Unclipping CLIP's Wings: Avoiding Robustness Pitfalls in Multimodal Image Classification

Wanqian Yang, Rajesh Ranganath

Keywords: robustness, CLIP, spurious correlations, contrastive learning, multimodality

Abstract: Despite being pretrained on large-scale data, multimodal models such as CLIP can still learn spurious correlations. However, CLIP does not seem to learn the same spurious correlations as standard vision models, performing worse on some benchmark datasets (Waterbirds) yet better on others (CelebA). We investigate this discrepancy and find that CLIP's robustness on these datasets is highly sensitive to the choice of class prompts. Worst-group accuracy can be arbitrarily improved or worsened by making minute, single-word changes to prompts. We further provide evidence that the root cause of this phenomenon is \textit{coverage} --- using class prompts that are out-of-distribution with respect to pretraining can worsen spurious correlations. Motivated by these findings, we propose using class prompts that are generated from a public image-to-text model, such as BLIP. We show that performing $k$-nearest neighbors on these prompt embeddings improve downstream robustness without needing to fine-tune CLIP.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

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.

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: 7958