| With the continuous development of deep learning,there are more and more structures of neural networks in the field of image recognition.Therefore,some scientists have used neural networks in other fields,such as traffic prediction and text analysis,and have achieved excellent results.However,when migrating neural networks to other fields,directly migrating the neural network structure of image recognition,in most cases,an excellent result cannot be obtained in the new field.Therefore,someone proposed the concept of neural architecture search.Early neural architecture search methods used the principle of back-propagation to search the neural network,and the neural network architecture of each sampling required a lot of time to retrain.At the same time,during training,the weight parameters and the architecture parameters of the neural network are updated at the same time,resulting in a high degree of coupling of the node weight parameters.Therefore,this paper proposes a neural architecture search based on evolutionary algorithms,which has the following advantages:(1)The parameter sharing strategy is used.Before the neural architecture search,train a super network.During the neural architecture search process,each sampling structure only needs to directly inherit the weight parameters from the super network and infer directly on the data set to obtain the accuracy rate.This avoids training every sampling structure and improves the speed of neural architecture search.(2)In this paper,when training a super network,only one path is randomly activated at a time for training.Such a training strategy can first reduce the high coupling between nodes due to simultaneous updates when training the super network.Secondly,because only one path is involved in back propagation at a time,the training speed has been greatly improved,and the consumption of memory and GPU resources has fallen sharply.(3)When the particle swarm optimization algorithm is used to optimize the structure of the neural network,the operation of back propagation is avoided,and the continuation of the structure parameters of the neural network is avoided,which makes the process of architecture search more accurate and fast.(4)In the process of neural architecture search,the multi-objective optimization algorithm NSGA-Ⅱ is introduced,which allows us to simultaneously search the neural network structure for multiple targets,which is closer to the actual problem and enables the neural network structure search to be applied to the broader field. |
