Research map / Label-efficient and robust learning / Self-supervised and contrastive
Directional Self-Supervised Learning for Heavy Image Augmentations
Yalong Bai, Yifan Yang, Wei Zhang, Tao Mei
Check 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 DSSL with heavy augmentations | SimSiam · CIFAR-10 linear evaluation | top-1 accuracy (%): 92.1 -> 94.5 | Improved |
| Add DSSL with heavy augmentations | BYOL · ImageNet linear evaluation | top-1 accuracy (%): 67.8 -> 68.3 | Improved |
| Add DSSL with heavy augmentations | SimSiam · ImageNet linear evaluation | top-1 accuracy (%): 67.4 -> 68.6 | Improved |
| Use RA(8,16) as the heavy augmentation with DSSL | BYOL · CIFAR-10 linear evaluation | top-1 accuracy (%): collapse -> 94.0 | Improved |
| Combine RA, UA, and Jigsaw as heavy augmentations with DSSL | BYOL · ImageNet linear evaluation | top-1 accuracy (%): 67.8 -> 69.3 | Improved |
| Add DSSL during pretraining | BYOL · COCO 2017 val object detection | AP: 36.9 -> 37.6 | Improved |
| Add DSSL during pretraining | SimSiam · COCO 2017 val instance segmentation | APm: 34.7 -> 35.1 | Improved |
| Add symmetric loss among heavily augmented views (ablation) | SimSiam with heavy augmentations · CIFAR-10 linear evaluation | top-1 accuracy (%): collapse | Hurt |
| Map standard-view features to heavily augmented-view targets (ablation) | SimSiam with heavy augmentations · CIFAR-10 linear evaluation | top-1 accuracy (%): 94.20 -> 59.74 | Hurt |
| Use directional prediction from heavily augmented views to their source standard views (ablation) | SimSiam with heavy augmentations · CIFAR-10 linear evaluation | top-1 accuracy (%): 59.74 -> 93.10 | Improved |
| Construct partially ordered views so each hard view is derived from a standard view (ablation) | SimSiam with heavy augmentations · CIFAR-10 linear evaluation | top-1 accuracy (%): 93.10 -> 94.20 | Improved |
Closest papers by meaning
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.87)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · similarity 0.87)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 2025 · similarity 0.87)
- Learning from Memory: Non-Parametric Memory Augmented Self-Supervised Learning of Visual Features (ICML 2024 · similarity 0.86)
- An Augmentation-Aware Theory for Self-Supervised Contrastive Learning (ICML 2025 · similarity 0.85)
- You Don’t Need Domain-Specific Data Augmentations When Scaling Self-Supervised Learning (NeurIPS 2024 · 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.
