Research map / Label-efficient and robust learning / Self-supervised and contrastive
The hidden uniform cluster prior in self-supervised learning
Mido Assran, Randall Balestriero, Quentin Duval, Florian Bordes, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Nicolas Ballas
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 |
|---|---|---|---|
| class-imbalanced mini-batch sampling instead of class-balanced sampling during pretraining (ablation) | SimCLR, MSN, VICReg (volume-maximization SSL) · ImageNet-1K and downstream transfer | downstream accuracy drop: -11.2 (SimCLR), -17.7 (MSN), -17.7 (VICReg) on ImageNet | Hurt |
| class-imbalanced mini-batch sampling instead of class-balanced sampling (ablation) | MAE, data2vec · ImageNet-1K and downstream transfer | downstream accuracy drop: -0.1 (MAE), -0.8 (data2vec) on ImageNet | No clear effect |
| replace uniform feature prior with the true iNat18 class prior (ablation) | PMSN on iNat18 · iNat18 pretraining, downstream transfer | downstream accuracy/score delta: +3.0 iNat18, +0.9 CIFAR100, +1.4 Places205, +2.1 Clevr/Count, +2.3 Clevr/Dist, +2.7 KITTI | Improved |
| power-law prior (tau=0.25) instead of uniform prior (ablation) | PMSN on iNat18 · iNat18 pretraining, downstream transfer | downstream accuracy delta: +1.0 iNat18, +0.7 CIFAR100, +0.8 Places205, +1.7 Clevr/Count, +2.1 Clevr/Dist, +0.0 KITTI | Improved |
| power-law prior (tau=0.25) instead of uniform prior (ablation) | PMSN on ImageNet-1K · ImageNet pretraining, downstream transfer | downstream accuracy delta: -19.9 iNat18, -17.8 CIFAR100, -15.7 Places205, -2.9 Clevr/Count, -6.7 KITTI; +8.0 Clevr/Dist | Hurt |
| power-law prior instead of uniform prior (ablation) | PMSN · Clevr/Dist and KITTI transfer | depth prediction delta: +8.0 Clevr/Dist on ImageNet but -6.7 KITTI on ImageNet | Mixed |
Closest papers by meaning
- Masked Siamese Networks for Label-Efficient Learning (ECCV 2022 · similarity 0.87)
- Efficient Unsupervised Visual Representation Learning with Explicit Cluster Balancing (ECCV 2024 · similarity 0.87)
- ClusterFit: Improving Generalization of Visual Representations (CVPR 2020 · similarity 0.87)
- A Theoretical Framework for Masked Pretraining (MPT) (JMLR 2026 · similarity 0.86)
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.85)
- Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.85)
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.
