Research map / Label-efficient and robust learning / Recognition and long-tail
Entropic Open-Set Active Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EOAL, an open-set active learning framework that combines a closed-set entropy (from class-aware binary classifiers) and a distance-based entropy (from clusters of queried unknowns) with diversity-based clustering to select known samples.
Problem
Active learning when the unlabeled pool contains unknown classes, which wastes annotation budget on unknown samples; existing open-set AL methods underuse the unknown samples obtained in AL rounds.
Approach
Two entropy scores: a closed-set entropy from K class-aware binary classifiers (pushed to maximal entropy on active unknowns) and a distance-based entropy over distances to FINCH-derived centers of active unknown clusters; their difference ranks samples, and per-cluster selection adds diversity.
Key results
- On CIFAR-100 at 20% mismatch ratio, EOAL outperforms MQNet by 3.88% and LfOSA by 6.00% in accuracy.
- Ablation on CIFAR-100: removing diversity clustering (w/o Diversity) drops final-cycle accuracy by 2.25%; diversity-only random selection drops by 11.33%.
- Removing the binary-classifier block (w/o BC) reduces accuracy by 3.21%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EOAL instead of MQNet | CIFAR-100, 20% mismatch ratio · CIFAR-100, 20% mismatch ratio | classification accuracy: +3.88% | Helped |
| Use EOAL instead of LfOSA | CIFAR-100, 20% mismatch ratio · CIFAR-100, 20% mismatch ratio | classification accuracy: +6.00% | Helped |
| Combine closed-set entropy and distance-based entropy in the query strategy (ablation) | EOAL · CIFAR-100, 20% mismatch ratio | classification accuracy | Helped |
| Add distance-based entropy and Tuplet loss to the closed-set-entropy-only variant (ablation) | EOAL · CIFAR-100, 20% mismatch ratio | classification accuracy | Helped |
| Add closed-set entropy and entropy-maximization loss to the distance-based-entropy-only variant (ablation) | EOAL · CIFAR-100, 20% mismatch ratio | classification accuracy | Helped |
| Train using active unknown samples (ablation) | EOAL · CIFAR-100, 20% mismatch ratio | classification accuracy | Helped |
| Train a separate target model on the labeled dataset for evaluation (ablation) | EOAL · CIFAR-100, 20% mismatch ratio | classification accuracy | Helped |
| Cluster samples and select low-score samples within each cluster instead of selecting globally (ablation) | EOAL · CIFAR-100, 20% mismatch ratio, last AL cycle | classification accuracy: 2.25% | Helped |
| Rank samples by the query score within clusters instead of randomly sampling from each cluster (ablation) | EOAL · CIFAR-100, 20% mismatch ratio | classification accuracy: 11.33% | Helped |
| Use binary classifiers to compute closed-set entropy instead of the first K logits of the classifier (ablation) | EOAL · CIFAR-100, 20% mismatch ratio | classification accuracy: 3.21% | Helped |
Limitations
- Requires the number of clusters to be fixed to the number of known classes K.
- Performance gap over standard AL shrinks or changes with mismatch ratio; LfOSA's advantage fades at high mismatch ratios.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 TinyImageNet
Benchmarks: CIFAR-10 CIFAR-100 TinyImageNet
Metrics: classification accuracy known-sample precision
Baselines: Random Entropy Certainty BALD Coreset BADGE MQNet LfOSA
Builds on: MQNet LfOSA FINCH Tuplet loss
Closest papers by meaning
- Rethinking Epistemic and Aleatoric Uncertainty for Active Open-Set Annotation: An Energy-Based Approach (CVPR 2025 · similarity 0.88)
- Inconsistency-Based Data-Centric Active Open-Set Annotation (AAAI 2024 · similarity 0.85)
- Provably Neural Active Learning Succeeds via Prioritizing Perplexing Samples (ICML 2024 · similarity 0.84)
- One-shot Active Learning Based on Lewis Weight Sampling for Multiple Deep Models (ICLR 2024 · similarity 0.83)
- Joint Out-of-Distribution Filtering and Data Discovery Active Learning (CVPR 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.