Research map / Label-efficient and robust learning / Self-supervised and contrastive
A Broad Study on the Transferability of Visual Representations With Contrastive Learning
Ashraful Islam, Chun-Fu (Richard) Chen, Rameswar Panda, Leonid Karlinsky, Richard Radke, Rogerio Feris
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 supervised contrastive + self-supervised contrastive loss instead of cross-entropy | ResNet-50 pretraining · 12 downstream datasets, linear evaluation | top-1 accuracy: +5.63% over CE | Improved |
| add supervised contrastive + self-supervised contrastive loss instead of cross-entropy | ResNet-18 pretraining · 12 downstream datasets, 5-shot | top-1 accuracy: 60.01 -> 63.47 | Improved |
| add strong MoCo-style augmentation to CE (ablation) | CE · 12 downstream datasets, linear evaluation | top-1 accuracy: 75.67 -> 75.91 | No clear effect |
| add strong MoCo-style augmentation to CE (ablation) | CE · 12 downstream datasets, full fine-tuning | top-1 accuracy: 88.13 -> 88.27 | No clear effect |
| add contrastive pretraining for full-network fine-tuning on full data | SupCon+SelfSupCon · 12 downstream datasets, full fine-tuning | top-1 accuracy: 88.13 -> 88.56 | No clear effect |
| add contrastive pretraining for fine-tuning with 1000 samples | CE+SelfSupCon, SupCon+SelfSupCon · 12 downstream datasets, 1000 training samples | top-1 accuracy: 69.79 -> 70.88, 71.27 | Improved |
| add self-supervised contrastive loss to SupCon | SupCon · MS COCO detection | AP: ~0.5% AP | Improved |
| add self-supervised contrastive loss to CE | CE · VOC0712 detection | AP50: 81.58 -> 83.06 | Improved |
| add self-supervised contrastive loss to CE | CE · ImageNet-A | top-1 accuracy: +5.18% | Improved |
| add self-supervised contrastive loss to CE | CE · ImageNet-R | top-1 accuracy: +4.38% | Improved |
| use SelfSupCon alone instead of CE | CE · ImageNet val | top-1 accuracy: 76.60 -> 66.07 | Hurt |
| increase weight alpha of self-supervised loss from 1 to 2 in CE+SelfSupCon (ablation) | CE+SelfSupCon · transfer accuracy (12 datasets) vs ImageNet | top-1 accuracy: best ImageNet at alpha=1, best transfer at alpha=2 | Mixed |
Closest papers by meaning
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.91)
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.91)
- What Should Not Be Contrastive in Contrastive Learning (ICLR 2021 · similarity 0.91)
- Contrasting Contrastive Self-Supervised Representation Learning Pipelines (ICCV 2021 · similarity 0.91)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.89)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · 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.
