Evolutionary Neural Network And Its Improvement

In recent years,artificial intelligence technology has been widely used in various fields such as computer vision,natural language processing,image classification and recognition,and intelligent robots.As two important branches of artificial intelligence technology,evolutionary algorithm and neural network have played an important role in various fields of life.The ability of large-scale parallel processing and learning of neural network make it possible to deal with complex practical engineering problems.However,the structure and parameters of models are still need to be adjusted manually,which consumes human labour and material resources.With the increase of network depth and the improvement of network performance,the parameters and network structure needed to be set manually become more and more complicated,Therefore,it is more important to design network structure and adjust parameters automatically.As a global optimization intelligent algorithm,evolutionary algorithm has the characteristics of parallel processing and high efficiency,so researchers combine neural network and evolutionary algorithm to propose evolutionary neural network algorithm.The thesis analyzes and optimizes the details of the evolutionary neural network algorithm,and applies it to the specific application of image classification.The main research work of the thesis is as follows:Firstly,it introduces the evolutionary neural network algorithm,including theory and related knowledge.Besides,it summarizes the research status at home and abroad and the main research directions.Then,it elaborates two kinds of evolutionary neural networks: the enhanced topology evolutionary neural network and evolutionary neural network based on deep learning in detail.Secondly,this thesis refines the improvements of deep evolutionary neural networks,and proposes an evolutionary convolutional neural network ECNN algorithm for image classification.This algorithm introduces ResNet module theory.Activation function,number of filters,filter size,step size,loss rate and other layer attributes are randomly selected and optimized by evolutionary algorithm to fully optimize the network.The ECNN algorithm was used in the malaria infection medical image data set officially released by the American Medical Library for image classification test.The results show that the improved algorithm has good stability and accuracy.Thirdly,in order to further study the deep evolution neural network,the thesis proposes a D-ECNN algorithm of the DenseNet evolutionary convolutional neural network for image classification,and introduces the DenseNet module which can effectively solve the gradient disappearance problem in the deep convolutional neural network.The activation function and the layer attributes such as filter number,filter size,loss rate,and number of neurons are constructed by evolutionary algorithm for random optimization.Beyond that,some parameters in the algorithm are fixed for selective randomization,which effectively reduces training parameters.Finally,D-ECNN algorithm is used in the public vehicle data set released by Stanford University for image classification test,and the performance of the model is compared with the advanced convolutional neural network algorithm.The results prove that the proposed algorithm is competitive and effective.