Research map / Architectures and efficiency / Attention and transformers
Less Is More: Pay Less Attention in Vision Transformers
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 |
|---|---|---|---|
| Replace early-stage self-attention with MLP blocks (ablation) | PVT-S architecture · ImageNet val | Top-1 accuracy: 79.8 -> 80.4 | Helped |
| Replace early-stage self-attention with MLP blocks (ablation) | Swin-Ti architecture · ImageNet val | Top-1 accuracy: 81.3 -> 81.3 | No clear effect |
| Replace uniform patch merging with deformable token merging (ablation) | PVT-S · ImageNet val | Top-1 accuracy: 79.8 -> 80.5 | Helped |
| Replace uniform patch merging with deformable token merging (ablation) | Swin-Ti · ImageNet val | Top-1 accuracy: 81.3 -> 81.6 | Helped |
| Remove self-attention from stage 1 (ablation) | PVT-S with standard MSA · ImageNet val | Top-1 accuracy: 80.9 -> 80.8 | No clear effect |
| Remove self-attention from stages 1 and 2 (ablation) | PVT-S with standard MSA · ImageNet val | Top-1 accuracy: 80.9 -> 80.4 | Hurt |
| Keep self-attention in stages 1 and 2 (ablation) | PVT-S · ImageNet val | Top-1 accuracy: 80.4 -> 80.9 | Helped |
| Remove self-attention from stages 1, 2, and 3 (ablation) | PVT-S with standard MSA · ImageNet val | Top-1 accuracy: 80.9 -> 75.0 | Hurt |
| Remove self-attention from all stages (ablation) | PVT-S with standard MSA · ImageNet val | Top-1 accuracy: 80.9 -> 66.8 | Hurt |
| Replace SRA with standard MSA in all stages (ablation) | PVT-S · ImageNet val | Top-1 accuracy: 79.8 -> 80.9 | Helped |
| Use the LIT-Ti architecture | ImageNet classification · ImageNet val | Top-1 accuracy: 79.8 -> 81.1 | Helped |
| Use LIT-Ti as the backbone | RetinaNet · COCO val2017 | AP: 40.4 -> 41.6 | Helped |
Closest papers by meaning
- MLP Can Be A Good Transformer Learner (CVPR 2024 · similarity 0.90)
- A Theoretical Understanding of Shallow Vision Transformers: Learning, Generalization, and Sample Complexity (ICLR 2023 · similarity 0.90)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.90)
- Vision Transformer With Deformable Attention (CVPR 2022 · similarity 0.89)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.89)
- Vision Transformers Need Registers (ICLR 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.