Research On The Augmentation Method Of Feature Data In The Middle Layer Of Model

In deep learning,data augmentation has become a standard operation.Most of the existing data augmentation methods augment the characteristics of the input layer of the model,and it is necessary to design and select suitable data augmentation methods for different training data sets.Organizations such as Google try to reduce manual design and trial costs by automatically searching for the best solution in the limited augmentation strategy space,but the search process is quite time-consuming.Compared with the input layer features,the middle layer features have undergone model processing,which is more abstract and more useful information.It is hoped that better results can be achieved by expanding the features of the middle layer.This paper studies the data augmentation technology for the features of the middle layer of the model.Based on the existing research,proposes a certain generality of data augmentation methods for the middle layer of the model: incremental mean Gaussian noise augmentation,mean interpolation augmentation and The mean is extrapolated and expanded.The incremental mean Gaussian noise augmentation method generates new data by adding Gaussian noise to the middle layer features.The mean interpolation/extrapolation augmentation method uses two "parent category" samples(one is the category mean and the other is a random sample)to generate a new "subcategory" sample through interpolation and extrapolation.The classification accuracy of the proposed incremental mean Gaussian noise augmentation and mean extrapolation augmentation methods on the MNIST,CIFAR-10 and CIFAR-100 datasets all exceed baseline,Disout and Dropout.In particular,on the CIFAR-100 data set,the incremental mean Gaussian noise augmentation method is 0.96 points higher than Disout and 2.07 points relative to Dropout;the mean extrapolation augmentation method is 2.96 points higher than Disout and4.07 points relative to Dropout.Experiments show that the proposed incremental mean Gaussian noise augmentation method and mean extrapolation augmentation method have good versatility;at different positions in the middle layer of the model,the augmentation effect is different,and the further the position is,the greater the augmentation effect is.Good;the larger the number of categories to be classified,the better the augmentation effect.