Data-Efficient Group Robustness via Out of Distribution Concepts

Yiwei Yang, Anthony Zhe Liu, Robert Wolfe, Aylin Caliskan, Bill Howe

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: Spurious Correlations, Robustness, Group Shifts

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

Abstract: Deep Neural Networks are prone to capturing correlations between spurious attributes and class labels, leading to low accuracy on some groups of the data. Existing methods rely on group labels either during training or validation to improve the model's robustness to spurious correlation. We propose Concept DRO, a method to mitigate spurious correlations even when both training and validation group labels are unknown. We first observe that data points can be separated into groups using concepts, sets of curated images representing the spurious attribute. We thus leverage concepts to estimate group labels, and then use these labels to train with a distributively robust optimization objective. We empirically show that Concept DRO outperforms existing methods that do not require group labels.

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: 9069