Research map / Detection and segmentation / CNN features
Rectifying Magnitude Neglect in Linear Attention
Qihang Fan, Huaibo Huang, Yuang Ai, Ran He
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 |
|---|---|---|---|
| discard Q magnitude by Q/||Q|| in softmax attention (ablation) | DeiT-T softmax · ImageNet-1K | top-1 accuracy: 72.2 -> 70.0 | Hurt |
| add magnitude-aware beta (ablation) | MAViT-T · ImageNet-1K | top-1 accuracy: 52.3 -> 82.9 | Improved |
| add gamma offset (ablation) | MAViT-T · ImageNet-1K | top-1 accuracy: NaN -> 82.9 | Improved |
| fixed beta,gamma instead of learnable (ablation) | MAViT-T · ImageNet-1K | top-1 accuracy: 71.7 -> 82.9 | Improved |
| swap kernel ELU+1 for ReLU (ablation) | MAViT-T · ImageNet-1K | top-1 accuracy: 82.9 -> 82.8 | No clear effect |
| swap kernel ELU+1 for exp (ablation) | MAViT-T · ADE20K | mIoU: 47.6 -> 47.4 | No clear effect |
| replace InLine attention with MALA | Swin-S · ImageNet-1K | top-1 accuracy: 83.6 -> 85.3 | Improved |
| replace InLine attention with MALA | DeiT-T · ImageNet-1K | top-1 accuracy: 74.5 -> 75.1 | Improved |
| replace vanilla linear attention with MALA | Conformer(S) · LibriSpeech test-other | WER: 10.2 -> 5.3 | Improved |
Closest papers by meaning
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.93)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.90)
- ZeroS: Zero‑Sum Linear Attention for Efficient Transformers (NeurIPS 2025 · similarity 0.90)
- Theory, Analysis, and Best Practices for Sigmoid Self-Attention (ICLR 2025 · similarity 0.89)
- MetaLA: Unified Optimal Linear Approximation to Softmax Attention Map (NeurIPS 2024 · similarity 0.88)
- The Hedgehog & the Porcupine: Expressive Linear Attentions with Softmax Mimicry (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.
