Research On Few-shot Learning Based On Matching Networks

In recent years,as the computer technology developing at a high rate of speed and a large number of data generated by the Internet,deep learning has made remark-able achievements in computer vision,auto machine translation,speech recognition,automaic drive and other fields.However,the current popular recognition algorithms,such as AlexNet^[1]?VGGNet^[2]?ResNet^[3],etc.,need to be trained on the data sets of at least thousands of training samples in each category to get better results in the corresponding categories.And it is difficult to continue to learn new categories through a small number of samples.These conditions limit the application of deep learning in some areas of data scarcity.Therefore,there are many research fields such as few/one/zero shot learning,which mainly study how to design recognition algorithm when each class is given only a few,one or no training samples,so as to maintain the strong generalization ability of the model.In this paper,the existing distance model Matchingnet^[4]of few/one shot learning is improved and visualized.Compared with the existing network structure of various recognition and target detection features extraction,combined with the visualization results,a new network structure and a new learning rate decay are creatively proposed without changing the network lightweight.Speed up the model training speed and improve the accuracy.The accuracy of omniglot and flower data set one shot learning tasks increased by 0.6 and 6.5 percentage points respectively.The research shows that dropout can increase the noise of the sample after comparing the input,which is helpful to alleviate the over fitting caused by too few samples.The gradient deconvolution visualization method is used to visualize the model,help to understand the working principle of the model,find problems,and provide follow-up research direction.