AI Paper Atlas

Research map / Label-efficient and robust learning / Self-supervised and contrastive

Unsupervised Visual Representation Learning by Synchronous Momentum Grouping

Bo Pang, Yifan Zhang, Yaoyi Li, Jia Cai, Cewu Lu

ECCV 2022 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Empirical contrastive visual learning · 25 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 linearly decreasing beta schedule 1.0->0.99 instead of fixed beta 0.99 (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 65.9 -> 67.2Improved
vary final beta of schedule (0.9, 0.99, 0.999) (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 67.0 / 67.2 / 67.1No clear effect
add periodical clustering to avoid collapse (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 0.1 -> 53.7Improved
add periodical reset of the momentum network on top of periodical clustering (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 53.7 -> 67.2Improved
reset momentum network periodically without clustering (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 0.1 -> 53.4Improved
increase number of groups from 300 to 3000 (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 65.2 -> 67.2Improved
increase number of groups from 3000 to 10000 (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 67.2 -> 67.2No clear effect
momentum update of group features instead of averaging update (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 65.8 -> 67.2Improved
momentum update instead of adopting latest instance features as group features (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 42.1 -> 67.2Improved
momentum update instead of randomly selected group features (ablation)SMoG · ImageNet linear (100 ep)top-1 accuracy: 30.2 -> 67.2Improved
add multi-crop trainingSMoG ResNet-50 · ImageNet linear evaltop-1 accuracy: 73.6 -> 76.4Improved
add multi-crop trainingSMoG Swin-T · ImageNet linear evaltop-1 accuracy: 74.5 -> 77.7Improved

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