Research map / Architectures and efficiency / Attention and transformers
On Separate Normalization in Self-supervised Transformers
Xiaohui Chen, Yinkai Wang, Yuanqi Du, Soha Hassoun, Liping Liu
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 |
|---|---|---|---|
| replace ShareNorm with SepNorm | MAE ViT-base · STL10 | top-1 accuracy: 92.01 -> 93.84 | Improved |
| replace ShareNorm with SepNorm | MAE ViT-base · Aircraft | top-1 accuracy: 52.54 -> 59.02 | Improved |
| replace ShareNorm with SepNorm | MAE ViT-base · SVHN | top-1 accuracy: 88.97 -> 89.18 | No clear effect |
| replace ShareNorm with SepNorm | MAE ViT-base · Flower | top-1 accuracy: 27.63 -> 32.51 | Improved |
| replace ShareNorm with SepNorm | BERT-base · STS unsupervised | average Spearman: 74.04 -> 75.87 | Improved |
| replace ShareNorm with SepNorm | RoBERTa-base · STS supervised | average Spearman: 80.70 -> 81.90 | Improved |
| replace ShareNorm with SepNorm | BERT-base · STS supervised | average Spearman: 81.00 -> 81.23 | No clear effect |
| replace ShareNorm with SepNorm | BERT-base · transfer tasks | average accuracy: 86.36 -> 86.14 | No clear effect |
| replace ShareNorm with SepNorm | RoBERTa-base · transfer tasks | average accuracy: 87.61 -> 88.37 | Improved |
| replace ShareNorm with SepNorm | Graphormer · ZINC | mean absolute error: 0.069 -> 0.052 | Improved |
| replace ShareNorm with SepNorm | Graphormer · MolHIV | AUC: 73.36% -> 75.64% | Improved |
| add uniformity loss on CLS and token embeddings with lambda=1 to SepNorm (BN+LN) (ablation) | MAE ViT-base · STL10 | accuracy: 92.80 -> 94.40 | Improved |
Closest papers by meaning
- HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization (NeurIPS 2025 · similarity 0.87)
- Transformers without Normalization (CVPR 2025 · similarity 0.87)
- Residual Stream Analysis with Multi-Layer SAEs (ICLR 2025 · similarity 0.86)
- Non-asymptotic Convergence of Training Transformers for Next-token Prediction (NeurIPS 2024 · similarity 0.86)
- Transformers are Universal In-context Learners (ICLR 2025 · similarity 0.86)
- What Improves the Generalization of Graph Transformers? A Theoretical Dive into the Self-attention and Positional Encoding (ICML 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.
