AI Paper Atlas

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

A Broad Study on the Transferability of Visual Representations With Contrastive Learning

Ashraful Islam, Chun-Fu (Richard) Chen, Rameswar Panda, Leonid Karlinsky, Richard Radke, Rogerio Feris

ICCV 2021 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Inductive bias and minimal representations · 83 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
add supervised contrastive + self-supervised contrastive loss instead of cross-entropyResNet-50 pretraining · 12 downstream datasets, linear evaluationtop-1 accuracy: +5.63% over CEImproved
add supervised contrastive + self-supervised contrastive loss instead of cross-entropyResNet-18 pretraining · 12 downstream datasets, 5-shottop-1 accuracy: 60.01 -> 63.47Improved
add strong MoCo-style augmentation to CE (ablation)CE · 12 downstream datasets, linear evaluationtop-1 accuracy: 75.67 -> 75.91No clear effect
add strong MoCo-style augmentation to CE (ablation)CE · 12 downstream datasets, full fine-tuningtop-1 accuracy: 88.13 -> 88.27No clear effect
add contrastive pretraining for full-network fine-tuning on full dataSupCon+SelfSupCon · 12 downstream datasets, full fine-tuningtop-1 accuracy: 88.13 -> 88.56No clear effect
add contrastive pretraining for fine-tuning with 1000 samplesCE+SelfSupCon, SupCon+SelfSupCon · 12 downstream datasets, 1000 training samplestop-1 accuracy: 69.79 -> 70.88, 71.27Improved
add self-supervised contrastive loss to SupConSupCon · MS COCO detectionAP: ~0.5% APImproved
add self-supervised contrastive loss to CECE · VOC0712 detectionAP50: 81.58 -> 83.06Improved
add self-supervised contrastive loss to CECE · ImageNet-Atop-1 accuracy: +5.18%Improved
add self-supervised contrastive loss to CECE · ImageNet-Rtop-1 accuracy: +4.38%Improved
use SelfSupCon alone instead of CECE · ImageNet valtop-1 accuracy: 76.60 -> 66.07Hurt
increase weight alpha of self-supervised loss from 1 to 2 in CE+SelfSupCon (ablation)CE+SelfSupCon · transfer accuracy (12 datasets) vs ImageNettop-1 accuracy: best ImageNet at alpha=1, best transfer at alpha=2Mixed

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