AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Robustifying Token Attention for Vision Transformers

Yong Guo, David Stutz, Bernt Schiele

ICCV 2023 · Architectures and efficiency › Attention and transformers · approach: Efficient and robust vision transformers · 38 citations (OpenAlex)

Read the paper

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 TAP and ADL together to FAN-B-HybridFAN-B-Hybrid · ImageNet-CmCE: 46.1 -> 43.7Improved
use dilated average pooling in TAP instead of multi-kernel pooling (ablation)TAP on FAN-B-Hybrid · ImageNet-CmCE: 45.5 -> 44.9Improved
replace adaptive TAP with fixed AvgPool5x5 (ablation)FAN-B-Hybrid · ImageNettop-1 accuracy: 83.9 -> 82.7Hurt
add TAP in every block rather than only shallow or middle+deep layers (ablation)FAN-B-Hybrid · ImageNet-CmCE: 45.7 -> 44.9Improved
increase ADL weight lambda to 10 (ablation)ADL · ImageNet-CmCEHurt
set ADL threshold tau=2 rather than 5 (ablation)ADL on FAN-B-Hybrid · ImageNet-CmCE: 45.8 -> 44.4Improved
increase TAP branches K beyond 4 (ablation)TAP · ImageNet-CmCENo clear effect
add ADL alone to FAN-B-Hybrid (ablation)FAN-B-Hybrid · ImageNettop-1 accuracy: 83.9 -> 84.0No clear effect
add TAP and ADL to Swin-BSwin-B · ImageNet-CmCE: 54.4 -> 51.9Improved
add TAP and ADL to FAN-B-Hybrid for segmentationFAN-B-Hybrid · ACDC testmIoU: 60.6 -> 63.6Improved
add ADL to reduce attention similarity between headsFAN-B-Hybrid · ImageNet head similaritycosine similarity between heads: 0.63 -> 0.27Improved

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