AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Active Token Mixer

Guoqiang Wei, Zhizheng Zhang, Cuiling Lan, Yan Lu, Zhibo Chen

AAAI 2023 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining · 21 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
Use ATMNet instead of scale-comparable backbonesImageNet-1K classification · ImageNet-1K val, 224²Top-1 accuracy (%): 82.0 vs 81.7Improved
Use ATMNet instead of Twins-LSemantic FPN · ADE20K valmIoU: 48.1 vs 46.7Improved
Use ATMNet-L instead of Twins-LUperNet · ADE20K val, multi-scalemIoU: 51.1 vs 50.2Improved
Use ATMNet-T instead of WaveMLP-SRetinaNet 1× · COCO val2017AP^b: 43.6 vs 43.4Improved
Use ATMNet-L instead of Twins-LMask R-CNN 1× · COCO val2017AP^b: 47.4 vs 45.9Improved
Add adaptive spatial token mixing rather than fixing all offsets to 0 (ablation)ATMNet · ImageNet-1K valTop-1 accuracy (%): 79.3 -> 82.0Improved
Add adaptive spatial token mixing rather than fixing all offsets to 0 (ablation)ATMNet · COCO, RetinaNet 1×AP^b: 36.0 -> 43.6Improved
Add adaptive spatial token mixing rather than fixing all offsets to 0 (ablation)ATMNet · ADE20K, Semantic FPNmIoU: 37.9 -> 45.8Improved
Add PEG for offset-generation position information (ablation)ATMNet · ImageNet-1K valTop-1 accuracy (%): 82.0 -> 82.0No clear effect
Add PEG for offset-generation position information (ablation)ATMNet · COCO, RetinaNet 1×AP^b: 43.4 -> 43.6Improved
Add PEG for offset-generation position information (ablation)ATMNet · ADE20K, Semantic FPNmIoU: 45.6 -> 45.8Improved
Replace FPN with ATMFPN (ablation)ResNet-50 · ADE20K, Semantic FPNmIoU: 37.3 -> 40.3 (+3.0)Improved

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