Research map / Optimization / Implicit bias and generalization
Understanding Batch Normalization
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 |
|---|---|---|---|
| Enable batch normalization so the network can use a larger learning rate | 110-layer ResNet · CIFAR10 | test accuracy | Helped |
| Use the same low learning rate for BN and unnormalized networks (ablation) | 110-layer ResNet · CIFAR10 | test accuracy | No clear effect |
| Add batch normalization only at the network top (ablation) | 110-layer ResNet · CIFAR10 | test accuracy: 90.1% | Helped |
| Add the affine transformation parametrized by gamma and beta (ablation) | Batch normalization in a 110-layer ResNet · CIFAR10 | final accuracy: roughly 0.6 % | Helped |
| Remove mean and variance normalization (ablation) | Batch normalization in a 110-layer ResNet · CIFAR10 | accuracy | Hurt |
| Keep variance normalization but remove mean normalization (ablation) | Batch normalization in a 110-layer ResNet · CIFAR10 | learning rate: 0.01 | Hurt |
| Keep mean normalization but remove variance normalization (ablation) | Batch normalization in a 110-layer ResNet · CIFAR10 | learning rate: 0.003 | Hurt |
| Remove both mean and variance normalization (ablation) | Batch normalization in a 110-layer ResNet · CIFAR10 | learning rate: 0.0001 | Hurt |
| Update BN mean and variance only every other batch (ablation) | 110-layer ResNet with batch normalization · CIFAR10 | maximum viable learning rate: 0.01 | Hurt |
| Use a warm start for the unnormalized network (ablation) | Unnormalized ResNet · CIFAR10 | maximum viable learning rate: 0.01 | Helped |
| Increase the learning rate for the unnormalized network | 110-layer ResNet without BN · CIFAR10 | convergence | Hurt |
| Use a shallower unnormalized ResNet (ablation) | Unnormalized ResNet · CIFAR10 | maximum viable learning rate: 32 layer Resnet: 0.1; 56 layer Resnet: 0.01 | Helped |
Closest papers by meaning
- Towards Understanding Regularization in Batch Normalization (ICLR 2019 · similarity 0.94)
- Four Things Everyone Should Know to Improve Batch Normalization (ICLR 2020 · similarity 0.93)
- Batch normalization is sufficient for universal function approximation in CNNs (ICLR 2024 · similarity 0.93)
- Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion (ICLR 2024 · similarity 0.91)
- Reducing Divergence in Batch Normalization for Domain Adaptation (AAAI 2025 · similarity 0.91)
- Batch Normalization Is Blind to the First and Second Derivatives of the Loss (AAAI 2024 · similarity 0.90)
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.