Research map / Label-efficient and robust learning / Recognition and long-tail
Understanding and Improving Robustness of Vision Transformers through Patch-based Negative Augmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add P-Rotate negative views with uniform loss | ViT-B/16 · ImageNet-C | top-1 accuracy: 50.8 -> 52.4 (+1.6) | Helped |
| Add P-Rotate negative views with uniform loss | ViT-B/16 · ImageNet-A | top-1 accuracy: 6.7 -> 7.0 (+0.3) | No clear effect |
| Use L2 rather than uniform loss for P-Rotate negative views (ablation) | ViT-B/16 · ImageNet-C | top-1 accuracy: 52.4 -> 51.6 | Hurt |
| Add P-Rotate negative views with L2 loss | ViT-B/16 · ImageNet-A | top-1 accuracy: 6.7 -> 6.7 (+0.0) | No clear effect |
| Add P-Rotate negative views with uniform loss to Rand-Augment | ViT-B/16 + Rand-Augment · ImageNet-C | top-1 accuracy: 55.2 -> 56.4 (+0.8) | Helped |
| Add P-Rotate negative views with L2 loss to Rand-Augment | ViT-B/16 + Rand-Augment · ImageNet-R | top-1 accuracy: 23.0 -> 22.6 (-0.4) | Hurt |
| Add P-Rotate negative views with L2 loss to AugMix | ViT-B/16 + AugMix · ImageNet-C | top-1 accuracy: 57.8 -> 58.8 (+1.0) | Helped |
| Add patch-based negative examples to contrastive loss | ViT-B/16 + Contrastive* · ImageNet-R | top-1 accuracy: 22.8 -> 23.6 (+0.8) | Helped |
| Apply P-Shuffle / Uniform in both pre-training and fine-tuning rather than fine-tuning only (ablation) | ViT-B/16 + Rand-Augment · ImageNet-A | top-1 accuracy: 7.1 -> 7.7 | Helped |
| Apply P-Shuffle / Contrastive in both pre-training and fine-tuning rather than fine-tuning only (ablation) | ViT-B/16 + Rand-Augment · ImageNet-C | top-1 accuracy: 56.2 -> 57.8 | Helped |
| Apply positive Rand-Augment at fine-tuning as well as pre-training (ablation) | ViT-B/16 pretrained on ImageNet-21k · ImageNet-C | top-1 accuracy: 67.2 -> 67.0 | Hurt |
| Increase pre-training batch size from 4096 to 8192 (ablation) | ViT-B/16 · ImageNet-R | accuracy: 20.3 -> 19.3 | Hurt |
Closest papers by meaning
- Vision Transformers Are Robust Learners (AAAI 2022 · similarity 0.92)
- Intriguing Properties of Vision Transformers (NeurIPS 2021 · similarity 0.91)
- Vision Transformers Beat WideResNets on Small Scale Datasets Adversarial Robustness (AAAI 2025 · similarity 0.89)
- Improving Interpretation Faithfulness for Vision Transformers (ICML 2024 · similarity 0.88)
- Interpreting vision transformers via residual replacement model (NeurIPS 2025 · similarity 0.88)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 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.