Research map / Architectures and efficiency / CNN design
Training BatchNorm and Only BatchNorm: On the Expressive Power of Random Features in CNNs
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 |
|---|---|---|---|
| Train only BatchNorm γ and β, freezing the other weights at random initialization | ResNet-110 · CIFAR-10 | test accuracy | Helped |
| Increase depth while training only BatchNorm γ and β | ResNet · CIFAR-10 | accuracy: 48% -> 82% | Helped |
| Increase width while training only BatchNorm γ and β | WRN · CIFAR-10 | accuracy: 48% -> 73% | Helped |
| Increase depth while training only BatchNorm γ and β | ResNet · ImageNet | top-5 accuracy: 17% -> 32% | Helped |
| Train two random parameters per channel instead of the two BatchNorm affine parameters (ablation) | ResNet · CIFAR-10 | accuracy: 17 to 21 percentage points lower | Hurt |
| Train two random parameters per channel instead of the two BatchNorm affine parameters (ablation) | ResNet · ImageNet | top-5 accuracy: never exceeds 4% | Hurt |
| Also train the output layer alongside BatchNorm (ablation) | ResNet · ImageNet | top-5 accuracy: about 25 percentage points; maximum value 57% | Helped |
| Train the output layer alone, without BatchNorm (ablation) | ResNet-200 · ImageNet | top-5 accuracy: 2.7% | Hurt |
| Initialize BatchNorm with β=1 instead of β=0 (ablation) | BatchNorm-only ResNets · CIFAR-10 and ImageNet | accuracy: 1 to 3 percentage points | Helped |
| Initialize BatchNorm with β=1 instead of β=0 (ablation) | ResNet-200 with all parameters trainable · ImageNet | top-5 accuracy: 0.5% | Hurt |
| Initialize γ=1, β=0 or γ sampled from [-1, 1], β=0 (ablation) | BatchNorm-only ResNets · CIFAR-10 and ImageNet | accuracy | No clear effect |
| Clamp BatchNorm γ values with |γ| < 0.01 to zero (ablation) | ResNets · CIFAR-10 and ImageNet | accuracy | No clear effect |
Closest papers by meaning
- Deconstructing the Regularization of BatchNorm (ICLR 2021 · similarity 0.90)
- Beyond BatchNorm: Towards a Unified Understanding of Normalization in Deep Learning (NeurIPS 2021 · similarity 0.90)
- How Does Batch Normalization Help Optimization? (NeurIPS 2018 · similarity 0.89)
- MetaNorm: Learning to Normalize Few-Shot Batches Across Domains (ICLR 2021 · similarity 0.89)
- Batch Normalization Is Blind to the First and Second Derivatives of the Loss (AAAI 2024 · similarity 0.86)
- Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion (ICLR 2024 · similarity 0.86)
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.