Research map / Label-efficient and robust learning / Self-supervised and contrastive
ArCL: Enhancing Contrastive Learning with Augmentation-Robust 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 |
|---|---|---|---|
| Replace average positive-pair alignment with ArCL worst-pair alignment | SimCLR · Modified CIFAR10 Aug 1, batch size 512 | accuracy (%): 88.90 -> 89.97 | Helped |
| Replace average positive-pair alignment with ArCL worst-pair alignment | SimCLR · Modified CIFAR100 Aug 1, batch size 512 | accuracy (%): 52.34 -> 53.40 | Helped |
| Add ArCL worst-pair alignment with 6 views rather than standard two-view alignment | SimCLR · MNIST-CIFAR OOD test | accuracy (%): 85.6 -> 88.4 | Helped |
| Increase ArCL views from 3 to 6 (ablation) | SimCLR · MNIST-CIFAR OOD test | accuracy (%): 86.0 -> 88.4 | Helped |
| Increase average-alignment views from 3 to 6 (ablation) | SimCLR + AAL · MNIST-CIFAR OOD test | accuracy (%): 85.4 -> 85.8 | No clear effect |
| Add ArCL to MoCo v2 | MoCo v2 · ImageNet transfer, 9 small datasets, linear evaluation, 800+50 epochs | Avg accuracy: 72.21 -> 74.35 | Helped |
| Add ArCL to MoCo v2 | MoCo v2 · ImageNet transfer, 9 small datasets, finetuning, 800+50 epochs | Avg accuracy: 83.59 -> 86.37 | Helped |
| Increase ArCL views from 2 to 4 (ablation) | MoCo v2 · ImageNet transfer, 9 small datasets, linear evaluation, 800+50 epochs | Avg accuracy: 73.91 -> 74.35 | Helped |
| Increase ArCL views from 2 to 4 (ablation) | MoCo v2 · ImageNet transfer, 9 small datasets, finetuning, 800+50 epochs | Avg accuracy: 86.10 -> 86.37 | Helped |
| Increase ArCL views from 4 to 8 (ablation) | SimCLR · Modified CIFAR100 Aug 5, batch size 512 | accuracy (%): 53.77 -> 53.21 | Hurt |
| Increase ArCL views from 2 to 4 (ablation) | MoCo v2 · Modified CIFAR10 Aug 1 | accuracy (%): 89.12 -> 90.77 | Helped |
| Use average alignment instead of ArCL worst-pair alignment (ablation) | SimCLR · Modified CIFAR10 Aug 1, batch size 512 | accuracy (%): 89.97 -> 88.90 | Hurt |
Closest papers by meaning
- An Augmentation-Aware Theory for Self-Supervised Contrastive Learning (ICML 2025 · similarity 0.91)
- Towards the Generalization of Contrastive Self-Supervised Learning (ICLR 2023 · similarity 0.90)
- Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap (ICLR 2022 · similarity 0.89)
- Data-Efficient Contrastive Self-supervised Learning: Most Beneficial Examples for Supervised Learning Contribute the Least (ICML 2023 · similarity 0.89)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.89)
- 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.