Research map / Label-efficient and robust learning / Self-supervised and contrastive
Improving Self-Supervised Learning by Characterizing Idealized Representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use asymmetric projection (remove one projection head) in contrastive SSL (CISSL vs SimCLR) | SimCLR · TinyImageNet linear probe | accuracy: 44.9 -> 45.8 | Helped |
| use DISSL objective instead of best distillation baseline (DINO) | DINO · TinyImageNet linear probe | accuracy: 43.5 -> 45.5 | Helped |
| increase representation dimensionality 512 to 2048 (ablation) | CISSL · TinyImageNet linear probe | accuracy: 45.8 -> 47.6 | Helped |
| increase representation dimensionality 512 to 2048 (ablation) | DISSL · TinyImageNet linear probe | accuracy: 45.5 -> 47.9 | Helped |
| train longer, 300 to 1000 epochs (ablation) | DISSL · TinyImageNet linear probe | accuracy: 47.9 -> 49.2 | Helped |
| use coarser augmentations (ablation) | CISSL · TinyImageNet linear probe | accuracy: 48.7 -> 51.0 | Helped |
| use coarser augmentations (ablation) | DISSL · TinyImageNet linear probe | accuracy: 49.2 -> 50.6 | Helped |
| stronger label-preserving augmentations (25% to 200%) (ablation) | DISSL · TinyImageNet downstream sample sizes | accuracy | Helped |
| evaluate with MLP probe instead of linear probe (ablation) | DISSL · TinyImageNet all data | accuracy: 45.9 -> 46.4 | Helped |
| train ISSL with MLP head matching MLP probe instead of linear (ablation) | DISSL · TinyImageNet all data | accuracy: 46.4 -> 47.5 | Helped |
| increase representation dimension 2048 to 8192 (ours) for CISSL | CISSL · ImageNet linear probe | accuracy: 65.1 -> 67.7 | Helped |
| DISSL instead of SwAV with multi-crop | SwAV · ImageNet 400 epochs | accuracy: 73.5 -> 74.0 | Helped |
Closest papers by meaning
- Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation (ICLR 2024 · similarity 0.89)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.89)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · similarity 0.88)
- Measuring Self-Supervised Representation Quality for Downstream Classification Using Discriminative Features (AAAI 2024 · similarity 0.88)
- Self-supervised Representation Learning from Random Data Projectors (ICLR 2024 · similarity 0.88)
- Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.87)
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.