Software Design Of Neural Network System Based On Memristor

A memristor is known as the fourth basic electronic component,and is a passive circuit component that describes the magnetic flux and charge dependence.As early as 1971,Cai Shaotang theoretically predicted the existence of memristors.In 2008,HP Labs first manufactured the first prototype memristor device in the industry,confirming Leon Chua’s theoretical prediction of memristors.Memristors have the characteristics of small size,low energy consumption,high speed,and high nonlinearity,and are mainly used in the fields of storing and processing information units,brain like units,true random number generators,resonators,and other fields.Therefore,the learning and memory characteristics of memristors can be effectively combined with matrix multiplication of neural networks,and the high storage capacity brought by neural networks based on memristors can also match the large amount of data required by neural networks.Therefore,memristor arrays are very suitable for approximate computation of neural networks.The neural network computing structure based on this can effectively solve the memory access bottleneck problem of traditional computers.However,the existing memristor models are basically based on the HP model proposed by HP Labs,and the research on neural networks is still limited to the traditional CNN convolutional neural networks and derivative frameworks.At the level of using memristor cells and arrays,it is still at the stage of constructing a cross array of memristors,with a relatively simple structure,simple functions,and low compatibility with the use of subsequent circuits,These all cause a large amount of energy and resource waste in the process of analog-to-digital and digital to analog data exchange.Therefore,this paper proposes a software design of a new type of neural network system based on memristors to achieve approximate computation of convolutional neural networks,achieving better performance and shorter execution time while ensuring the accuracy of neural network computation.The following work has been completed in this article:(1)Starting from the basic theory of memristors,the mathematical expression principle of memristors is emphasized;This article introduces the Ti O2 memristor model from HP Lab.Through modeling the HP type memristor model,simulation experiments and performance analysis under sinusoidal excitation signals are implemented in the SPICE environment to verify the characteristics of the memristor.(2)This paper proposes a neural network computing and processing fusion architecture based on memristors.Firstly,an improved memristor cross array circuit architecture is designed to implement the process of one-time convolution algorithm through the framework.The architecture-based optimization unit design shortens the time for analog-to-digital conversion between pooling and convolution operations,effectively improving the efficiency of resource utilization.(3)This paper proposes to combine the entire fusion framework with image recognition algorithms to design a three-layer memristor neural network framework: preprocessing input layer,memristor neural network recognition layer,and decision output layer.Compared to the algorithm design of traditional neural network frameworks,the deep combination of this memristor neural network framework and scene optimization algorithms has significantly improved the ability to recognize images compared to traditional Hopfield network models.