Research On Convolutional Neural Network Architecture Search Based On Genetic Algorith

Convolutional Neural Network(CNN)is one of the most representative models in deep learning algorithms,but the design process of CNN is very time-consuming and labor-intensive.To address this problem,researchers proposed the deep neural network architecture search(NAS)which automates the design and training process of CNNs.Due to the advantages of relatively low resource consumption and high search performance,evolutionary computation has become one of the important search frameworks in NAS.Based on these theoretical approaches,we analyze the advantages and shortcomings of current evolutionary computation frameworks and explores efficient evolutionary computation methods to obtain CNN model architectures with shorter training time,fewer number of parameters and better adaptation to complex datasets.Therefore,our main work is summarized as follows:(1)A novel network of attention mechanism based on confidence.Most current attention mechanisms do not adequately model the human visual system,i.e.,screening objects by differences.To address this problem,we propose a novel attention mechanism network based on the disparity between local and global perceptual fields.The network extracts local and global contextual information and corrects the input using the disparity between them.Experiments showed that the confidence-based attention mechanism uses fewer parameters to achieve higher performance gains.(2)A genetic algorithm-based search method for attention network architecture.To address the problems of time-consuming and low performance of the traditional attention mechanism to manually design the local receptive field size,we propose a genetic algorithm-based attentional receptive field design method.The method encodes the receptive field size of the whole network into individuals and populations,simulates population evolution,and performs crossover and mutation operations to form the optimal receptive field combinations.Experiments showed that the receptive field size combination searched by the genetic algorithm-based attention mechanism network architecture search method is more suitable for the network model.(3)An effective sleep stage classification method based on the neural network architecture search of genetic algorithm.To address the problems of poor scalability and time-consuming design training process of traditional machine learning methods for sleep staging classification,we propose a CNN model with the addition of trainable kernel coefficients for sleep data;meanwhile,an efficient deep convolutional neural network architecture search algorithm based on genetic algorithm searches out such an efficient CNN architecture.We first bind a trainable coefficient on each convolutional kernel,and then use a deep convolutional neural network search framework based on genetic algorithm to generate such an efficient CNN architecture.Experiments showed that this method can significantly improve the performance of sleep signal processing and the accuracy of sleep stage classification,while reducing the time consumption of network design.