Memristor-based Neural Networks With Perturbation Algorithms Training And Its Circuit Implementations

In recent years,with the advent of AlphaGo and AlphaGo Zero,neural networks have been attracting more and more attention.It can be seen in the fields of image recognition,speech recognition and automatic control.Up to now,most of neural networks have been realized by software,which leads to the lack of the inherent parallelism of neural networks.However,circuit operations are inherently parallel and capable of providing high speed computation.Therefore,it is necessary to research hardware implementations of neural networks.When implementing neural networks via CMOS technology,non-volatile storage of weights is a major bottleneck.The emergence of memristor provides a new approach for the hardware implementations of neural networks.The non-volatile characteristics of memristors are similar to those of biological synapses,which spurs that memristor-based neural networks have attracted people 's interests.This paper first introduces the research background and significance of this topic,analyzes the research status of memristor-based neural networks at home and abroad,and then introduces the basic theories and methods related to memristor-based neural networks.Based on the theoretical knowledge of memristor-based neural networks,this paper proposes a memristor-based neural networks with weight simultaneous perturbation training and another memristor-based neural networks with weight summed perturbation training.The main contributions and innovations of this dissertation are summarized as follows:(1)A memristor-based neural network with weight simultaneous perturbation algorithm training is proposed.Firstly,in order to solve the problem of weight storage in the neural network,a memristor-based synapse circuit is designed.The circuit contains two memristors that can achieve positive and negative weights.The synaptic circuit has a larger weight adjustment range than a single memristor-based synapse circuit.In Matlab,a 2x2 single-layer neural network based on the designed memristor-based synapse circuit is constructed to verify the feasibility of the proposed synapse circuit.Secondly,A memristor-based neural network with weight simultaneous perturbation algorithm training is proposed.Compared with the error back propagation algorithm,the algorithm does not involve error back propagation and activation function derivation,and is easy to implement in hardware.In order to make full use of the characteristics of the algorithm,the circuit operation protocol of the algorithm is designed,and the feasibility of the circuit operation protocol is proved.Finally,the feasibility and correctness of our proposed memristor-based neural network with weight simultaneous perturbation algorithm training are verified by two tasks.(2)A memristor-based neural network with weight summed perturbation training is proposed.Compared with the error back propagation algorithm,the algorithm does not involve error back propagation and activation function derivation,and is easy to implement in hardware.Compared with the weight simultaneous perturbation algorithm,the algorithm only applies perturbations to neurons,and the circuit operations are more concise.Based on the characteristics of weight summed perturbation algorithm,the circuit operation protocol of the algorithm is redesigned,and the feasibility and rationality of the circuit operation protocol are proved mathematically.Finally,the feasibility and correctness of our proposed memristor-based neural network with weight summed perturbation training are verified by two tasks.