Research map / Architectures and efficiency / Attention and transformers
Towards Robust Vision Transformer
Check 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 |
|---|---|---|---|
| combine robust architectural components and add PAAS plus patch-wise augmentation | RVT-S · ImageNet-1K | FGSM accuracy: 51.3 -> 51.8 | Helped |
| combine robust architectural components and add PAAS plus patch-wise augmentation | RVT-S · ImageNet-1K | Top-1 accuracy: 81.7 -> 81.9 | Helped |
| add a convolutional FFN (ablation) | DeiT-Ti · ImageNet-1K | robust accuracy: 22.3 -> 31.9 | Helped |
| replace global self-attention with local self-attention (ablation) | DeiT-Ti · ImageNet-1K | robust accuracy: 22.3 -> 21.0 | Hurt |
| replace CLS token with output-token pooling (ablation) | DeiT-Ti · ImageNet-1K | robust accuracy: 22.3 -> 28.4 | Helped |
| replace linear patch embedding with convolutional patch embedding (ablation) | DeiT-Ti · ImageNet-1K | robust accuracy: 22.3 -> 23.2 | Helped |
| replace linear patch embedding with T2T embedding (ablation) | DeiT-Ti · ImageNet-1K | robust accuracy: 22.3 -> 25.4 | Helped |
| increase attention heads from 1 to 8 (ablation) | DeiT-Ti · ImageNet-1K | robust accuracy: 17.6 -> 25.2 | Helped |
| increase attention heads from 8 to 12 (ablation) | DeiT-Ti · ImageNet-1K | robust accuracy: 25.2 -> 24.7 | Hurt |
| move two transformer blocks to the later, lower-resolution stage (V2) (ablation) | DeiT-Ti · ImageNet-1K | robust accuracy: 22.3 -> 24.3 | Helped |
| use learned absolute position encoding instead of no position encoding (ablation) | DeiT-Ti · ImageNet-1K | robust accuracy: 15.8 -> 22.3 | Helped |
| replace learned absolute position encoding with input-conditioned position encoding (ablation) | DeiT-Ti · ImageNet-1K | FGSM accuracy: 22.3 -> 21.5 | Hurt |
Closest papers by meaning
- Vision Transformers Are Robust Learners (AAAI 2022 · similarity 0.93)
- 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)
- Towards Understanding and Improving Adversarial Robustness of Vision Transformers (CVPR 2024 · similarity 0.88)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.88)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · 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.