Research map / Architectures and efficiency / Attention and transformers
Selection-as-Nonlinearity: Bridging Attention and Activation via a Joint Game-Decision Lens for Interpretable, Discriminative Visual Representations
Sudong Cai, Shuai Yuan, Bingzhi Chen, Rui Mao, Bing Wang
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 |
|---|---|---|---|
| remove FFN layers (attention-only stack) (ablation) | Swin-Min · ImageNet-1K val | top-1 accuracy (%): 72.2 -> 62.8 | Hurt |
| widen the attention-only model to match parameters (ablation) | attention-only Swin-Min · ImageNet-1K val | top-1 accuracy (%): 62.8 -> 67.5 | Improved |
| apply CSaN to attention-only stack (ablation) | attention-only Swin-Min · ImageNet-1K val | top-1 accuracy (%): 62.8 -> 72.4 | Improved |
| add CSaN to Swin | Swin-Min · ImageNet-1K val | top-1 accuracy (%): 72.2 -> 75.0 | Improved |
| add CSaN to Swin | Swin-Tiny · ImageNet-1K val | top-1 accuracy (%): 81.3 -> 82.7 | Improved |
| add CSaN to ViT | ViT-Base-Slim/16 · ImageNet-1K val | top-1 accuracy (%): 80.7 -> 82.1 | Improved |
| add CSaN to Hiera | Hiera-Small · ImageNet-1K val | top-1 accuracy (%): 81.3 -> 82.5 | Improved |
| add CSaN backbone for detection | Swin-Tiny with RetinaNet · COCO 1x | mAP (%): 37.3 -> 38.0 | Improved |
| add 7x7 depth-wise convolution after CSaN output (CSaN-V2) | Swin-Min with CSaN · ImageNet-1K val | top-1 accuracy (%): 75.0 -> 77.1 | Improved |
| add 7x7 depth-wise convolution after CSaN output (CSaN-V2) | Swin-Tiny with CSaN · ImageNet-1K val | top-1 accuracy (%): 82.7 -> 83.0 | Improved |
| use unit-wise instead of head-wise weights in CSaN (ablation) | Swin-Tiny CSaN · ImageNet-1K val | top-1 accuracy (%): 82.5 -> 82.7 | Improved |
Closest papers by meaning
- Towards Interpretable and Efficient Attention: Compressing All by Contracting a Few (NeurIPS 2025 · similarity 0.88)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.88)
- Selective Attention: Enhancing Transformer through Principled Context Control (NeurIPS 2024 · similarity 0.88)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.87)
- STILL: Selecting Tokens for Intra-Layer Hybrid Attention to Linearize LLMs (NeurIPS 2026 · similarity 0.87)
- The Hedgehog & the Porcupine: Expressive Linear Attentions with Softmax Mimicry (ICLR 2024 · 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.
