Research map / Architectures and efficiency / RNNs and normalization
Norm matters: efficient and accurate normalization schemes 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 |
|---|---|---|---|
| Replace L² batch norm with calibrated L¹ batch norm | ResNet18 and ResNet50 · ImageNet | accuracy: 69.8% -> 69.74%; 75.3% -> 75.32% | No clear effect |
| Replace L² layer norm with calibrated L¹ layer norm | Transformer · WMT14 | perplexity: 5.1 ppl -> 5.2 ppl | Hurt |
| Replace L² batch norm with L∞ batch norm (ablation) | ResNet18 and ResNet50 · ImageNet | accuracy: lower accuracy, by 3% | Hurt |
| Use Top(10) instead of maximum absolute deviation for L∞ normalization (ablation) | ResNet18 and ResNet50 · ImageNet | accuracy: difference < 0.2% | Helped |
| Run L² batch norm in half precision rather than L¹ batch norm (ablation) | ResNet18 · ImageNet | validation accuracy | Hurt |
| Remove the L¹ calibration constant C_L1 (ablation) | L¹ batch norm · CIFAR-10 | validation error: 20% | Hurt |
| Use bounded weight-norm instead of unbounded weight-norm (ablation) | ResNet50 · ImageNet | accuracy: 67% -> 73.8% | Helped |
| Use bounded weight-norm instead of unbounded weight-norm (ablation) | ResNet56 · CIFAR-10 | accuracy: 92.5% -> 92.88% | Helped |
| Replace layer norm with bounded weight-norm (ablation) | Transformer · WMT14 | BLEU: 27.3 BLEU -> 26.5 BLEU | Hurt |
| Use bounded weight-norm instead of layer norm (ablation) | 2-layer LSTM · WMT14 de-en | BLEU: 21.5 BLEU -> 21.2 BLEU | Hurt |
| Replace weight-decay training with the paper's learning-rate correction (ablation) | VGG11 with batch norm · CIFAR-10 | test accuracy | No clear effect |
| Replace learning-rate scheduling with norm scheduling (ablation) | VGG11 with batch norm · CIFAR-10 | test accuracy | No clear effect |
Closest papers by meaning
- Four Things Everyone Should Know to Improve Batch Normalization (ICLR 2020 · similarity 0.92)
- Demystifying Batch Normalization in ReLU Networks: Equivalent Convex Optimization Models and Implicit Regularization (ICLR 2022 · similarity 0.92)
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.91)
- Weight Normalization: A Simple Reparameterization to Accelerate Training of Deep Neural Networks (NeurIPS 2016 · similarity 0.91)
- Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion (ICLR 2024 · similarity 0.89)
- Why Do We Need Weight Decay in Modern Deep Learning? (NeurIPS 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.