AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Improving Generalization of Batch Whitening by Convolutional Unit Optimization

Yooshin Cho, Hanbyel Cho, Youngsoo Kim, Junmo Kim

ICCV 2021 · Architectures and efficiency › CNN design · approach: Large and dynamic kernel design · 4 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 proposed Convolutional Unit (whitening before convolution) without linear transform (ablation)IterNorm · CIFAR-10 (ResNet20)top-1 test error (%): 9.01 -> 8.29Improved
use proposed Convolutional Unit (whitening before convolution) without linear transform (ablation)IterNorm · CIFAR-100 (ResNet56)top-1 test error (%): 28.35 -> 27.3Improved
add gamma scaling to the proposed unit (ablation)IterNorm · CIFAR-100 (ResNet56)top-1 test error (%): 27.3 -> 27.64Hurt
use proposed unit with BN keeping gamma and beta (ablation)BN · CIFAR-10 (ResNet20)top-1 test error (%): 8.18 -> 8.43Hurt
replace original unit with proposed unit plus Grouped ShiftIterNorm · CIFAR-10 (ShiftResNet56)top-1 test error (%): 7.03 -> 6.66Improved
replace original unit with proposed unit plus Grouped ShiftIterNorm · CIFAR-100 (ShiftResNet56)top-1 test error (%): 28.38 -> 27.20Improved
replace original unit with proposed unit plus Grouped ShiftDBN · CIFAR-100 (ShiftResNet56)top-1 test error (%): 29.87 -> 27.85Improved
replace original unit with proposed unit at 10x larger learning rate (lr 1.0)IterNorm · CIFAR-10 (ShiftResNet56)top-1 test error (%): 8.28 -> 6.04Improved
replace original unit with proposed unit at 10x larger learning rate (lr 1.0)IterNorm · CIFAR-100 (ShiftResNet56)top-1 test error (%): 30.63 -> 25.51Improved
replace original unit with proposed unitIterNorm · ImageNet (ShiftNet-A)top-1 test error (%): 28.25 -> 27.37Improved
replace original unit with proposed unitIterNorm · ImageNet (ShiftNet-A-1.5)top-1 test error (%): 23.87 -> 23.15Improved
replace BN with IterNorm using proposed unitShiftNet-A · CUB-200-2011 transfertop-1 test error (%): 17.64 -> 14.10Improved

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