Research map / Architectures and efficiency / Attention and transformers
Improving Vision Transformers by Revisiting High-Frequency Components
Jiawang Bai, Li Yuan, Shu-Tao Xia, Shuicheng Yan, Zhifeng Li, Wei Liu
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 |
|---|---|---|---|
| add HAT adversarial training | ViT-B · ImageNet-1K val | top-1 accuracy (%): 82.0 -> 83.2 | Improved |
| add HAT adversarial training | ViT-T · ImageNet-1K val | top-1 accuracy (%): 72.2 -> 73.3 | Improved |
| add HAT adversarial training | Swin-B · ImageNet-1K val | top-1 accuracy (%): 83.5 -> 84.0 | Improved |
| add HAT adversarial training | Swin-S · ImageNet-1K val | top-1 accuracy (%): 83.0 -> 83.3 | Improved |
| add HAT adversarial training | VOLO-D5 (512 resolution) · ImageNet-1K val | top-1 accuracy (%): 87.1 -> 87.3 | Improved |
| add high-frequency adversarial perturbations (filter size 10) to training (ablation) | ViT-B · ImageNet-1K val | top-1 accuracy (%): 82.0 -> 83.0 | Improved |
| add low-frequency adversarial perturbations to training (ablation) | ViT-B · ImageNet-1K val | top-1 accuracy (%) | No clear effect |
| combine HAT with knowledge distillation | ViT-B · ImageNet-1K val | top-1 accuracy (%): 83.2 -> 84.3 | Improved |
| train ViT-B with HAT instead of longer normal training at 2.3x cost (ablation) | ViT-B · ImageNet-1K val | top-1 accuracy (%): 82.4 -> 83.2 | Improved |
| increase epsilon from 2/255 to 5/255 (ablation) | HAT ViT-B · ImageNet-1K val | top-1 accuracy (%): 83.2 -> 83.0 | Hurt |
| add HAT with small training set (16K samples) (ablation) | ViT-B · ImageNet subset (1/8 of training set) | top-1 accuracy (%): +7.1 | Improved |
| add HAT to fine-tuning under SimMIM pre-training | ViT-B (SimMIM) · ImageNet-1K val | top-1 accuracy (%): 83.6 -> 83.9 | Improved |
Closest papers by meaning
- Early Convolutions Help Transformers See Better (NeurIPS 2021 · similarity 0.89)
- Scaling Vision Transformers (CVPR 2022 · similarity 0.89)
- Anti-Oversmoothing in Deep Vision Transformers via the Fourier Domain Analysis: From Theory to Practice (ICLR 2022 · similarity 0.88)
- When Vision Transformers Outperform ResNets without Pre-training or Strong Data Augmentations (ICLR 2022 · similarity 0.88)
- Towards Efficient Adversarial Training on Vision Transformers (ECCV 2022 · similarity 0.88)
- HIRI-ViT: Scaling Vision Transformer With High Resolution Inputs (TPAMI 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.
