AI Paper Atlas

Research map / Optimization / Optimizers

Delving into Variance Transmission and Normalization: Shift of Average Gradient Makes the Network Collapse

Yuxiang Liu, Jidong Ge, Chuanyi Li, Jie Gui

AAAI 2021 · Optimization › Optimizers · approach: Memory-efficient LLM optimizers · 1 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 output normalization with PWS on convolution filtersSSD · VOC07 test, batch size 2mAP: 72.97 -> 76.93Improved
Replace output normalization with PWS on convolution filtersSSD · VOC07 test, batch size 8mAP: 76.74 -> 77.66Improved
Use PWS instead of BNMask R-CNN with ResNet-50 · COCO, 1x scheduleAP bbox: 38.0 -> 38.9Improved
Use PWS instead of BNSSD with VGG-16 · COCO, 2x scheduleAP bbox: 26.6 -> 27.2Improved
Replace BN with PWSResNet-50 · CIFAR-10 residual, LR 5e-2top-1 error: 9.14 -> 6.75Improved
Replace BN with PWSResNet-50 · ImageNet valtop-5 error: 23.6 -> 24.0Hurt
Add trainable alpha to PWS (ablation)PWS with VGG-16 · VOC07 test, LR 1e-2mAP: 74.96 -> 77.52Improved
Add the sqrt(2/n_l) variance scaling factor (ablation)PWS with VGG-16 · VOC07 test, LR 1e-2mAP: NaN -> 77.52Improved
Set gamma to 1e-3 instead of 1e-5 (ablation)PWS with VGG-16 · VOC07 test, LR 1e-3mAP: 72.99 -> 70.32Hurt
Set gamma to 1e-5 instead of 1e-3 (ablation)PWS with VGG-16 · VOC07 test, LR 1e-2mAP: 77.52 -> NaNHurt
Set gamma to 1e-4 instead of 1e-3 (ablation)PWS with VGG-16 · VOC07 test, LR 1e-2mAP: 77.52 -> 76.50Hurt

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