eduzhai > Applied Sciences > Engineering >

Curriculum Manager for Source Selection in Multi-Source Domain Adaptation

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 17 pages

Abstract: The performance of Multi-Source Unsupervised Domain Adaptation dependssignificantly on the effectiveness of transfer from labeled source domainsamples. In this paper, we proposed an adversarial agent that learns a dynamiccurriculum for source samples, called Curriculum Manager for Source Selection(CMSS). The Curriculum Manager, an independent network module, constantlyupdates the curriculum during training, and iteratively learns which domains orsamples are best suited for aligning to the target. The intuition behind thisis to force the Curriculum Manager to constantly re-measure the transferabilityof latent domains over time to adversarially raise the error rate of the domaindiscriminator. CMSS does not require any knowledge of the domain labels, yet itoutperforms other methods on four well-known benchmarks by significant margins.We also provide interpretable results that shed light on the proposed method.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×