Research On Transfer Learning Based On Generative Adversarial Networks

The ability of dealing with a wide variety of machine learning tasks has been greatly improved due to recent advances in deep learning algorithms such as Convolutional Neural Networks(CNNs).Unfortunately,the impressive performance mostly comes when a large amount of labeled data are available.Meanwhile,the cost of annotating the large amount of data is extremely expensive or even infeasible.The heavy cost of data labeling motivates us to establish effective algorithms to automatically perform annotation,typically by leveraging rich labeled data from a different but related source domain.For the problem of insufficient annotation data,this paper uses the adversarial method in generative adversarial networks to align the distribution of the two domains,and solves the problem of using the large amount of annotated data of the source domain to improve the classification task of the target domain when there is no labeled data in the target domain.Based on the generative adversarial networks,two transfer learning algorithms are proposed.The two algorithms are summarized as follows:(1)we present the Residual Adversarial Networks(RAN),which aligns features of the two domains by adversarial learning and supports classifier adaptation.Firstly,the deep features are extracted through the neural network,and the features of the two domains are matched by adversarial learning.The feature of the two domains are then fed to classifiers,and the residual modules are used to correlate the classifiers of the two domains.The overall framework can be solved efficiently by end-to-end back-propagation and generates more discriminative features.Experiments show that this method can achieve better results compared with traditional adversarial learning methods.(2)we present the Feature and Label Adversarial Networks(FLAN),which largely extends the ability of deep adversarial adaptation.FLAN considers the joint distributions of features and labels in both domains,generates domain-invariant features by generator and the features are discriminative enough.This method firstly extracts the deep features through the neural network,and then the classifier of the target domain generates a pseudo label of the target domain under the constraint of the entropy minimization.The feature information and the label information of the cross-domain are fused and then the two domains are matched by adversarial learning with feature and label information.Experimental results on standard unsupervised domain adaptation benchmarks have demonstrated that FLAN can outperform the state of art domain invariant representation learning methods.