Research map / Architectures and efficiency / RNNs and normalization
Group Normalization
Yuxin Wu, Kaiming 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 |
|---|---|---|---|
| replace BN with GN at batch size 2 (ablation) | ResNet-50 · ImageNet val | top-1 error (%): 34.7 -> 24.1 | Improved |
| replace BN with GN at batch size 32 | ResNet-50 · ImageNet val | top-1 error (%): 23.6 -> 24.1 | Hurt |
| replace BN with GN at batch size 8 (ablation) | ResNet-50 · ImageNet val | top-1 error (%): 24.8 -> 24.0 | Improved |
| use GN instead of Batch Renorm at batch size 4 (ablation) | ResNet-50 · ImageNet val | top-1 error (%): 26.3 -> 24.2 | Improved |
| use grouped channels (2 per group) instead of per-channel IN (ablation) | GN ResNet-50 · ImageNet val | top-1 error (%): 28.4 -> 25.6 | Improved |
| use G=32 groups instead of G=1 (LN) (ablation) | GN ResNet-50 · ImageNet val | top-1 error (%): 25.3 -> 24.1 | Improved |
| replace BN with GN on ResNet-101 at batch size 2 (ablation) | ResNet-101 · ImageNet val | top-1 error (%): 31.9 -> 23.0 | Improved |
| replace frozen BN with GN in C4 Mask R-CNN (ablation) | Mask R-CNN C4 · COCO val2017 | box AP: 37.7 -> 38.8 | Improved |
| replace GN with LN in C4 Mask R-CNN (ablation) | Mask R-CNN C4 · COCO val2017 | box AP: 1.9 box AP worse than GN | Hurt |
| add GN to the box head with BN* backbone (ablation) | Mask R-CNN FPN · COCO val2017 | box AP: 38.6 -> 39.5 | Improved |
| add GN to the backbone (pretrained GN model) (ablation) | Mask R-CNN FPN · COCO val2017 | box AP: 39.5 -> 40.0 | Improved |
| train GN longer (180k to 270k iterations) (ablation) | R50 GN Mask R-CNN FPN · COCO val2017 | box AP: 40.3 -> 40.8 | Improved |
Closest papers by meaning
- Group Whitening: Balancing Learning Efficiency and Representational Capacity (CVPR 2021 · similarity 0.93)
- Filter Response Normalization Layer: Eliminating Batch Dependence in the Training of Deep Neural Networks (CVPR 2020 · similarity 0.90)
- Kalman Normalization: Normalizing Internal Representations Across Network Layers (NeurIPS 2018 · similarity 0.90)
- Towards Stabilizing Batch Statistics in Backward Propagation of Batch Normalization (ICLR 2020 · similarity 0.90)
- Beyond BatchNorm: Towards a Unified Understanding of Normalization in Deep Learning (NeurIPS 2021 · similarity 0.90)
- Dynamic Group Normalization: Spatio-Temporal Adaptation to Evolving Data Statistics (CVPR 2025 · 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.
