Research And Implementation Of A Neural Network Evolutionary Algorithms

Convolutional neural network has developed rapidly in recent years,and various network structures have appeared.It needs a lot of time for professionals to design a suitable network structure for the problem.How to reduce the cost of neural network structure design is a problem that neural network must face when it is widely used.In recent years,some scholars have proposed using evolutionary algorithms to search for the appropriate neural network structure,but these algorithms have too much computational cost to calculate in less resources.This thesis mainly discusses the existing evolutionary neural network algorithm,and proposes an evolutionary neural network algorithm based on the existing network substructure to search the neural network structure,in order to reduce the calculation cost of structure search.Based on the substructure,the algorithm defines structured super parameters,including depth,width and dropout probability.In the evolutionary mutation operation,the neural network structure evolution is realized by random mutation and instantiation of the super parameters of the substructure.In this thesis,the design and implementation of the algorithm are described in detail,and three measures are proposed to reduce the operation cost: first,by periodically limiting the number of offspring evolution mechanism,limiting the number of population reproduction,thus limiting the depth of structural search.Secondly,an early stop algorithm based on the evaluation of training period and accuracy gain is designed to evaluate the evolutionary efficiency according to the training efficiency,so that the unwell offspring can stop faster.Thirdly,the neural network fitness calculation method with the negative correlation factor of neural network scale is designed to restrain the expansion of parameter scale.Based on keras,distributed database,shared storage and other open-source components,this thesis designs and implements the experimental system,realizes the validation of the data set based on fashion MNIST,and achieves 92.66% accuracy under 221 k parameters.After full training,the selected model reaches 93.1%.The result is the same as that of 2 conv + 3 FC(1.8m parameter)neural network.Combined with the problem of 4K quality judgment of ultrahigh definition video image,the image preprocessing is integrated,and the engineering application is verified.The optimal accuracy rate is 92%,the accuracy rate is 100%,the recall rate is 84%,and the F1 measure is 0.913.