Quantum And Deep Neural Architecture Search By Quantum Evolutionary Algorithm For Image Classification

With the development of neural network-related research,in the field of deep learning,the Deep Neural Networks(DNN)model has shown strong performance advantages in image classification and other tasks.In the field of quantum computing,due to the advantages of parallel characteristics of quantum systems,the Quantum Neural Networks(QNN)generated by the combination of quantum computing and neural network has also become a field of in-depth research by more and more researchers.In recent years,more and more neural networks with excellent performance have been proposed and applied in different tasks.A neural network with excellent performance often depends on its excellent artificially designed network architecture.However,whether it is DNN or QNN,the architecture design of neural network still needs to rely on very rich prior knowledge and professional experience.Therefore,how to automatically design,build and search for an excellent neural network architecture for the same task such as image classification has very important and far-reaching practical significance.This thesis explores and researches neural architecture search methods on quantum neural networks and traditional deep neural networks.The quantum evolutionary strategy is used to improve the search ability in the huge discrete search space,so as to help the neural architecture search method to search for a better network structure.In this thesis,different search spaces are designed according to the different network characteristics of QNN and DNN.A new neural architecture search is proposed to solve the problems of no clear guiding experience in network design,low accuracy and high complexity of manually designed models,insufficient search ability of existing methods in discrete space,high time cost,and low correlation of networks performance.The main work of this thesis is as follows:Aiming at the problems in the field of quantum neural network,such as the lack of clear guiding experience in the network design method,the poor performance of the manually designed network model,the high circuit complexity of the network model,the traditional network architecture search method can not fully utilize the quantum characteristics,this thesis proposes an evolutionary quantum neural architecture search method based on quantum circuit.Through the comparison experiment with the original artificially designed network model,it can be concluded that the searched quantum neural network has improved model accuracy compared with the original network model,and effectively reduced the complexity of the quantum circuit.It’s very helpful to build and run quantum neural networks on the quantum computer.In terms of deep neural network,aiming at the problems that artificial design of network model requires a lot of a priori knowledge,the existing neural architecture search methods based on evolution have insufficient search ability in a huge discrete search space,and the time cost is large,this thesis proposes a Oneshot neural architecture search method based on quantum evolutionary algorithm,which improves the search ability and reduces the search time.By comparing experiments with artificially designed network models and existing neural architecture search methods on the benchmark datasets,such like CIFAR10 and CIFAR100,it can be found that the neural network searched by the method proposed in this thesis has better model performance.Aiming at the problem of low correlation between network performance evaluation and real performance in the neural architecture search method,this thesis analyzes and deeply thinks about the problems and contradictions between the traditional neural architecture search method and the Oneshot method based on weight sharing,and the conflicts and problems between different supernets training sampling methods in the current Oneshot method.Finally,a neural architecture search method based on quantum evolutionary algorithm and balanced pool strategy is proposed on the basis of maintaining the advantages brought by quantum evolution.In comparative experiments,the neural network searched by the proposed method has better model performance than other artificially designed networks and networks searched by other methods.