Learning a Reusable Meta Denoiser for Learning with Noisy Labels on Multiple Target Domains

Ming-Kun Xie, Bo-Shi Zou, Gang Niu, Chen-Chen Zong, Kun-Peng Ning, Masashi Sugiyama, Sheng-Jun Huang

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation 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: Learning with noisy labels, meta denoising, reusable meta denoiser

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Learning with noisy labels (LNL) is a classification problem, where some training data are mislabeled. To identify which data are mislabeled, many denoising (e.g., sample selection) methods have been proposed by exploring meta features (e.g., loss values) during training. They are successful, since the used meta features are informative for identifying mislabeled data. However, the useful meta features are discarded after training, which is a waste of resources if LNL is needed on more datasets. In this paper, we work on LNL with one clean source domain and multiple noisy target domains and propose a general framework called meta denoising (MeDe), where the input spaces and/or label sets can be different for the source and target domains. Specifically, we find that some meta features are nearly transferable across datasets; thus, we train a reusable meta denoiser, which is a binary classifier to identify mislabeled data given meta features, by simulating noisy labels on the source domain; then, we can run the meta denoiser on any target domain by extracting its own meta features. Experiments show that MeDe can denoise datasets with different label sets and outperform denoising methods applied on each dataset separately.

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: 5158