Research map / Architectures and efficiency / RNNs and normalization
Layer-wise Conditioning Analysis in Exploring the Learning Dynamics of DNNs
Lei Huang, Jie Qin, Li Liu, Fan Zhu, Ling Shao
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 |
|---|---|---|---|
| add one BN layer before the last linear layer | ResNet 1202-layer · CIFAR-10 | test error: 9.42 ± 3.10 -> 5.68 ± 0.14 | Improved |
| add one BN layer before the last linear layer | ResNet 56-layer · CIFAR-10 | test error: 7.52 ± 0.30 -> 6.50 ± 0.22 | Improved |
| add one BN layer before the last linear layer | ResNet 110-layer · CIFAR-10 | test error: 6.89 ± 0.52 -> 6.10 ± 0.09 | Improved |
| add one BN layer before the last linear layer | ResNet 1202-layer · CIFAR-100 | test error: 30.49 ± 4.44 -> 25.51 ± 0.27 | Improved |
| add one BN layer before the last linear layer | ResNet 56-layer · CIFAR-100 | test error: 29.60 ± 0.41 -> 28.82 ± 0.38 | Improved |
| add the extra BN layer before final pooling (ablation) | PreResNet-style 1202-layer · CIFAR-10 | test error: 7.89 ± 0.58 -> 6.07 ± 0.10 | Improved |
| add the extra BN layer before final pooling (ablation) | PreResNet-style 1202-layer · CIFAR-100 | test error: 30.07 ± 2.04 -> 26.23 ± 0.26 | Improved |
| add the extra BN layer before final pooling (ablation) | PreResNet-style 56-layer · CIFAR-10 | test error: 6.75 ± 0.26 -> 6.89 ± 0.09 | No clear effect |
| add one BN layer before the last linear layer | ResNet 18-layer · ImageNet val | top-1 error: 29.78 -> 29.38 | Improved |
| add one BN layer before the last linear layer | ResNet 18-layer with WD=0.0002 · ImageNet val | top-1 error: 30.07 -> 28.96 | Improved |
| add one BN layer before the last linear layer | ResNet 50-layer · ImageNet val | top-1 error: 23.97 -> 23.76 | Improved |
| add one BN layer before the last linear layer | VGG-style without BN, 56-layer · CIFAR-10 | test error: 89.93 ± 0.12 -> 23.40 ± 1.74 | Improved |
Closest papers by meaning
- Batch Normalization Preconditioning for Neural Network Training (JMLR 2022 · similarity 0.88)
- Balancing Learning Rates Across Layers: Exact Two-Step Dynamics and Optimal Scaling in Linear Neural Networks (ICML 2026 · similarity 0.88)
- On The Concurrence of Layer-wise Preconditioning Methods and Provable Feature Learning (ICML 2025 · similarity 0.88)
- A Loss Curvature Perspective on Training Instabilities of Deep Learning Models (ICLR 2022 · similarity 0.88)
- Towards Understanding Regularization in Batch Normalization (ICLR 2019 · similarity 0.87)
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 2025 · 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.
