Research On Unsupervised Cross-Domain Learning Methods

With the rapid growth of data scale and computing resources,machine learning has become one of the leading technical cornerstones of big data analysis,computer vision,and natural language processing tasks.Most machine learning methods usually assume that the data comes from the same distribution.However,data can often be divided into different but related subgroups according to their distribution in practical application scenarios.Such subgroups are often called domains.It is not difficult for humans to map and translate data and knowledge from some domains(called source domain)to other domains(called target domain).For machine learning,achieving such cross-domain data mapping and knowledge transfer usually requires a large amount of label information to train a cross-domain mapping/translation model.However,obtaining the label consumes a lot of human resources and material resources,and in some cases,the label cannot be obtained at all.Therefore,how to realize cross-domain data mapping and knowledge transfer without label information has become an urgent problem to be solved in machine learning.Unsupervised cross-domain learning is proposed to solve this problem.Unsupervised cross-domain learning relaxes the constraint that data must come from the same domain in traditional machine learning,which brings many opportunities to industry and academia and faces many challenges.This thesis focuses on the tasks of domain mapping,domain adaptation,and domain generalization in unsupervised cross-domain learning,and studies how to mine domain-invariant essential features and structures among different domains,as well as the unique characteristics of each domain,so that data translation and knowledge transfer can be achieved across domains.The main research contents and innovations of this paper can be summarized as follows:(1)An unsupervised video domain mapping method is proposed based on motion guidance.Unsupervised video domain mapping aims to learn a mapping such that the video can be transformed from the source domain to the target domain without any paired training samples.This video domain mapping requires not only the visual appearance of each frame of the mapped video to be realistic but also the motion between consecutive frames.To this end,this thesis proposes a new motion-guided recurrent generative adversarial network,which innovatively introduces motion estimation into the unsupervised domain mapping task.The network uses three constraints:1)Adversarial constraints transform the source domain video frames to the target domain through a max-min training strategy of two networks of generator and discriminator.2)Video frame and inter-frame motion loop consistency constraints implicitly guarantee that the generated target domain video retains the semantic and motion information of the input source domain video.3)The motion transition constraint further enhances the temporal continuity of the generated target domain videos.Experimental results show that this method can effectively improve video domain mapping quality,and achieve better performance than existing unsupervised video domain mapping models on multiple standard datasets.(2)An unsupervised domain adaptation method is proposed based on transferable contrastive learning.Unsupervised domain adaptation aims to learn a model that performs well in the target domain using labeled source domain data and many unlabeled target domain data.A recent advance in domain adaptation is using self-supervised learning to improve the cross-domain invariance of unlabeled target domain data features.However,most existing domain adaptation methods treat self-supervised learning as an independent auxiliary component without considering that its own goal is to eliminate domain distribution discrepancies.Instead,this thesis proposes a self-supervised learning paradigm tailored specifically for domain adaptation tasks,i.e.,transferable contrastive learning.This paradigm organically combines self-supervised learning and domain adaptation,which can simultaneously improve the classification model’s crossdomain transferability and discrimination ability.Experimental results show that the thesis achieves state-of-the-art performance on multiple unsupervised domain adaptation benchmark datasets.(3)A domain generalization method is proposed based on style and semantic memory mechanism.Domain generalization aims to learn a model with strong generalization ability from several source domain data to achieve better results on target domain data that has not been seen during training.Current state-of-the-art domain generalization algorithms often prioritize the assumption of semantic invariance across domains,while ignoring the inherent style invariance within domains.This thesis finds that exploiting intra-domain style invariance is crucial to improving the ability of the model to generalize over the domain.This thesis uses a style and semantic memory mechanism to store the style and semantic features of different samples in different domains and proposes an instance-level style contrast loss and a "ury" mechanism for intra-domain invariant style feature learning and inter-domain invariant semantic feature learning.Ultimately,the semantic features learned by the model can generalize well to the target domain.The experimental results show that the performance of this method on multiple domain generalization benchmark datasets has reached the top level.