Research map / Label-efficient and robust learning / Self-supervised and contrastive
Unleashing the Power of Contrastive Self-Supervised Visual Models via Contrast-Regularized Fine-Tuning
Yifan Zhang, Bryan Hooi, Dapeng Hu, Jian Liang, Jiashi Feng
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 |
|---|---|---|---|
| Use Core-tuning instead of CE-tuning | MoCo-v2 ResNet-50 fine-tuning · Image classification, 9-dataset Avg. | top-1 accuracy: 87.76 -> 90.47 | Improved |
| Use Core-tuning instead of CE-tuning | MoCo-v2 ResNet-50 fine-tuning · iNaturalist18 | top-1 accuracy: 61.72+/-0.18 -> 63.57+/-0.09 | Improved |
| Use Core-tuning instead of CE fine-tuning | MoCo-v2 ResNet-50 semantic segmentation · PASCAL VOC val2012 | MIoU: 78.42 -> 79.62 | Improved |
| Use Core-tuning instead of CE-Con | MoCo-v2 ResNet-50 domain generalization · PACS Avg. | accuracy: 85.65 -> 88.08 | Improved |
| Use Core-tuning instead of CE-Con | MoCo-v2 ResNet-50 domain generalization · VLCS Avg. | accuracy: 77.67 -> 80.34 | Improved |
| Use Core-tuning adversarial training instead of AT-CE-Con | MoCo-v2 ResNet-50 · CIFAR10, PGD-10 ℓ2 attack ε=0.5 | robust accuracy: 90.74 -> 92.97 | Improved |
| Add the original contrastive loss to CE-only fine-tuning (ablation) | MoCo-v2 ResNet-50 · ImageNet20 | top-1 accuracy: 88.28 -> 89.29 | Improved |
| Use manifold mixup instead of CE-only fine-tuning (ablation) | MoCo-v2 ResNet-50 · ImageNet20 | top-1 accuracy: 88.28 -> 90.67 | Improved |
| Use hardness-directed mixup with original contrastive loss instead of manifold mixup alone (ablation) | MoCo-v2 ResNet-50 · ImageNet20 | top-1 accuracy: 90.67 -> 92.20 | Improved |
| Replace original contrastive loss with focal contrastive loss while retaining hardness-directed mixup (ablation) | MoCo-v2 ResNet-50 · ImageNet20 | top-1 accuracy: 92.20 -> 92.73 | Improved |
| Add Core-tuning to fine-tuning across the nine image-classification datasets | ResNet-50 with contrastive self-supervised pre-training · Image classification, 9 natural image datasets | top-1 accuracy | Improved |
Closest papers by meaning
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.89)
- A Broad Study on the Transferability of Visual Representations With Contrastive Learning (ICCV 2021 · similarity 0.89)
- On the duality between contrastive and non-contrastive self-supervised learning (ICLR 2023 · similarity 0.88)
- Self-Contrastive Learning: Single-Viewed Supervised Contrastive Framework Using Sub-network (AAAI 2023 · similarity 0.88)
- Decoupled Contrastive Learning (ECCV 2022 · similarity 0.88)
- 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.
