AI Paper Atlas

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

When Does Self-supervision Improve Few-shot Learning?

Jong-Chyi Su, Subhransu Maji, Bharath Hariharan

ECCV 2020 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Data-efficient and robust self-supervision · 165 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 jigsaw puzzle auxiliary loss (ablation)ProtoNet · 7 few-shot datasets (5-way 5-shot)accuracy: relative error reduced by 4.0%, 8.7%, 19.7%, 8.4%, 4.7%, 15.9%, 27.8%Improved
add rotation prediction auxiliary loss (ablation)ProtoNet · aircrafts and flowers (5-way 5-shot)accuracy: 91.4 -> 91.4 (aircrafts); 89.2 -> 89.0 (flowers)No clear effect
add jigsaw and rotation together (ablation)ProtoNet · 7 few-shot datasets (5-way 5-shot)accuracy: tiered-ImageNet 78.0 -> 77.2; Birds 89.8 -> 90.2Mixed
add jigsaw loss to other meta-learners (ablation)Softmax, MAML, ProtoNet · 5 fine-grained datasets avg (5-way 5-shot)accuracy: 85.5 -> 86.6; 82.6 -> 83.8; 88.5 -> 90.4Improved
add jigsaw loss on degraded images (Birds greyscale, Cars low-res) (ablation)ProtoNet · degraded datasets (5-way 5-shot)accuracy: 82.2 -> 85.4 (Birds greyscale); 84.8 -> 87.0 (Cars low-res)Improved
add jigsaw loss with only 20% of base images (ablation)ProtoNet · 20% Cars (5-way 5-shot)accuracy: 75.8 -> 82.8; gain 0.7% -> 7.0%Improved
add jigsaw loss with only 20% of base images (ablation)ProtoNet · 20% Aircrafts (5-way 5-shot)accuracy: 77.7 -> 78.4No clear effect
use self-supervision alone instead of supervised labels (ablation)ResNet-18 · fine-grained avg (5-way 5-shot)accuracy: 85.5 -> 32.9 (jigsaw); 85.5 -> 33.7 (rotation)Hurt
initialize with SSL before meta-learning instead of random init (ablation)ProtoNet · few-shot benchmarksaccuracyNo clear effect
replace fraction of SSL images with out-of-domain images (ablation)ProtoNet + Jigsaw · 20% base set, 5 datasetsaccuracyHurt
add more unlabeled in-domain images for SSL (20% to 100%) (ablation)ProtoNet + Jigsaw · 20% base setaccuracyImproved
add random pool images for SSL (ablation)ProtoNet + Jigsaw · 20% Cars/Dogs/Flowersaccuracy: 82.1 -> 78.4 (Cars); 71.8 -> 68.5 (Dogs); 85.9 -> 83.5 (Flowers)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