutional neural networks with non-ideal characteristics,and studies and designs binary memristor based convolutional neural network that has good recognition performance and can tolerate non-ideal characteristics of the device.The main work of the paper is as follows:In Chapter 2,a memristor based binary convolutional neural network structure without pooling layer is designed.At first,the research on the structure design of the binarized convolutional neural network is carried out,the improved network structure,and the binarization method of the network and the binary training method of the network are designed(Section 2.1).Subsequently,the mapping method of the network weights to the memristor array is proposed(Section 2.2).Then,the implementation of network in memristor array is proposed(Section 2.3).Finally,the network performance is simulated and analyzed,and the network recognition rate,power consumption,and hardware resource consumption are verified(Section 2.4).The simulation results show that the binary memristor convolutional neural network structure designed in this chapter is simple to implement and consumes less resources.Compared with the existing memristor-based convolutional neural network,its recognition rate is only reduced by 0.8%.In Chapter 3,the mechanism of the mixing application of two types of neurons,type 01 and ±1,that can improve the tolerance of the memristor based binary convolutional network to the non-ideal characteristics of the device is designed.At first,a binary convolutional neural network with the mixed application of two kinds of neurons is designed(Section 3.1).Subsequently,the non-ideal characteristic model of the memristive device is constructed,which mainly includes the yield and resistance fluctuation of the device(Section 3.2).The output fluctuation model of neuron is constructed(Section 3.3).Then,the tolerance of the neural network with type 01 and± 1 neuron adopted separately to non-ideal characteristics is analyzed through simulation(Section 3.4).Finally,the binarized convolutional neural network with the mixed application of two types of neurons is simulated and analyzed for its tolerance to non-ideal characteristics(Section 3.5).The simulation results show that the mixed application mechanism of 01 type and ±1 type neurons designed in this chapter effectively improves the tolerance of the network to non-ideal characteristics of the device.In Chapter 4,a training method that can improve the tolerance of the memristor based binary convolutional neural network to non-ideal characteristics of devices is designed.At first,the problem of the existing training methods of memristor based neural network are explained(Section 4.1).Subsequently,a network training method considering the non-ideal characteristics of the device is proposed,and the specific implementation details are explained(Section 4.2).Then,the specific implementation process of the optimized training algorithm is introduced(Section 4.3).Finally,the optimized network training method is simulated and verified(Section 4.4).The simulation results show that the optimized training method can improve the tolerance of the memristor based binarized convolutional neural network to array yield,device resistance fluctuation and neuron output fluctuation.