AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

GFNet: Global Filter Networks for Visual Recognition

Yongming Rao, Wenliang Zhao, Zheng Dong Zhu, Jie Zhou, Jiwen Lu

TPAMI 2023 · Architectures and efficiency › Attention and transformers · approach: Efficient and robust vision transformers · 579 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 self-attention with global filter layer (GFNet-Ti vs DeiT-Ti)DeiT-Ti · ImageNet-1k valtop-1 accuracy (%): 72.2 -> 74.6Improved
replace spatial MLP with global filter (GFNet-Ti vs gMLP-Ti)gMLP-Ti · ImageNet-1k valtop-1 accuracy (%): 72.0 -> 74.6Improved
use global filter instead of ResMLP token mixingResMLP-12 · ImageNet-1k valtop-1 accuracy (%): 76.6 -> 78.6Improved
use global filter instead of self-attention at DeiT-S scaleDeiT-S · ImageNet-1k valtop-1 accuracy (%): 79.8 -> 80.0No clear effect
finetune at 384 resolution for 30 epochsGFNet-B · ImageNet-1k valtop-1 accuracy (%): 80.7 -> 82.1Improved
finetune at 384 resolution for 30 epochsGFNet-XS · ImageNet-1k valtop-1 accuracy (%): 78.6 -> 80.6Improved
use hierarchical GFNet-H instead of ResNet-50-scale CNNResNet-50 · ImageNet-1k valtop-1 accuracy (%): 76.1 -> 81.5Improved
use GFNet-H-B instead of Swin-SSwin-S · ImageNet-1k valtop-1 accuracy (%): 83.0 -> 82.9No clear effect
use GFNet-H-Ti backbone instead of PVT-TiPVT-Ti · ADE20KmIoU: 35.7 -> 41.0Improved
use GFNet-H-B backbone instead of Swin-SSwin-S · ADE20KmIoU: 45.2 -> 44.8Hurt
use GFNet-S instead of DeiT-S for robustnessDeiT-S · ImageNet FGSM/PGDaccuracy (%): 40.7 -> 42.6; 16.7 -> 21.0Improved
use GFNet-S instead of DeiT-S on ImageNet-ADeiT-S · ImageNet-Atop-1 accuracy (%): 18.9 -> 14.3Hurt

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