AI Paper Atlas

Research map / Label-efficient and robust learning / Self-supervised and contrastive

How Mask Matters: Towards Theoretical Understandings of Masked Autoencoders

Qi Zhang, Yifei Wang, Yisen Wang

NeurIPS 2022 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Masked autoencoder pretraining

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
add uniformity regularizer (U-MAE)MAE ViT-Tiny · CIFAR-10 linear probingaccuracy (%): 59.6 -> 68.9Improved
add uniformity regularizer (U-MAE)MAE ViT-Base · ImageNet-100 linear probingaccuracy (%): 61.2 -> 67.5Improved
add uniformity regularizer (U-MAE)MAE ViT-Large · ImageNet-100 linear probingaccuracy (%): 64.4 -> 72.8Improved
add uniformity regularizer (U-MAE)MAE ViT-Base · ImageNet-1K linear probingaccuracy (%): 55.4 -> 58.5Improved
add uniformity regularizer (U-MAE)MAE ViT-Large · ImageNet-1K linear probingaccuracy (%): 62.2 -> 65.8Improved
add uniformity regularizer (U-MAE)MAE · CIFAR-10 fine-tuning ViT-Tinyaccuracy (%): 89.6 -> 89.4No clear effect
add uniformity regularizer (U-MAE)MAE ViT-Base · ImageNet-1K fine-tuningaccuracy (%): 82.9 -> 83.0No clear effect
add uniformity regularizer to SimMIM (U-SimMIM)SimMIM ViT-Base · ImageNet-100 linear probingaccuracy (%): 54.3 -> 61.1Improved
set uniformity coefficient to 1e-3 (ablation)U-MAE · ImageNet-100 linear probingaccuracy (%): 61.2 -> 65.9Improved
set uniformity coefficient to 1e-2 (ablation)U-MAE · ImageNet-100 linear probingaccuracy (%): 61.2 -> 67.5Improved
set uniformity coefficient to 1e-1 (too large) (ablation)U-MAE · ImageNet-100 linear probingaccuracy (%): 67.5 -> 45.1Hurt
set uniformity coefficient to 1e0 (ablation)U-MAE · ImageNet-100 linear probingaccuracy (%): 67.5 -> 47.0Hurt

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