Prototype-Guided Feature Learning for Unsupervised Domain Adaptation

Abstract

Unsupervised Domain Adaptation transfers knowledge from the source domain to the target domain. It makes remarkable progress in alleviating the label-shortage problem in machine learning. Existing methods focus on aligning the two domain distributions directly. However, due to domain discrepancy, there may be some samples in the source domain being unnecessary or even harmful to the target tasks. Avoiding transferring knowledge from these samples is crucial. Existing researches are limited in this area. To this end, we propose a new unsupervised domain adaptation approach named the prototype-guided feature learning. The proposed method contains three main innovations. Firstly, we propose to utilize the more representative source-domain samples, class prototypes, to learn a domain-invariant subspace with the target samples. Secondly, the modified nearest class prototype method is proposed to predict the target samples by exploiting the structural information of the target domain efficiently. Thirdly, a multi-stage label filtering method is proposed to alleviate the mislabeling problem during training. Extensive experiments manifest that our method is competitive compared to the current mainstream unsupervised domain adaptive methods.