AI Paper Atlas

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

Concept Generalization in Visual Representation Learning

Mert Bulent Sariyildiz, Yannis Kalantidis, Diane Larlus, Karteek Alahari

ICCV 2021 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Invariance-based self-supervised learning · 38 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 self-supervised pretraining (DINO, SwAV, BYOL, OBoW, SimCLR-v2) instead of supervised trainingResNet50 · ImageNet-CoG L1-L5 (linear, all data)top-1 accuracyImproved
replace ResNet50 by ResNet50-sized visual transformers (DeiT-S, T2T-ViT-t-14)ResNet50 · ImageNet-CoG L5 (linear, all data)top-1 accuracyHurt
use transformer-based models in few-shot linear classificationResNet50 · ImageNet-CoG L5 (few-shot, N<=128)top-1 accuracy: consistent 3-4% accuracy gainImproved
pretrain on additional noisy web data (Semi-Sup, Semi-Weakly-Sup, CLIP)ResNet50 · ImageNet-CoG L1-L5 (linear, all data)top-1 accuracyImproved
add distillation (MEAL-v2, DeiT-S-distilled)ResNet50 / DeiT-S · ImageNet-CoG L1-L5 (linear, all data)top-1 accuracyImproved
add adversarial training (Adv-Robust)ResNet50 · ImageNet-CoG L1-L5 (linear, all data)top-1 accuracyNo clear effect
use NAS-designed backbones (EfficientNet-B1/B4, NAT-M4)ResNet50 · ImageNet-CoG L1-L5 (linear, all data)top-1 accuracyImproved
add label-associated augmentation (MixUp, Manifold-MixUp, CutMix, ReLabel)ResNet50 · ImageNet-CoG L1-L5 (linear, all data)top-1 accuracyHurt
build levels with word2vec text similarity instead of WordNet Lin similarity (ablation)ImageNet-CoG levels · ImageNet-CoG L3-L5 (ResNet50 linear)top-1 accuracyImproved
class-balance levels by subsampling to 732 training images per concept (ablation)ImageNet-CoG · ImageNet-CoG (subset of models)top-1 accuracyNo clear effect

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