Research map / Label-efficient and robust learning / Self-supervised and contrastive
What Makes for Good Views for Contrastive Learning?
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 InfoMin Aug. rather than SimCLR pre-training | ResNet-50 ImageNet linear readout · ImageNet | Top-1 accuracy (%): 69.3 -> 73.0 | Helped |
| Train InfoMin Aug. for 800 rather than 200 epochs | ResNet-50 linear readout · ImageNet | Top-1 accuracy (%): 70.1 -> 73.0 | Helped |
| Learn views with semi-supervised loss rather than supervised-only loss (ablation) | STL-10 view generation (RGB) · STL-10 | accuracy: 79.9 ± 1.5 -> 86.0 ± 0.6 | Helped |
| Learn views with semi-supervised loss rather than supervised-only loss (ablation) | STL-10 view generation (YDbDr) · STL-10 | accuracy: 78.5 ± 2.3 -> 87.0 ± 0.3 | Helped |
| Add the unsupervised MI-reduction term to supervised view learning (ablation) | STL-10 view generation · STL-10 | accuracy: 79.9 ± 1.5 -> 86.0 ± 0.6; 78.5 ± 2.3 -> 87.0 ± 0.3 | Helped |
| Add a JigSaw branch to RA-CJ-Blur augmentation (ablation) | InfoMin pre-training · ImageNet linear readout | Top-1 accuracy (%): 63.6% to 65.9% | Helped |
| Replace the linear projection head with a 2-layer MLP (ablation) | InfoMin pre-training · ImageNet linear readout | Top-1 accuracy (%): 65.9% to 67.3% | Helped |
| Share digit and position between the two views instead of sharing background | Contrastive representation learning · Colorful Moving-MNIST digit classification | error rate (%): 88.6 -> 14.5 | Helped |
| Share the downstream-relevant factor between views | Contrastive representation learning · Colorful Moving-MNIST background classification | error rate (%): 88.6 -> 51.7 | Helped |
| Increase cropping strength beyond the usual low-area bound of 0.08 toward 0.2 (ablation) | MoCo augmentation · ImageNet linear classification | accuracy: more than 1% | Helped |
| Reduce view mutual information without preserving task-relevant information (ablation) | Unsupervised view learning · STL-10 view learning | downstream accuracy | Mixed |
Closest papers by meaning
- Rethinking Minimal Sufficient Representation in Contrastive Learning (CVPR 2022 · similarity 0.90)
- Decomposed Mutual Information Estimation for Contrastive Representation Learning (ICML 2021 · similarity 0.89)
- What Should Not Be Contrastive in Contrastive Learning (ICLR 2021 · similarity 0.89)
- Self-Contrastive Learning: Single-Viewed Supervised Contrastive Framework Using Sub-network (AAAI 2023 · similarity 0.89)
- An Augmentation-Aware Theory for Self-Supervised Contrastive Learning (ICML 2025 · similarity 0.88)
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 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.