Research And Application Of Convolutional Neural Network Model With Small-world Features

For the past few years,convolutional neural network has got great attention due to its good performance in deep learning.To improve convolutional neural network,many skills were proposed like going deeper,DropOut and DropConnect,batch normalize and so on.A good convolutional neural network model needs many experiments,seldom theory can be referenced.Based on lots of convolutional neural network models,an obvious trend was that using shortcut connection to deal with gradient vanishing problem.In terms of network's structure,shortcut connection brings long connection,so traditional convolutional neural network structures are changing from regular network to small world network.Is it feasible according to construction of small world network to design a workable convolutional neural network? A new convolutional neural network model was proposed on the basis of small world network and tested on some data sets.The main work was that based on DenseNet model,to design an improved convolutional neural network with random shortcut connection.The new model's shortcut connections were not dense like DenseNet model and its topological structure was not regular network or random network,so to a certain extent it had small world network feature.In order to analyse the influence of shortcut connection,a lot of new models were tested on Mnist and Cifar10 data sets.The results indicated that better model should use long shortcut connections and use more front layers' feature maps.At last,the improved model was tested on a medical image data set,and the result is better than its control model,so it got expected result.