Research map / Label-efficient and robust learning / Recognition and long-tail
Towards Controlled Data Augmentations for Active Learning
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 |
|---|---|---|---|
| Add CAMPAL augmentation-induced acquisition and training to AL (CAMPAL* with Chamfer BADGE) | active learning on CIFAR-10 · CIFAR-10, NL=2000 | test accuracy: +13.34 over BADGE | Helped |
| Add CAMPAL to score-based and representation-based acquisitions (Entropy, LC, Margin, Core-set, BADGE) (ablation) | active learning on CIFAR-10 and CIFAR-100, FashionMNIST, SVHN · Table 3 comparisons at 2000 labeled (CIFAR-10) and other datasets | test accuracy gain over non-augmented baseline: +0.28 to +13.34 across methods and datasets | Helped |
| Apply augmentation only to the labeled pool (LA) for Entropy (ablation) | Entropy AL on CIFAR-10 · CIFAR-10, NL=2000 | test accuracy: +8.35 over baseline 65.95 | Helped |
| Apply augmentation only to the unlabeled pool (UA) for Entropy (ablation) | Entropy AL on CIFAR-10 · CIFAR-10, NL=2000 | test accuracy: +1.54 over baseline 65.95 | Helped |
| Apply augmentation to both labeled and unlabeled pools versus only one (ablation) | Entropy AL on CIFAR-10 · CIFAR-10, NL=2000 | test accuracy: 76.90 vs 74.30 (LA only) and 67.49 (UA only) | Helped |
| Integrate CAMPAL* into FixMatch SSL (ablation) | FixMatch on CIFAR-10 · CIFAR-10, NL=40/250/1000 | test accuracy: +0.55, +0.61, +0.15 | Helped |
| Integrate CAMPAL* into Pseudo Label SSL (ablation) | Pseudo Label on CIFAR-10 · CIFAR-10, NL=40/250/1000 | test accuracy: +5.05, +13.20, +9.98 | Helped |
| Integrate CAMPAL* into Mean Teacher and UDA (ablation) | Mean Teacher, UDA on CIFAR-10 · CIFAR-10, NL=40/250/1000 | test accuracy: Mean Teacher +4.45/+2.43/+1.05; UDA +1.99/+1.76/+0.79 | Helped |
| Integrate CAMPAL* with FixMatch under class imbalance (ablation) | FixMatch on imbalanced CIFAR-10 · CIFAR-10 imbalanced, IF=0.1, 1000 samples | test accuracy: 54.72 baseline to 67.34 with both (+12.62) | Helped |
| Use CAMPAL on imbalanced CIFAR-10 with a different imbalance factor (ablation) | CAMPAL* alone on imbalanced CIFAR-10 · CIFAR-10 imbalanced, IF=0.01, 1000 samples | test accuracy: 46.74 to 51.02 (+4.28) | Helped |
| Apply CAMPAL to LADA-style augmentation on SVHN versus CAMPAL at SVHN scale | Existing augmentation-based AL (LADA, CAL, BGADL) · CIFAR-10 NL=500 and NL=2000; SVHN | test accuracy | No clear effect |
| Use unlabeled-pool strength larger than labeled-pool strength (learned strengths) | CAMPAL strength optimization on CIFAR-10 · CIFAR-10, 2000 labeled | learned augmentation strength | No clear effect |
Closest papers by meaning
- Bayesian Generative Active Deep Learning (ICML 2019 · similarity 0.87)
- LADA: Look-Ahead Data Acquisition via Augmentation for Deep Active Learning (NeurIPS 2021 · similarity 0.85)
- Learning Active Learning from Data (NeurIPS 2017 · similarity 0.85)
- The Power of Comparisons for Actively Learning Linear Classifiers (NeurIPS 2020 · similarity 0.85)
- Learning Data Manipulation for Augmentation and Weighting (NeurIPS 2019 · similarity 0.85)
- Understanding when Dynamics-Invariant Data Augmentations Benefit Model-free Reinforcement Learning Updates (ICLR 2024 · similarity 0.84)
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.