Research map / Architectures and efficiency / RNNs and normalization
Revisiting Batch Norm Initialization
Jim Davis, Logan Frank
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 |
|---|---|---|---|
| initialize BN gamma at 0.1 with gamma lr divided by 100 | ResNet-18 · CIFAR-10, lr 0.1 | test accuracy (%): 89.44 -> 90.80 | Improved |
| initialize BN gamma at 0.05 with gamma lr divided by 100 | ResNet-18 · CIFAR-10, lr 0.01 | test accuracy (%): 84.64 -> 88.84 | Improved |
| initialize BN gamma at 0.01 with gamma lr divided by 100 | ResNet-18 · CIFAR-10, lr 0.001 | test accuracy (%): 71.32 -> 80.37 | Improved |
| add gamma below 1 initialization on top of gamma lr reduction (ablation) | ResNet-18 · CIFAR-10, lr 0.1 | test accuracy (%): 89.81 -> 90.80 | Improved |
| add gamma lr reduction (c=100) to gamma<1 initialization (ablation) | ResNet-18 · CIFAR-10 | test accuracy (%) | Improved |
| increase learning rate for BN parameters only (ablation) | BN parameters · CIFAR-10 | test accuracy (%) | Hurt |
| scale only normalized data (standard form) instead of scaling beta too (A1, A2) (ablation) | BN affine · CIFAR-10, low lr | test accuracy (%): up to 7% gain | Improved |
| apply method on ResNet-101 | ResNet-101 · CIFAR-10, lr 0.1 | test accuracy (%): 88.28 -> 91.58 | Improved |
| apply method on ResNeXt-50 | ResNeXt-50 · CIFAR-10, lr 0.1 | test accuracy (%): 88.60 -> 92.04 | Improved |
| apply method to Stanford Cars at lr 0.01 | ResNet-50 · Stanford Cars, lr 0.01 | test accuracy (%): 51.86 -> 51.18 | No clear effect |
| use proposed gamma init instead of RBN scale | ResNet-34 · CIFAR-100, lr 0.01 | test accuracy (%): 58.95 -> 64.01 | Improved |
| use instance-specific statistics in RBN/IEBN scaling (ablation) | RBN/IEBN · CIFAR-10 | test accuracy (%): 84.72 -> 84.50 (RBN vs RBN-, lr 0.01) | No clear effect |
Closest papers by meaning
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.94)
- BNET: Batch Normalization With Enhanced Linear Transformation (TPAMI 2023 · similarity 0.91)
- Re-Thinking the Effectiveness of Batch Normalization and Beyond (TPAMI 2023 · similarity 0.91)
- Towards Understanding Regularization in Batch Normalization (ICLR 2019 · similarity 0.91)
- Momentum Batch Normalization for Deep Learning with Small Batch Size (ECCV 2020 · similarity 0.91)
- Reducing Divergence in Batch Normalization for Domain Adaptation (AAAI 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.
