Research Of Deep Adversarial Domain Adaptation

Transfer learning is the main development direction of machine learning in the future,and unsupervised domain adaptation is an important part of transfer learning.In the case of lacking or even no label in the target domain,it is of great research significance and application value to solve the problem that the feature space and label space are the same but the distribution is different.Compared with the traditional machine learning,deep learning can extract the abstract features at the hierarchical level,and it does not need to design the feature extractor manually through engineering technology and professional domain knowledge,which simplifies the design and greatly strengthens the generalization ability of the model.The emergence of generative adversarial network makes full use of the computing power,which further extracts the abstract features and improves the classification accuracy.This thesis conducts a study on the unsupervised domain adaptation method based on deep adversarial learning,including the following two aspects.Firstly,self-corrected unsupervised domain adaptation(SCUDA)is proposed.After using the labeled samples of the source domain to train the model for several iterations,the model has the ability to preliminarily discriminate the target domain samples.The predicted label is not accurate enough,which called pseudo label in this paper,and additional means are needed to make the pseudo label become a "true label".Pseudo labels are only converted into weights or confidence to measure the degree of classification of the target domain samples and do not update actively.In this method,the pseudo label is regarded as a vector variable to minimize the inverse relative-entropy between the variable and the predicted value.Secondly,classifier-improved unsupervised domain adaptation(CIUDA)is proposed.In general,the output dimension of the classifier is the same as the actual number of categories in the dataset.Most of the existing methods follow this setting,so that the classifier only undertakes the pure classification task,and then adds other network components to complete the domain alignment task.In this method,a(K+1)-dimensional classifier is used to simplify the classical model,and the target domain samples are weighted to realize the simple sample priority strategy.At the same time,the adversarial learning is used to reject difficult samples and achieve the goal of domain alignment.