AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

What to Hide from Your Students: Attention-Guided Masked Image Modeling

Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, Nikos Komodakis

ECCV 2022 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining · 137 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 random block-wise masking with attention-guided masking of most attended tokens (AttMask-High) (ablation)iBOT ViT-S, 20% ImageNet · ImageNet-1k k-NNtop-1 accuracy: 46.7 -> 49.7Improved
replace random block-wise masking with AttMask-High (ablation)iBOT ViT-S, 20% ImageNet · ImageNet-1k linear probingtop-1 accuracy: 56.4 -> 57.9Improved
mask least attended tokens (AttMask-Low) instead of random block-wise (ablation)iBOT ViT-S, 20% ImageNet · ImageNet-1k k-NNtop-1 accuracy: 46.7 -> 44.0Hurt
use aggressive 75% random masking instead of block-wise 10-50% (ablation)iBOT ViT-S, 20% ImageNet · ImageNet-1k linear probingtop-1 accuracy: 56.4 -> 55.5Hurt
use random patch masking (10-50%) instead of block-wise (ablation)iBOT ViT-S, 20% ImageNet · ImageNet-1k k-NNtop-1 accuracy: 46.7 -> 47.8Improved
reveal a few most attended tokens as hints (AttMask-Hint) rather than hiding all (AttMask-High) (ablation)iBOT ViT-S, 20% ImageNet · ImageNet-1k k-NNtop-1 accuracy: 49.7 -> 49.5No clear effect
add AttMask-High masking to DINO (no MIM loss) (ablation)DINO ViT-S, 20% ImageNet · ImageNet-1k k-NNtop-1 accuracy: 43.0 -> 43.5Improved
use AttMask-Low masking in DINO (ablation)DINO ViT-S, 20% ImageNet · ImageNet-1k k-NNtop-1 accuracy: 43.0 -> 42.7Hurt
use AttMask-High with less training data (5% of ImageNet) (ablation)iBOT ViT-S · ImageNet-1k k-NNtop-1 accuracy: 15.7 -> 17.5Improved
train on 100% ImageNet for 300 epochs with AttMask (ablation)iBOT ViT-S · ImageNet-1k k-NNtop-1 accuracy: 74.6 -> 75.0Improved
add AttMask-High to iBOT with few examples per class (nu=1) (ablation)iBOT ViT-S, 100 ep · ImageNet-1k k-NN few-shottop-1 accuracy: 32.9 -> 37.1Improved
use masking probability 0.5 rather than 0 (no masking) for attention-guided masking (ablation)iBOT+AttMask · ImageNet-1k k-NNtop-1 accuracy: 43.4 -> 49.4Improved

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