Research map / Architectures and efficiency / RNNs and normalization
Understanding the Failure of Batch Normalization for Transformers in NLP
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 RBN mean/variance discrepancy regularization instead of BN | Post-Norm Transformer · IWSLT14 | BLEU: 34.0 -> 35.5 | Helped |
| Add RBN mean/variance discrepancy regularization instead of BN | Post-Norm Transformer · WMT16 | BLEU: 25.0 -> 26.5 | Helped |
| Add RBN mean/variance discrepancy regularization instead of BN | Post-Norm Transformer · PTB | perplexity: 45.9 -> 44.6 | Helped |
| Add RBN mean/variance discrepancy regularization instead of BN | Post-Norm Transformer · WikiText-103 | perplexity: 17.2 -> 17.1 | Helped |
| Add RBN mean/variance discrepancy regularization instead of BN | Post-Norm Transformer · CoNLL2003 | F1: 90.9 -> 91.4 | Helped |
| Add RBN mean/variance discrepancy regularization instead of BN | Post-Norm Transformer · IMDB | accuracy: 84.0 -> 84.5 | Helped |
| Replace LN with BN | Post-Norm Transformer · PTB | perplexity: 53.2 -> 45.9 | Helped |
| Replace LN with BN | Post-Norm Transformer · WMT16 | BLEU: 27.3 -> 25.0 | Hurt |
| Add mean-discrepancy penalty and variance-discrepancy penalty, each tested separately (ablation) | RBN in Pre-Norm Transformer on IWSLT14 · IWSLT14 | BLEU | Helped |
| Combine mean and variance penalties with moderate coefficients (ablation) | RBN in Pre-Norm Transformer on IWSLT14 · IWSLT14 | BLEU | Helped |
| Use batch statistics instead of EMA population statistics at test time (ablation) | Post-BN · IWSLT14 | BLEU: 34.0 -> 34.8 | Helped |
| Use batch statistics instead of EMA population statistics at test time (ablation) | Pre-BN · IWSLT14 | BLEU: 34.8 -> 27.2 | Hurt |
Closest papers by meaning
- PowerNorm: Rethinking Batch Normalization in Transformers (ICML 2020 · similarity 0.95)
- Four Things Everyone Should Know to Improve Batch Normalization (ICLR 2020 · similarity 0.91)
- Normalizing the Normalizers: Comparing and Extending Network Normalization Schemes (ICLR 2017 · similarity 0.90)
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.90)
- Reducing Divergence in Batch Normalization for Domain Adaptation (AAAI 2025 · similarity 0.89)
- The Curse of Depth in Large Language Models (NeurIPS 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.