The Research And Application Of Neural Network Architecture Search Base On Multi-objective Evolutionary Algorithm

With the rapid development of deep neural networks,the development of deep neural networks in modern computer vision has made remarkable achievements in image classification,key point detection and other tasks.However,designing and building a neural network model for different task scenarios usually requires a lot of specialized domain knowledge for human manual design and consumes a lot of computational resources and time for validation.Moreover,with the advent of the information explosion,the forms and sources of data are more abundant,and the application scenarios become more complex.Therefore,it is an important challenge to design suitable neural network architecture for different application scenarios.Neural Architecture Search(NAS)has attracted a lot of attention as part of Automated Machine Learning(Auto ML)in recent years.The neural network architecture search models the optimization problem of the network architecture,provides a solution for the automatic design of the network architecture,and adaptively modifies the architecture composition through the search strategy to achieve high performance on the target data set Although NAS methods have made great progress,there are still some challenges,especially how to reduce the number of GPUs required for deployment,memory footprint and search time costs.Based on the theoretical foundations related to multiobjective evolutionary algorithms,this paper proposes a novel search space and search strategy for the performance optimization,computational efficiency and specific application scenario optimization of neural network architecture search based on an evolutionary algorithm.The main research work and contributions of this paper are divided into the following two aspects:(1)An adaptive segmentation multi-objective evolutionary neural network architecture search algorithm(ASMEvo NAS)is proposed,which aims to address the computationally expensive problem of MOEAs requiring separate evaluation of multiple objectives at each generation,and thus improve their generalization and migration capabilities.First,an adaptive segmentation evaluation strategy is designed to adaptively select different but more appropriate targets to efficiently evaluate candidate architectures at different evolutionary stages,rather than evaluating them with all considered targets simultaneously.As a result,the computational cost and complexity of the search process can be controlled and reduced.Secondly,a preferencebased pre-selection strategy is designed to filter initialized architectures with a high number of parameters to reduce the total number of parameters for the whole population,thus accelerating convergence.Finally,a novel ideal desirable gene reservation-based crossover(DGRX)and directed connection-based mutation(DCM)is proposed to generate progeny.Experimental results show that the proposed algorithm is not only low in search time and memory cost,but also has good generalization and migration ability.(2)In terms of practical application,this paper proposes a COVID-19 image classification(MOEvo NAS)method based on a multi-objective neural network architecture search.This method takes classification accuracy,recall,and architectural complexity as optimization goals,and individual diversity as constraints,aiming to design a lightweight,high-performance CNN for the classification and recognition of COVID-19 computed tomography images.First of all,this paper designs two new convolution operation operators to reduce the number of convolution operation types that need to be selected in the supernet search space,thereby reducing the number of parameters and memory usage of the network architecture,and realizing the lightweight of the searched network architecture.Second,considering the particularity of the COVID-19 application itself,the recall rate is introduced in the evaluation performance index to avoid the adverse effects caused by missing positive COVID-19 cases.Experimental results show that the proposed MOEvo NAS exhibits better performance on both the small COVID-CT dataset and the more complex large COVIDx-CT and COVIDx-CT 2A datasets compared to other comparative networks.