Research map / Detection and segmentation / CNN features
SAAL: Sharpness-Aware Active Learning
Yoon-Yeong Kim, Youngjae Cho, JoonHo Jang, Byeonghu Na, Yeongmin Kim, Kyungwoo Song, Wanmo Kang, Il-chul Moon
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 SAAL acquisition (perturbed pseudo-label loss) instead of baselines with Adam optimizer | active learning on image classification · CIFAR-10 | test accuracy (%): Random 50.7 -> SAAL 54.4 (Adam, with k-means++ variant) | Improved |
| Add SAM optimizer to SAAL acquisition (ablation) | SAAL · CIFAR-10 | test accuracy (%): SAAL Adam 52.3 -> SAAL SAM 56.0; on Fashion SAM lowers SAAL 85.6 -> 85.0 | Mixed |
| Add k-means++ seeding diversity to SAAL acquisition (ablation) | SAAL · CIFAR-10 | test accuracy (%): Adam 52.3 -> 54.4; SAM 56.0 -> 57.0 (reported as w/ k-means++) | Improved |
| Use SAAL acquisition instead of BADGE with Adam optimizer | active learning on image classification · Fashion-MNIST | test accuracy (%): BADGE 85.4 -> SAAL 85.6 | Improved |
| Use SAAL acquisition instead of BADGE with SAM optimizer | active learning on image classification · Fashion-MNIST | test accuracy (%): BADGE 86.2 -> SAAL 85.0 | Hurt |
| Use SAAL acquisition instead of baselines on long-tailed CIFAR-10 (ablation) | active learning on imbalanced data · Long-tailed CIFAR-10 | test accuracy (%): BADGE 21.69 -> SAAL 23.03 | Improved |
| Use SAAL score (perturbed loss) for object detection image selection with SSD | SSD object detection · PASCAL VOC 2007+2012 | mAP: BADGE 0.7493 -> SAAL 0.7541 | Improved |
| Combine SAAL score with diversity score in region-based annotation (RI-SAAL) for domain adaptive segmentation | SYNTHIA to CityScapes segmentation · SYNTHIA to CityScapes | mIOU: RIPU 70.2 -> RI-SAAL 70.6 | Improved |
| Use SAM-perturbation sensitivity radius rho=0.05 vs 0.01 or 0.10 (ablation) | SAAL · CIFAR-10 | test accuracy (%) | No clear effect |
| Batch-wise perturbation (batch size 10) instead of per-instance perturbation in SAAL (ablation) | SAAL · CIFAR-10 | test accuracy (%): Adam: BS1 54.4 vs BS10 54.0; SAM: 57.0 vs 57.7; wall time drops from 49.6 s to 16.8 s | No clear effect |
| Use SAM optimizer instead of Adam for SAAL (ablation) | SAAL · CIFAR-100 | test accuracy (%): 46.6 -> 48.4 | Improved |
| Use SAAL acquisition in high-budget setting (1,000 per step) (ablation) | active learning on image classification · Fashion, SVHN, CIFAR-10, CIFAR-100 (high budget) | test accuracy: small margin over baselines | Improved |
Closest papers by meaning
- Active Learning for Convolutional Neural Networks: A Core-Set Approach (ICLR 2018 · similarity 0.86)
- A Cross-Domain Benchmark for Active Learning (NeurIPS 2024 · similarity 0.85)
- Plug and Play Active Learning for Object Detection (CVPR 2024 · similarity 0.83)
- Sharpness-Aware Minimization: General Analysis and Improved Rates (ICLR 2025 · similarity 0.83)
- Adam Reduces a Unique Form of Sharpness: Theoretical Insights Near the Minimizer Manifold (NeurIPS 2025 · similarity 0.83)
- AutoAL: Automated Active Learning with Differentiable Query Strategy Search (ICML 2025 · similarity 0.83)
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.
