Research map / Label-efficient and robust learning / Self-supervised and contrastive
Do SSL Models Have Déjà Vu? A Case of Unintended Memorization in Self-supervised Learning
Casey Meehan, Florian Bordes, Pascal Vincent, Kamalika Chaudhuri, Chuan Guo
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 changed | Where | Measured | Effect |
|---|---|---|---|
| increase VICReg training epochs from 250 to 1000 (ablation) | VICReg · ImageNet 300K train set | deja vu score: from under 10% to over 30% | Hurt |
| increase VICReg training epochs beyond 250 (ablation) | VICReg · ImageNet 300K train set | linear probe train-test gap | No clear effect |
| increase training set size from 100K to 500K (ablation) | VICReg · ImageNet, 1000 epochs | deja vu score | No clear effect |
| increase training set size (ablation) | VICReg · ImageNet, 1000 epochs | linear probe train-test gap | Improved |
| increase VICReg invariance coefficient lambda above 25 (ablation) | VICReg · ImageNet 100K train set | deja vu score | Improved |
| vary SimCLR temperature (ablation) | SimCLR · ImageNet | deja vu score | Mixed |
| switch SSL criterion from VICReg or Barlow Twins to SimCLR or BYOL (ablation) | SSL training criterion · ImageNet | deja vu score | Improved |
| use SSL instead of supervised training (ablation) | ResNet101 on ImageNet · ImageNet | deja vu score | Hurt |
| use larger model (ResNet50/101 vs ResNet18/34; larger ViT) (ablation) | VICReg · ImageNet | deja vu score | Hurt |
| use backbone embedding instead of projector embedding (guillotine regularization) (ablation) | VICReg · ImageNet | deja vu score | Improved |
| fine-tune pretrained VICReg for the first 5 epochs (ablation) | VICReg pretrained 1000 epochs · ImageNet set A classification fine-tuning | deja vu score | Improved |
| continue fine-tuning to 20 epochs (ablation) | VICReg pretrained 1000 epochs · ImageNet set A classification fine-tuning | deja vu score: 5 epochs, 20 epochs | Hurt |
Closest papers by meaning
- Déjà Vu Memorization in Vision–Language Models (NeurIPS 2024 · similarity 0.91)
- Memorization in Self-Supervised Learning Improves Downstream Generalization (ICLR 2024 · similarity 0.90)
- Localizing Memorization in SSL Vision Encoders (NeurIPS 2024 · similarity 0.89)
- Measuring Dejavu Memorization Efficiently (NeurIPS 2024 · similarity 0.88)
- Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation (ICLR 2024 · similarity 0.87)
- Captured by Captions: On Memorization and its Mitigation in CLIP Models (ICLR 2025 · similarity 0.86)
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.
