Research map / Architectures and efficiency / Attention and transformers
Sparsifiner: Learning Sparse Instance-Dependent Attention for Efficient 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 |
|---|---|---|---|
| Add instance-dependent sparse attention alongside EViT token pruning | DeiT-S · ImageNet, 224×224; token keep rate 0.7 | MHSA FLOPs: 193.1 -> 113.3 MFLOPs | Helped |
| Add instance-dependent sparse attention alongside EViT token pruning | DeiT-S · ImageNet, 224×224; token keep rate 0.7 | top-1 accuracy: 79.5 -> 79.5 | No clear effect |
| Add instance-dependent sparse attention alongside EViT token pruning | DeiT-S · ImageNet, 224×224; token keep rate 0.5 | MHSA FLOPs: 149.1 -> 86.6 MFLOPs | Helped |
| Add instance-dependent sparse attention alongside DynamicViT token pruning | DeiT-S · ImageNet, 224×224; token keep rate 0.5 | top-1 accuracy: 77.3 -> 78.4 | Helped |
| Use sparse full-rank attention reconstruction instead of direct low-rank attention (ablation) | DeiT-S · ImageNet, 224×224 | top-1 accuracy: 77.54 -> 79.79 | Helped |
| Reduce the attention budget to keep rate 0.1 (ablation) | Sparsifiner-S · ImageNet, 224×224 | top-1 accuracy: 79.82 -> 79.42 | Hurt |
| Reduce the attention budget to keep rate 0.05 (ablation) | Sparsifiner-S · ImageNet, 224×224 | top-1 accuracy: 79.82 -> 78.75 | Hurt |
| Reduce the attention budget to keep rate 0.01 (ablation) | Sparsifiner-S · ImageNet, 224×224 | top-1 accuracy: 79.82 -> 73.03 | Hurt |
| Reduce the attention budget to keep rate 0.25 (ablation) | DeiT-T · ImageNet, 384×384 | top-1 accuracy: 75.45 -> 75.35 | No clear effect |
| Reduce the attention budget to keep rate 0.1 (ablation) | DeiT-T · ImageNet, 384×384 | top-1 accuracy: 75.45 -> 74.58 | Hurt |
Closest papers by meaning
- Chasing Sparsity in Vision Transformers: An End-to-End Exploration (NeurIPS 2021 · similarity 0.90)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.89)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.89)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.88)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.88)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (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.