Research map / Detection and segmentation / CNN features
Attentive Normalization
Xilai Li, Wei Sun, Tianfu Wu
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 |
|---|---|---|---|
| use RSD attention input instead of mean only (ablation) | ResNet50 AN (w/ BN) · ImageNet-1k val | top-1 error (%): 21.85 -> 21.59 | Improved |
| use RSD instead of concatenated (mean,std) (ablation) | ResNet50 AN (w/ BN) · ImageNet-1k val | top-1 error (%): 21.73 -> 21.59 | Improved |
| add BN after fc (Choice 2 instead of Choice 1) (ablation) | ResNet50 AN (w/ BN) · ImageNet-1k val | top-1 error (%): 21.76 -> 21.59 | Improved |
| use hsigmoid instead of softmax (ablation) | ResNet50 AN (w/ BN) · ImageNet-1k val | top-1 error (%): 21.72 -> 21.59 | Improved |
| use hsigmoid instead of relu (ablation) | ResNet50 AN (w/ BN) · ImageNet-1k val | top-1 error (%): 21.89 -> 21.59 | Improved |
| use hsigmoid instead of sigmoid (ablation) | ResNet50 AN (w/ BN) · ImageNet-1k val | top-1 error (%): 21.96 -> 21.59 | Improved |
| use K=(10,10,20,20) instead of (5,5,10,10) (ablation) | ResNet50 AN (w/ BN) · ImageNet-1k val | top-1 error (%): 21.92 -> 21.59 | Improved |
| use K=(10,10,20,20) instead of larger (20,20,40,40) (ablation) | ResNet50 AN (w/ BN) · ImageNet-1k val | top-1 error (%): 21.62 -> 21.59 | No clear effect |
| replace only the BN after last 3x3 conv instead of all BNs (ablation) | ResNet50 AN (w/ BN) · ImageNet-1k val | top-1 error (%): 22.15 -> 21.59 | Improved |
| replace BN with AN | ResNet50 · ImageNet-1k val (vanilla setup) | top-1 error (%): 23.01 -> 21.59 | Improved |
| replace BN with AN | DenseNet121 · ImageNet-1k val (vanilla setup) | top-1 error (%): 25.35 -> 22.62 | Improved |
| replace BN with AN | ResNet50 · ImageNet-1k val (state-of-the-art setup) | top-1 error (%): 21.08 -> 19.92 | Improved |
Closest papers by meaning
- Filter Response Normalization Layer: Eliminating Batch Dependence in the Training of Deep Neural Networks (CVPR 2020 · similarity 0.86)
- ACNe: Attentive Context Normalization for Robust Permutation-Equivariant Learning (CVPR 2020 · similarity 0.85)
- Towards More Robust Interpretation via Local Gradient Alignment (AAAI 2023 · similarity 0.85)
- Normalize Filters! Classical Wisdom for Deep Vision (NeurIPS 2025 · similarity 0.84)
- Domain-Aware Fine-Tuning: Enhancing Neural Network Adaptability (AAAI 2024 · similarity 0.84)
- EAN: An Efficient Attention Module Guided by Normalization for Deep Neural Networks (AAAI 2024 · similarity 0.84)
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.
