AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Adaptive Frequency Filters As Efficient Global Token Mixers

Zhipeng Huang, Zhizheng Zhang, Cuiling Lan, Zheng-Jun Zha, Yan Lu, Baining Guo

ICCV 2023 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining · 136 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
add AFF token mixers to a baseline with no spatial token mixer (ablation)AFFNet baseline · ImageNet-1Ktop-1 accuracy: 77.9 -> 79.8Improved
use AFF token mixer instead of 3x3 conv token mixer (ablation)AFFNet · ImageNet-1Ktop-1 accuracy: 78.6 -> 79.8Improved
apply adaptive filtering in the frequency domain instead of the original domain (ablation)AFF token mixer · ImageNet-1Ktop-1 accuracy: 78.4 -> 79.8Improved
use Hadamard product instead of summation in frequency filtering (ablation)AFF token mixer · ImageNet-1Ktop-1 accuracy: 78.8 -> 79.8Improved
use AFF design instead of AFNO frequency operation (ablation)AFFNet · ImageNet-1Ktop-1 accuracy: 78.8 -> 79.8Improved
use AFF design instead of FNO (trainable matrices) (ablation)AFFNet · ImageNet-1Ktop-1 accuracy: 79.7 -> 79.8, params 141.0M -> 5.5MImproved
use channel-specific masks instead of channel-shared masks (ablation)AFF token mixer · ImageNet-1Ktop-1 accuracy: 79.3 -> 79.8Improved
increase spatial kernel size in mask network to 3x3 (ablation)AFF mask network · ImageNet-1Ktop-1 accuracy: 79.8 -> 79.8, params 5.5M -> 7.9MNo clear effect
use regular linear layers (N_group=1) in mask network (ablation)AFF mask network · ImageNet-1Ktop-1 accuracy: 79.8 -> 79.9, params 5.5M -> 7.7MNo clear effect
use depthwise setting N_group=C in mask network (ablation)AFF mask network · ImageNet-1Ktop-1 accuracy: 79.8 -> 79.4Hurt
use AFFNet backbone instead of EdgeNext in SSDEdgeNext · COCO detectionmAP: 27.9 -> 28.4Improved
use AFFNet backbone instead of EdgeNext in DeepLabv3EdgeNext · PASCAL VOCmIoU: 80.2 -> 80.5Improved

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