Data Efficient Learning Methods Based On Latent Variable Augmentation For Image Classification

In the past years,people have made great progress in image categorization.Especially,the convolutional neural networks(CNN)-based image categorization methods achieve the state-of-the-art performance in the image categorization task.Existing CNN-based method are on the basis of two assumptions.The first assumption is that there is a large-scale image training set to train the CNN.The second one is that the data from both the training set and the test set is independent and identically distributed(i.i.d.).However,in some real-life application,there is not a large-scale training set.Namely,people have to use the insufficient training sample to training a CNN.This problem is called data efficient learning.In this paper,we pay more attention to the data efficient image categorization problem,i.e.,we train an image classifier with insufficient training samples to carry image categorization.To address this problem,one intuition is to increase the number of training samples so as to reduce the influence of insufficient training samples.To this end,we innovatively propose latent variables augmentation method,that is,we increase the number of training samples by latent variables augmentation indirectly.In this paper,we firstly propose two data efficient image categorization methods based on latent variables augmentation method.The first algorithm is called latent variable augmentation method based on generative adversarial networks(Lavagan).The second one is called classification method based on variational auto-encoders(Cevae).These methods are composed of two tasks mainly.One task of these two tasks is latent variables augmentation based on adaptive latent variable distributions.By this task,on one hand,we are able to construct a set of latent variables distributions;on the other hand,we can sample a large number of latent variables from these latent variable distributions and we can enhance the generalization ability of CNN model.The other task is utilizing these augmented latent variables for classification task so that we can obtain a predictive model.Secondly,we take the above two tasks into consideration and propose a uniform objective function for the latent variables augmentation method mentioned above in order that the above tasks can cooperate with each other.Thirdly,to determine the value of parameters in the CNN models,we exploit an alternative two-player minimization game optimization method and Variational statistical gradient decent(VSGD)to minimize the above objective functions with respect to the Lavagan method and the Cevae method.Furthermore,to investigate the effectiveness of the proposed latent variables augmentation mothed,we analyze the empirical error upper bound based on Hoeffding's Inequality and Chernoff Bounding method.Lastly,we implement the proposed Lavagan method and Cevae method based on the existing image categorization data set to estimate the performance.Moreover,we also compare the proposed method with famous image categorization methods.In the experimental research,we find that the proposed methods are able to make prediction with the data efficient settings;these phenomenon manifests that the proposed methods are of high feasibility.And after the comparison with the existing methods,the proposed Lavagan method and Cevae method return higher performance,which investigate the effectiveness of proposed methods.The contributions of this paper are composed of the following four folds:Firstly,we propose two latent variables augmentation methods to address the problem of data efficient image categorization problem.By latent variable augmentation,we can improve the generalization ability of CNN with insufficient training samples.To the best of our knowledge,this is the first study that improves the classification performance with data efficient settings by latent variables augmentation.Secondly,we propose a uniform objective model.In this model,we focus on two tasks.The first one is that we draw a large number of latent variables from a set of constrained and adaptive latent variable distributions.The second task is we use the above sampled latent variables for training an image classifier.To optimize the objective model,we utilize the alternative two-player minimization game optimization method and VSGD method to obtain the parameters of CNN.Thirdly,to investigate the feasibility and effectiveness of the proposed latent variables augmentation method theoretically,we analyze the empirical error upper bound based on Chernoff Bounding method and the empirical error upper bound of traditional CNN based on Hoeffding's Inequality.We compare these two empirical error upper bounds,and we can discover that the empirical error upper bound of the proposed method is lower than that of the traditional CNN models.This manifests that the proposed latent variables augmentation method is of feasibility and effectiveness.Lastly,we also investigate the feasibility and effectiveness of the proposed latent variables augmentation method in an experimental view.We implement the proposed method based on public existing images data sets.Also,we compare the proposed method with other state-of-the-art(SOTA)image categorization method for insufficient training samples settings and analyze the result.After this comparison,we can draw a conclusion that the proposed latent variable augmentation outperforms over other SOTA methods and the proposed latent variables augmentation methods are of feasibility.