Research map / Label-efficient and robust learning / Self-supervised and contrastive
Contrastive Learning with Stronger Augmentations
Xiao Wang, Guo-Jun Qi
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 stronger augmentation with DDM loss to MoCo v2 | MoCo v2 · ImageNet linear eval (200 epochs) | top-1 accuracy: 67.5 -> 69.4 | Improved |
| add multiple stronger augmentations (multi-crop) | CLSA · ImageNet linear eval (200 epochs) | top-1 accuracy: 69.4 -> 73.3 | Improved |
| use strong query instead of weak query in MoCo v2 (ablation) | MoCo v2 · ImageNet linear eval (200 epochs) | top-1 accuracy: 67.5 -> 67.7 | No clear effect |
| use strong query and strong key in MoCo v2 (ablation) | MoCo v2 · ImageNet linear eval (200 epochs) | top-1 accuracy: 67.5 -> 67.0 | Hurt |
| add DDM loss instead of plain contrastive loss (ablation) | CLSA · ImageNet linear eval (200 epochs) | top-1 accuracy: 68.0 -> 69.4 | Improved |
| train CLSA longer (100 to 200 epochs) (ablation) | CLSA · ImageNet linear eval | top-1 accuracy: 67.2 -> 69.4 | Improved |
| use CLSA at equal running time versus MoCo v2 | CLSA · ImageNet linear eval | top-1 accuracy: 67.5 -> 68.3 | Improved |
| add stronger augmentation strength s=5 versus none (ablation) | CLSA · ImageNet KNN (100 epochs, K=20) | KNN accuracy: 51.9 -> 53.0 | Improved |
| increase strength s from 5 to 7 (ablation) | CLSA · ImageNet KNN (100 epochs, K=20) | KNN accuracy: 53.0 -> 52.6 | No clear effect |
| add CLSA to BYOL | BYOL · ImageNet linear eval (200 epochs) | top-1 accuracy: 66.5 -> 67.3 | Improved |
| add CLSA to MoCo v2 (generalization) | MoCo v2 · ImageNet linear eval (200 epochs) | top-1 accuracy: 67.5 -> 69.4 | Improved |
| train longer with multi-crop CLSA (800 epochs) | CLSA* · ImageNet linear eval | top-1 accuracy: 76.2 vs 76.5 | Improved |
Closest papers by meaning
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 2025 · similarity 0.89)
- Enhancing Contrastive Learning with Variable Similarity (NeurIPS 2025 · similarity 0.89)
- RényiCL: Contrastive Representation Learning with Skew Rényi Divergence (NeurIPS 2022 · similarity 0.87)
- MetAug: Contrastive Learning via Meta Feature Augmentation (ICML 2022 · similarity 0.87)
- An Augmentation-Aware Theory for Self-Supervised Contrastive Learning (ICML 2025 · similarity 0.86)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.85)
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.
