Semi-supervised Domain Adaptation via Joint Error based Triplet Alignment
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: domain adaptation, semi-supervised learning, joint error
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Existing domain adaptation methods are very effective in aligning feature distributions. However, these techniques usually do not improve the performance that much when a few annotated examples are available in the target domain. To address this semi-supervised domain adaptation scenario, we propose a novel joint error based triplet alignment approach that simultaneously optimizes the classification loss as well as the joint error among the source, labeled and unlabeled target domains. Besides, we propose a novel dissimilarity measurement between two classifiers, namely maximum cross margin discrepancy, which can asymptotically bridge the gap between the theory and algorithm. We empirically demonstrate the superiority of our method over several baselines.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3932
Loading