Generalized Representation Learning Method With Reduced Domain Gap

The optimization of deep neural networks is inseparable from the supervised training of labeled data,which undoubtedly requires a lot of resources to collect and label large-scale data.Recently,domain adaptation has received widespread attention,which transfers the knowledge learned from the labeled source domain data to the unlabeled target domain.This paper mainly focuses on reducing the domain gap and extracting generalized features and proposes a corresponding solution for the semi-supervised domain adaptation and multi-source unsupervised domain adaptation,respectively.Aiming at the problem that a few labeled target domain samples in the semisupervised domain adaptation affect the alignment of the source domain and the target domain,this paper constructs a graph-in-graph module to model the internal relationship of the input features,to extract rich and generalized features.In addition,for a large amount of unlabeled data in the target domain,this paper uses contrastive loss to optimize the network to extract generalized representations.This method solves the problem of generalized feature extraction,and the experimental results on multiple public datasets strongly prove its effectiveness.Towards the problem that the domain gap in source domains hinders the alignment between source domains and the target domain in multi-source unsupervised domain adaptation,this paper proposes a progressive alignment framework.For low-level features,this paper designs a 2-way domain classifier to distinguish the source domains from the target domain.For high-level features,this paper uses a multi-class domain classifier to specifically distinguish which source domain or target domain the sample comes from.Combined with adversarial training,the domain gap between the multisource domain and the target domain is gradually reduced.In addition,this paper designs a sample importance learning module to select source domain samples that are conducive to knowledge transfer.This method realizes the alignment of multiple source domains and the target domain,and sufficient experiments prove that the method proposed in this paper is superior to other baseline methods.