BOT: Bootstrapped Optimal Transport for Multi-label Noise Learning
Keywords: multi-label learning, label noise
TL;DR: This paper presents the Bootstrapped Optimal Transport method to improve multi-label learning with label noise, outperforming existing methods on various datasets through a flexible Optimal Transport formulation.
Abstract: Multi-label learning with label noise is a practical but more challenging problem, as the underlying label dependency complicates the modeling from clean labels to noisy variants. Progress in this area is usually explored from the perspectives of semi-supervised learning, robust loss functions, or noise transition, which are less effective on complicated datasets or highly sensitive to transition matrix estimation. To refine the noisy labels in a general framework, we propose a simple but effective method, named Bootstrapped Optimal Transport method (BOT). Unlike the \emph{explicit} linear transition matrix with stringent conditions, BOT considers the modeling between true labels and noisy labels as an \emph{implicit} optimal transport procedure which has a more powerful degree of freedom. We show that with the proper reference by bootstrapping and adversarial orientation, the underlying true labels can be effectively estimated for training by the Sinkhorn-Knopp algorithm. Despite the simplicity, extensive experiments on a range of benchmark datasets prove that BOT consistently outperforms state-of-the-art methods, and comprehensive ablations explain the success behind BOT.
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
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: 3026
Loading