AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Dynamic Spatial Sparsification for Efficient Vision Transformers and Convolutional Neural Networks

Yongming Rao, Zuyan Liu, Wenliang Zhao, Jie Zhou, Jiwen Lu

TPAMI 2023 · Architectures and efficiency › Attention and transformers · approach: Efficient and robust vision transformers · 80 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
use dynamic (input-dependent) token sparsification instead of structural downsampling (ablation)DynamicViT (DeiT-S) · ImageNet valtop-1 accuracy (%): 78.2 -> 79.3Improved
use dynamic rather than static sparsification (ablation)DynamicViT (DeiT-S) · ImageNet valtop-1 accuracy (%): 73.4 -> 79.3Improved
use learned prediction module instead of random token removal (ablation)DynamicViT (DeiT-S) · ImageNet valtop-1 accuracy (%): 77.5 -> 79.3Improved
use learned prediction module instead of class-attention-based removal (ablation)DynamicViT (DeiT-S) · ImageNet valtop-1 accuracy (%): 78.1 -> 79.3Improved
use three progressive stages instead of a single stage (ablation)DynamicViT (DeiT-S) · ImageNet valtop-1 accuracy (%): 77.4 -> 79.3Improved
use three stages instead of two (ablation)DynamicViT (DeiT-S) · ImageNet valtop-1 accuracy (%): 79.2 -> 79.3Improved
add distillation and KL losses (ablation)DynamicViT-S/0.7 (LV-ViT-S) · ImageNet valtop-1 accuracy (%): 82.5 -> 83.0Improved
add distillation and KL losses (ablation)DynamicViT-S/0.7 (DeiT-S) · ImageNet valtop-1 accuracy (%): 79.2 -> 79.3Improved
use linear projection fast path instead of zero mask (ablation)DynamicCNN-S/0.9 · ImageNet valtop-1 accuracy (%): 53.2 -> 83.1Improved
use linear projection instead of learnable mask (ablation)DynamicCNN-S/0.9 · ImageNet valtop-1 accuracy (%): 55.5 -> 83.1Improved
use linear projection instead of bottleneck MLP fast path (ablation)DynamicCNN-S/0.9 · ImageNet valtop-1 accuracy (%): 82.9 -> 83.1Improved
sparsify only stage 3 instead of also stage 4 (ablation)DynamicCNN-T · ImageNet valtop-1 accuracy (%): 79.7 -> 81.6Improved

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