Research map / Architectures and efficiency / Attention and transformers
An Impartial Take to the CNN vs Transformer Robustness Contest
Francesco Pinto, Philip H. S. Torr, Puneet K. Dokania
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 |
|---|---|---|---|
| swap CNN for transformer architecture (BiT-R152x4 to ViT-B) | ViT-B/16 · ImageNet-O OoD | AUROC: 80.07 -> 79.89 | Hurt |
| swap Swin-L for ConvNeXt-L | ConvNeXt-L · ImageNet-O OoD | AUROC: 87.76 -> 89.07 | Improved |
| swap top CNN for top transformer | ViT-L/16 · ImageNet-O OoD | AUROC: 90.04 -> 90.60 | No clear effect |
| flip positive class and use AUPR instead of AUROC (ablation) | ConvNeXt-XL · ImageNet-O OoD | AUPR: 97.19 -> 68.50 | Mixed |
| rebalance in/out samples (ablation) | ViT-B/16 · ImageNet-O OoD | AUPR: 36.77 -> 73.77 | Mixed |
| use ViT-B/16 instead of Swin-B | ViT-B/16 · ImageNet-1K | ECE: 8.40 -> 1.40 | Improved |
| swap ConvNeXt-XL for ViT-L16 under ImageNet-R shift | ViT-L16 · ImageNet-R | ECE: 2.92 -> 3.07 | Hurt |
| scale up BiT-R50x1 to BiT-R152x4 | BiT · ImageNet-1K | ECE: 3.49 -> 9.38 | Hurt |
| use ViT-L/16 instead of ViT-B/16 for simplicity bias | ViT-L/16 · R-CIFAR | accuracy: 99.79 -> 95.66 | Hurt |
| swap ViT-L16 for Swin-L under ImageNet-R | Swin-L · ImageNet-R | PRR: -34.12 -> 36.53 | Improved |
Closest papers by meaning
- Can CNNs Be More Robust Than Transformers? (ICLR 2023 · similarity 0.94)
- Transformers Learn Low Sensitivity Functions: Investigations and Implications (ICLR 2025 · similarity 0.90)
- Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness (ICML 2024 · similarity 0.89)
- Viewing Transformers Through the Lens of Long Convolutions Layers (ICML 2024 · similarity 0.88)
- Adventures of Trustworthy Vision-Language Models: A Survey (AAAI 2024 · similarity 0.88)
- ConvNet vs Transformer, Supervised vs CLIP: Beyond ImageNet Accuracy (ICML 2024 · similarity 0.88)
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.
