Research map / Architectures and efficiency / Attention and transformers
Poly-NL: Linear Complexity Non-Local Layers With 3rd Order Polynomials
Francesca Babiloni, Ioannis Marras, Filippos Kokkinos, Jiankang Deng, Grigorios Chrysos, Stefanos Zafeiriou
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 Poly-NL before last block of Res4 | Mask R-CNN ResNet-50 · COCO | AP box: 37.9 -> 39.2 | Improved |
| add Poly-NL before last block of Res4 | Mask R-CNN ResNet-50 · COCO | AP mask: 34.6 -> 35.4 | Improved |
| use Poly-NL instead of TESA | Mask R-CNN ResNet-50 · COCO | AP box: 39.5 -> 39.2 | Hurt |
| add Poly-NL at Res3, Res4 and Res5 instead of Res4 only (ablation) | Mask R-CNN ResNet-50 · COCO | AP box: 39.2 -> 39.8 | Improved |
| add Poly-NL at Res3 only instead of Res4 (ablation) | Mask R-CNN ResNet-50 · COCO | AP box: 39.2 -> 38.6 | Hurt |
| add Poly-NL at Res5 only instead of Res4 (ablation) | Mask R-CNN ResNet-50 · COCO | AP box: 39.2 -> 38.7 | Hurt |
| add Poly-NL at Res3 only instead of none (ablation) | Mask R-CNN ResNet-50 · COCO | AP mask: 34.6 -> 35.2 | Improved |
| add Poly-NL to ResNet-50 | ResNet-50 · ImageNet | top-1 accuracy: 75.62 -> 76.30 | Improved |
| use Poly-NL instead of Non-local | ResNet-50 · ImageNet | top-1 accuracy: 76.09 -> 76.30 | Improved |
| use Poly-NL instead of TESA | ResNet-50 · ImageNet | top-1 accuracy: 76.49 -> 76.30 | Hurt |
| add Latent-GNN to ResNet-50 | ResNet-50 · ImageNet | top-1 accuracy: 75.62 -> 75.28 | Hurt |
| add Poly-NL to ResNet-50 | RetinaNet-style detector · WIDER FACE Hard | AP: 89.87 -> 92.76 | Improved |
Later papers that compared with this method
1 later accepted paper reports PolyNL as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- QT-ViT: Improving Linear Attention in ViT with Quadratic Taylor Expansion (NeurIPS 2024 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Linear Complexity Self-Attention with $3^{\text{rd}}$ Order Polynomials (TPAMI 2023 · similarity 0.90)
- The Multi-Block DC Function Class: Theory, Algorithms, and Applications (NeurIPS 2026 · similarity 0.84)
- On the Hardness of Training Deep Neural Networks Discretely (AAAI 2025 · similarity 0.84)
- AsCAN: Asymmetric Convolution-Attention Networks for Efficient Recognition and Generation (NeurIPS 2024 · similarity 0.84)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.84)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.83)
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.
