Research map / Architectures and efficiency / RNNs and normalization
Kalman Normalization: Normalizing Internal Representations Across Network Layers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Kalman Normalization at micro-batch size 4 | ResNet50 · ImageNet val, bs=4, moving stats | top-1 accuracy: 72.7 -> 76.1 | Helped |
| replace BN with Kalman Normalization at batch size 32 | ResNet50 · ImageNet val, bs=32 | top-1 accuracy: 76.4 -> 76.8 | Helped |
| replace BN with Group Normalization at batch size 32 | ResNet50 · ImageNet val, bs=32 | top-1 accuracy: 76.4 -> 75.9 | Hurt |
| replace BN with Kalman Normalization | Inceptionv2 · ImageNet val, bs=32 | steps to reach 73.1%: 170k -> 100k | Helped |
| replace BN* (frozen) with KN during fine-tuning | Mask RCNN ResNet50-conv4 · COCO 2017 val | APbbox / APmask: 36.7 -> 37.8 box, 32.1 -> 33.1 mask | Helped |
| add Kalman Normalization to Batch Norm (BKN) at bs=2 (ablation) | ResNet CIFAR-10 · CIFAR-10 | accuracy: 1.5% gain over BN | Helped |
| add Kalman Normalization to Group Norm (GKN) at bs=2 (ablation) | ResNet CIFAR-10 · CIFAR-10 | accuracy: 0.4% gain over GN | Helped |
| use online batch statistics instead of moving statistics at inference (ablation) | BN, batch size 1 · ImageNet InceptionV2, bs=1 | top-1 accuracy: 0.4% -> 47.99% (KN) | Helped |
| use online batch statistics instead of moving statistics at inference (ablation) | KN, CIFAR-10 bs=2 · CIFAR-10 Inception | accuracy: 90.9 -> 90.9 | No clear effect |
| use online batch statistics instead of moving statistics at inference (ablation) | BN, CIFAR-10 bs=128 · CIFAR-10 Inception | accuracy: 92.1 -> 90.0 | Hurt |
| replace BN with KN at small batch | Inception · CIFAR-100, bs=2 | accuracy: 63.8 -> 67.3 | Helped |
| replace BN with KN at small batch | Inception · SVHN, bs=2 | accuracy: 98.06 -> 98.16 | Helped |
Closest papers by meaning
- Four Things Everyone Should Know to Improve Batch Normalization (ICLR 2020 · similarity 0.90)
- Filter Response Normalization Layer: Eliminating Batch Dependence in the Training of Deep Neural Networks (CVPR 2020 · similarity 0.90)
- Normalizing the Normalizers: Comparing and Extending Network Normalization Schemes (ICLR 2017 · similarity 0.90)
- Towards Stabilizing Batch Statistics in Backward Propagation of Batch Normalization (ICLR 2020 · similarity 0.89)
- Exploring Normalization in Deep Residual Networks with Concatenated Rectified Linear Units (AAAI 2017 · similarity 0.89)
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.89)
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.