Research map / Label-efficient and robust learning / Self-supervised and contrastive
A Closer Look at Benchmarking Self-supervised Pre-training with Image Classification
Markus Marks, Manuel Knott, Neehar Kondapaneni, Elijah Cole, Thijs Defraeye, Fernando Pérez‐Cruz, Pietro Perona
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 |
|---|---|---|---|
| add batch norm before the linear classifier (ablation) | MaskFeat ViT-B/16 linear probing · ImageNet-1k | top-1 accuracy: 3.7 -> 51.7 | Improved |
| add z-score normalization to kNN probing (ablation) | MAE ViT-B/16 · ImageNet-1k | top-1 accuracy: 26.6 -> 47.7 | Improved |
| add z-score normalization to kNN probing (ablation) | DINO ViT-B/16 · ImageNet-1k | top-1 accuracy: 74.7 -> 74.9 | No clear effect |
| add final batch norm to 10% fine-tuning (ablation) | DINO ViT-B/16 · ImageNet-1k 10% labels | top-1 accuracy: 75.4 -> 76.6 | Improved |
| add final batch norm to 1% fine-tuning (ablation) | DINO ViT-B/16 · ImageNet-1k 1% labels | top-1 accuracy: 65.6 -> 69.2 | Improved |
| switch backbone from ResNet-50 to ViT-B/16 (ablation) | DINO · ImageNet-1k fine-tuning | top-1 accuracy: 76.0 -> 81.3 | Improved |
| switch backbone from ResNet-50 to ViT-B/16 (ablation) | MoCo v3 · ImageNet-1k fine-tuning | top-1 accuracy: 76.2 -> 80.8 | Improved |
| use linear probing instead of full fine-tuning as in-domain predictor of OOD ranking (ablation) | 26 SSL models · OOD average (11 datasets) | Spearman r: 0.85 vs 0.79 | Improved |
| use in-domain probing instead of fine-tuning to predict OOD kNN ranking under categorical shift (ablation) | 26 SSL models · categorical shift | Spearman r: 0.96 vs 0.62 | Improved |
| use style shift instead of categorical shift for OOD fine-tuning ranking (ablation) | 26 SSL models · OOD FT | Spearman r: 0.73 -> 0.61 | Hurt |
| use 10% fine-tuning as predictor of OOD fine-tuning (ablation) | 26 SSL models · OOD FT | Spearman r | Improved |
| use ResNet to ViT backbone for generative vs discriminative comparison (ablation) | SSL families · ImageNet | fine-tune vs linear probe gap | Improved |
Closest papers by meaning
- When Does Self-supervision Improve Few-shot Learning? (ECCV 2020 · similarity 0.92)
- Dynamic Data Selection for Efficient SSL via Coarse-to-Fine Refinement (ECCV 2024 · similarity 0.92)
- Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation (ICLR 2024 · similarity 0.88)
- Measuring Self-Supervised Representation Quality for Downstream Classification Using Discriminative Features (AAAI 2024 · similarity 0.88)
- Probing the Mid-level Vision Capabilities of Self-Supervised Learning (CVPR 2025 · similarity 0.88)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · similarity 0.88)
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.
