Research map / Architectures and efficiency / Attention and transformers
Are Vision Transformers Robust to Spurious Correlations?
Soumya Suvra Ghosal, Yixuan Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ViT-B/16 instead of BiT-M-R50x3 (both ImageNet-21k) | ViT-B/16 · Waterbirds | worst-group accuracy (%): 80.51 -> 89.30 | Improved |
| use ViT-B/16 instead of BiT-M-R50x3 (both ImageNet-21k) | ViT-B/16 · CelebA | worst-group accuracy (%): 89.80 -> 94.10 | Improved |
| pre-train ViT-B on ImageNet-21k instead of ImageNet-1k (DeiT-B) (ablation) | ViT-B/16 · Waterbirds | worst-group accuracy (%): 44.6 -> 89.3 | Improved |
| pre-train BiT-R50x3 on ImageNet-21k instead of ImageNet-1k (ablation) | BiT-R50x3 · Waterbirds | worst-group accuracy (%): 60.3 -> 80.5 | Improved |
| pre-train ViT-B on ImageNet-21k instead of ImageNet-1k (ablation) | ViT-B/16 · CelebA | worst-group accuracy (%): 88.0 -> 94.0 | Improved |
| pre-train BiT-M-R50x3 on ImageNet-21k instead of ImageNet-1k (ablation) | BiT-R50x3 · CelebA | worst-group accuracy (%): 88.3 -> 89.8 | Improved |
| use DeiT-B (ImageNet-1k transformer) instead of BiT-S-R50x3 CNN | DeiT-B/16 · Waterbirds | worst-group accuracy (%): 60.3 -> 44.6 | Hurt |
| scale model from ViT-Ti/16 to ViT-B/16 (ablation) | ViT · Waterbirds | worst-group accuracy (%): 71.65 -> 89.30 | Improved |
| use ViT-B/16 instead of BiT-M-R50x3 for spurious OOD detection | ViT-B/16 · Waterbirds OOD (energy score) | AUROC: 59.0 -> 91.0 | Improved |
| restrict self-attention to local neighbourhood at inference (masked attention) (ablation) | ViT-B/16 · Waterbirds and CelebA | worst-group accuracy (%) | Hurt |
| remove minority samples from fine-tuning data (data imbalance) | ViT-S/16 and BiT-M-R50x1 · Waterbirds | worst-group accuracy (%) | Hurt |
| fine-tune ViT longer (ablation) | ViT-S/16 · Waterbirds | minority accuracy | Improved |
Closest papers by meaning
- Towards Robust Vision Transformer (CVPR 2022 · similarity 0.91)
- Vision Transformers Are Robust Learners (AAAI 2022 · similarity 0.91)
- Vision Transformers Beat WideResNets on Small Scale Datasets Adversarial Robustness (AAAI 2025 · similarity 0.89)
- Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness (ICML 2024 · similarity 0.88)
- Improving Interpretation Faithfulness for Vision Transformers (ICML 2024 · similarity 0.87)
- Random Entangled Tokens for Adversarially Robust Vision Transformer (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
