Robust Multi-modal Learning with Shifted Feature Reweighting Against Spurious Correlations

Chenyu You, Weicheng Dai, Yifei Min, Tiansheng Wen, Lawrence Hamilton Staib, James s Duncan

Keywords: spurious correlations; multi-modal learning; shortcut learning;

TL;DR: In this paper, we propose Shifted Feature Reweighting (SFR), a robust multi-modal learning method to mitigate the reliance on spurious feature with theoritical grounding.

Abstract:

Pre-trained multi-modal models have recently garnered significant attention due to their adaptability to diverse downstream tasks via fine-tuning.
However, their resilience to certain group shift issues, i.e., spurious correlations, remains imperative yet relatively under-investigated.
We study this problem in vision-language models (VLMs), and we observe potential vulnerabilities in pre-trained VLMs, such as CLIP, when confronted with spurious correlations.
While recent studies have been exploited to address unimodal group-imbalances by minority group up-sampling or creating group-balanced subsets, we posit that true robustness can be achieved by debiasing the training process through feature reweighting.
In this paper, we propose Shifted Feature Reweighting (SFR), a robust multi-modal learning method to mitigate the reliance on spurious features.
Specifically, we introduce a novel disagreement-based importance weight that allocates distinct weights to individual instances within the training data.
This contrasts with existing group rebalance weight strategies, which uniformly weigh all instances within a group.
Our reweighting strategy adeptly addresses disparities in instance-level learning difficulty.
Moreover, our empirical results unveil that representation collapse may arise during fine-tuning.
To address this, we proposed to introduce feature dropout and show that this simple method can further regularize the training on the majority groups and encourage the training on the minority groups.
Empirical results on multiple benchmarks verify our claims and confirm the effectiveness of our proposed SFR.
Theoretically, we analyze the performance of our SFR and confirm its superiority in mitigating spurious correlations.
Our codes will be here.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

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/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 781