Research map / Architectures and efficiency / RNNs and normalization
Batch Normalization Biases Residual Blocks Towards the Identity Function in Deep Networks
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 |
|---|---|---|---|
| Add a learnable residual-branch scalar initialized to 0 | Unnormalized n-2 Wide-ResNet · CIFAR-10, depth 1000 | trainability | Helped |
| Replace batch normalization with SkipInit α=0 | n-2 Wide-ResNet · CIFAR-10, depth 1000 | test accuracy: 94.6 ± 0.1 -> 94.3 ± 0.2 | Hurt |
| Use SkipInit α=0 instead of batch normalization | n-2 Wide-ResNet · CIFAR-10, depth 1000 | trainability | Helped |
| Initialize the SkipInit scalar to 1 rather than a small value (ablation) | n-2 Wide-ResNet · CIFAR-10, depth 100 and 1000 | trainability | Hurt |
| Scale the merged residual block by 1/√2 instead of downscaling the residual branch (ablation) | Unnormalized n-2 Wide-ResNet · CIFAR-10, depth 1000 | trainability | Hurt |
| Remove L2 regularization from SkipInit α=0 (ablation) | n-2 Wide-ResNet · CIFAR-10, depth 1000 | trainability | No clear effect |
| Add dropout and biases to SkipInit (ablation) | ResNet50-V2 · ImageNet validation, batch size 256 | top-1 accuracy: 74.9 -> 75.6 | Helped |
| Add dropout to Fixup (ablation) | ResNet50-V2 · ImageNet validation, batch size 256 | top-1 accuracy: 74.8 -> 75.8 | Helped |
| Use SkipInit instead of batch normalization | ResNet50-V2 · ImageNet validation, batch size 4096 | top-1 accuracy: 74.9 -> 70.8 | Hurt |
| Use batch normalization rather than SkipInit | 16-4 Wide-ResNet · CIFAR-10, small batch sizes | optimal learning rate | No clear effect |
| Increase batch size with batch normalization | 16-4 Wide-ResNet · CIFAR-10, large batch sizes | maximum stable learning rate | Helped |
Closest papers by meaning
- Four Things Everyone Should Know to Improve Batch Normalization (ICLR 2020 · similarity 0.91)
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.91)
- Norm matters: efficient and accurate normalization schemes in deep networks (NeurIPS 2018 · similarity 0.90)
- Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion (ICLR 2024 · similarity 0.90)
- Reducing Divergence in Batch Normalization for Domain Adaptation (AAAI 2025 · similarity 0.89)
- Batch Normalization Is Blind to the First and Second Derivatives of the Loss (AAAI 2024 · similarity 0.87)
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.