Research map / Optimization / Implicit bias and generalization
Batch Normalization Preconditioning for Neural Network Training
Susanna Lange, Kyle Helfrich, Qiang Ye
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 BNP instead of BN at mini-batch size 6 | fully connected network · CIFAR10 FC (batch 6) | test accuracy | Improved |
| use BNP instead of BN at mini-batch size 2 | 5-layer CNN · CIFAR10 CNN (batch 2) | test accuracy | Improved |
| use BNP at mini-batch size 1 where BN does not train | 5-layer CNN · CIFAR10 CNN (batch 1) | test accuracy | Improved |
| use BNP instead of BN with large mini-batch (60) | fully connected network · MNIST FC (batch 60) | test accuracy | No clear effect |
| use BNP instead of BN with large mini-batch (60) | fully connected network · CIFAR10 FC (batch 60) | test accuracy | Mixed |
| use BNP instead of BN at mini-batch size 128 | 5-layer CNN · CIFAR10 CNN (batch 128) | test accuracy | No clear effect |
| add BNP preconditioning to a vanilla network | fully connected network · CIFAR10 FC (batch 60) | Hessian condition number | Improved |
| add BNP to a vanilla network | fully connected network · CIFAR10 FC | test accuracy and training loss | Improved |
| add GN to BNP on ResNet-110 (BNP+GN vs BNP alone) (ablation) | ResNet-110 · CIFAR10 | test accuracy at learning rate decay | Improved |
| use BNP+GN instead of BN, before first learning rate decay | ResNet-110 · CIFAR10 | test accuracy | Improved |
| use BNP+GN instead of BN, after learning rate decay | Preactivation ResNet-110 · CIFAR100 | test accuracy | Hurt |
| use BNP+GN instead of BN, after second decay | ResNet-18 · ImageNet | test accuracy | Hurt |
Closest papers by meaning
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.92)
- 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)
- Demystifying Batch Normalization in ReLU Networks: Equivalent Convex Optimization Models and Implicit Regularization (ICLR 2022 · similarity 0.91)
- Generalized Batch Normalization: Towards Accelerating Deep Neural Networks (AAAI 2019 · similarity 0.90)
- Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion (ICLR 2024 · 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.
