Research map / Label-efficient and robust learning / Recognition and long-tail
Cleaning the Pool: Progressive Filtering of Unlabeled Pools in Deep Active Learning
Denis Huseljic, Marek Herde, Lukas Rauch, Paul Hahn, Bernhard Sick
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 |
|---|---|---|---|
| increase filtering rounds R from 1 to 7 (ablation) | progressive filtering + random sampling · CIFAR-10 | AULC improvement [%]: 3.02 -> 3.81 | Improved |
| increase filtering rounds R from 1 to 9 (ablation) | progressive filtering + random sampling · Snacks | AULC improvement [%]: 6.91 -> 8.43 | Improved |
| increase R from 7 to 9 (ablation) | progressive filtering · CIFAR-10 | AULC improvement [%]: 3.81 -> 3.78 | No clear effect |
| use very low sampling ratio alpha=0.2 instead of 0.3 (ablation) | progressive filtering · CIFAR-10 | AULC improvement [%]: 3.89 -> 3.33 | Hurt |
| vary alpha between 0.3 and 0.9 (ablation) | progressive filtering · Snacks | AULC improvement [%]: 7.60 -> 8.40 | No clear effect |
| increase batches per strategy J from 1 to 20 (ablation) | progressive filtering · CIFAR-10 | AULC improvement [%]: 3.57 -> 3.93 | Improved |
| increase J from 1 to 20 (ablation) | progressive filtering · Snacks | AULC improvement [%]: 7.88 -> 7.69 | No clear effect |
| apply progressive filtering as preprocessing to individual AL strategies | random sampling, Margin, BAIT, AlfaMix, UHerding · CIFAR-100 (DINOv2/DINOv3) | AULC improvement [%]: +0.7% to +5.4% | Improved |
| apply progressive filtering before BAIT | BAIT · CIFAR-100 (DINOv3) | AULC improvement [%]: +4.9% | Improved |
| apply progressive filtering before Coreset | Coreset · ESC50 (EAT base) | accuracy: up to 10% | Improved |
| use FIR or expected-error batch ranking instead of coverage-based second stage (ablation) | REFINE · Dopanim and CIFAR-100 (DINOv2) | relative accuracy | No clear effect |
| increase batch size b from 10 to 100 at fixed budget (ablation) | REFINE · CIFAR-100 (DINOv2) | accuracy | Improved |
Closest papers by meaning
- Active Learning Guided by Efficient Surrogate Learners (AAAI 2024 · similarity 0.85)
- AutoAL: Automated Active Learning with Differentiable Query Strategy Search (ICML 2025 · similarity 0.85)
- One-shot Active Learning Based on Lewis Weight Sampling for Multiple Deep Models (ICLR 2024 · similarity 0.85)
- Entropic Open-Set Active Learning (AAAI 2024 · similarity 0.84)
- Provably Neural Active Learning Succeeds via Prioritizing Perplexing Samples (ICML 2024 · similarity 0.84)
- Instance-wise Supervision-level Optimization in Active Learning (CVPR 2025 · 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.
