Research On Universal Domain Adaptation Based On Unknown Detection

Unsupervised Domain Adaptation(UDA)aims to leverage a fully labeled source domain to assist learning on unlabeled target domain,and has received widespread attention in the field of machine learning.Traditional UDA learning assumes that the category space of source and target domain data is exactly same,and only the distribution shift cross domains needs to be addressed.However,in reality,it is difficult for the categories of target domain to be completely consistent with that of source domain,i.e.,there may exist private categories in either the source or target domain.In such cases,there exist both distribution shift and category discrepancy between domains,and directly aligning the domain distribution may misclassify the private categories in the target domain to the known source domain categories,leading to negative transfer.To address both distribution shift and category discrepancy in UDA,i.e.,the Universal Domain Adaptation(Uni DA)learning scenario,this paper mainly proposes the following two parts of research:Firstly,an OSDA method called “Cross-domain Self-supervised Open-set Domain Adaptation based on Contrastive Learning(CSCL)” is proposed.The CSCL method simultaneously addresses the distribution shift and category difference between the source and target domains within a unified contrastive learning framework.The discriminative and well-clustered feature representations are obtained based on contrastive learning for both domains.Then,a similarity criterion is used to guide the detection of unknown classes in the target domain,and discriminative features for shared and unknown classes are further learned.Secondly,an Uni DA method called “Towards Adaptive Unknown Authentication for Universal Domain Adaptation by Classifier Paradox(UACP)” is proposed.The UACP method utilizes a multiclass classifier to classify samples into one of the source domain classes,and the corresponding binary classifier to further verify the sample’s class.By leveraging the prediction paradox between the two classifiers,UACP adaptively identifies unknown class samples in the target domain.Additionally,implicit domain alignment is performed in the output space to enable cross-domain samples to share the same decision boundary.Experiments on datasets MNIST-USPS,Office-31,Office-Home,Domain Net and Vis DA show the effectiveness of the UACP and CSCL under different scenarios.