AI Paper Atlas

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

NeurIPS 2023 · Architectures and efficiency › Attention and transformers · approach: Transformer block and depth design · 1 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace ShareNorm with SepNormMAE ViT-base · STL10top-1 accuracy: 92.01 -> 93.84Improved
replace ShareNorm with SepNormMAE ViT-base · Aircrafttop-1 accuracy: 52.54 -> 59.02Improved
replace ShareNorm with SepNormMAE ViT-base · SVHNtop-1 accuracy: 88.97 -> 89.18No clear effect
replace ShareNorm with SepNormMAE ViT-base · Flowertop-1 accuracy: 27.63 -> 32.51Improved
replace ShareNorm with SepNormBERT-base · STS unsupervisedaverage Spearman: 74.04 -> 75.87Improved
replace ShareNorm with SepNormRoBERTa-base · STS supervisedaverage Spearman: 80.70 -> 81.90Improved
replace ShareNorm with SepNormBERT-base · STS supervisedaverage Spearman: 81.00 -> 81.23No clear effect
replace ShareNorm with SepNormBERT-base · transfer tasksaverage accuracy: 86.36 -> 86.14No clear effect
replace ShareNorm with SepNormRoBERTa-base · transfer tasksaverage accuracy: 87.61 -> 88.37Improved
replace ShareNorm with SepNormGraphormer · ZINCmean absolute error: 0.069 -> 0.052Improved
replace ShareNorm with SepNormGraphormer · MolHIVAUC: 73.36% -> 75.64%Improved
add uniformity loss on CLS and token embeddings with lambda=1 to SepNorm (BN+LN) (ablation)MAE ViT-base · STL10accuracy: 92.80 -> 94.40Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy