AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Learning Versatile Convolution Filters for Efficient Visual Recognition

Kai Han, Yunhe Wang, Chang Xu, Chunjing Xu, Enhua Wu, Dacheng Tao

TPAMI 2021 · Architectures and efficiency › CNN design · approach: Filter design and trained-filter analysis · 12 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 separate instead of shared learnable masks (ablation)L-Versatile (s=2) ResNet-56 · CIFAR-10accuracy: 92.8% -> 93.4%Improved
use separate instead of shared learnable masks (ablation)L-Versatile (s=4) ResNet-50 · ImageNettop-1 error: 26.8% -> 25.5%Improved
use separate instead of shared learnable masks (ablation)L-Versatile (s=4) AlexNet · ImageNettop-1 error: 43.7% -> 43.1%Improved
set channel overlap and stride to 8 instead of 1 (ablation)C-Versatile ResNet-56 · CIFAR-10accuracy: 92.4% -> 93.4%Improved
increase channel overlap and stride from 8 to 12 (ablation)C-Versatile ResNet-56 · CIFAR-10accuracy: 93.4% -> 92.4%Hurt
reduce filters further with g=2 instead of g=8 at c-c_hat=8 (ablation)C-Versatile ResNet-56 · CIFAR-10accuracy: 93.4% -> 91.2%Hurt
add channel versatile filters to spatial versatile filtersS-Versatile ResNet-56 · CIFAR-10accuracy: 93.2% -> 92.1%Hurt
increase s from 2 to 4 in separate L-Versatile (ablation)ResNet-56 · CIFAR-10accuracy: 93.4% -> 92.8%Hurt
increase s to 6 in separate L-Versatile (ablation)ResNet-56 · CIFAR-10accuracy: 93.4% -> 92.4%Hurt
use spatial versatile filters instead of vanilla filtersAlexNet · ImageNettop-1 error: 42.9% -> 42.1%Improved
use spatial versatile filters instead of vanilla filtersResNet-50 · ImageNettop-1 error: 24.7% -> 24.5%Improved
increase s from 4 to 32 in separate L-Versatile (ablation)ResNet-50 · ImageNettop-5 error: 8.0% -> 8.7%Hurt

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