Research map / Label-efficient and robust learning / OOD and anomaly detection
Joint Out-of-Distribution Filtering and Data Discovery 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 Joda, an active learning method that jointly filters out-of-distribution samples and discovers novel categories using only the labeled pool, with no auxiliary models or unlabeled-pool access.
Problem
Active learning in open-world settings where the unlabeled pool contains both OOD data and novel (undiscovered) categories, a combination not handled by prior open-set AL or category-discovery AL methods, which also often need auxiliary models and unlabeled-pool access.
Approach
Train one ResNet classifier with cross-entropy on inlier labeled data plus an outlier-exposure loss on OOD samples in the labeled pool, filter unlabeled samples by an energy-score threshold chosen via Youden's J, then select samples with a SISOMe-based score combined with class balancing.
Key results
- Joda achieves the highest accuracy in 17 of 18 evaluated combinations, per the authors.
- Joda reaches a selection precision of almost 1.0 in 8 of 10 CIFAR-100/TinyImageNet combinations while providing the fastest class discovery in all settings.
- Joda uses zero additional models and no unlabeled-pool data, compared with 1 to 2 auxiliary models in prior open-set AL methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Joda instead of competing active-learning methods | CIFAR-10, CIFAR-100, and TinyImageNet OSDAL settings · 18 OSDAL configurations | accuracy | Helped |
| Use Joda for sample selection and OOD filtering | CIFAR-10, CIFAR-100, and TinyImageNet OSDAL settings · OSDAL settings | class discovery | Helped |
| Use Joda for sample selection and OOD filtering | CIFAR-10, CIFAR-100, and TinyImageNet OSDAL settings · OSDAL settings | selection precision | Helped |
| Add outlier exposure (OE) loss (ablation) | Joda · CIFAR-100 with Places365 OOD | accuracy | Helped |
| Use outlier exposure (OE) rather than energy exposure (EE) (ablation) | Joda · CIFAR-100 with Places365 OOD | accuracy | Helped |
| Add energy-based OOD filtering (ablation) | Joda · CIFAR-100 with Places365 OOD | accuracy | Helped |
| Add energy-based OOD filtering (ablation) | Joda · CIFAR-100 with Places365 OOD | class discovery | Helped |
| Add energy-based OOD filtering (ablation) | Joda · CIFAR-100 with Places365 OOD | selection precision | Helped |
| Use SISOMe selection instead of pure energy-based selection (ablation) | Joda · CIFAR-100 with Places365 OOD | accuracy | Helped |
| Use SISOMe selection instead of pure energy-based selection (ablation) | Joda · CIFAR-100 with Places365 OOD | class discovery | Helped |
| Use SISOMe selection instead of pure energy-based selection (ablation) | Joda · CIFAR-100 with Places365 OOD | selection precision | Helped |
| Set the OE loss weight λ to 0.25 instead of 0.5 (ablation) | Joda · CIFAR-100 with Places365 OOD | accuracy | No clear effect |
Surprising findings
- Classic uncertainty-based AL methods such as LLoss and Ent perform competitively on some OOD settings, including LLoss on TinyImageNet.
Limitations
- Evaluation limited to image classification on three datasets; the authors note future work could add more datasets and OOD scenarios.
- Only the ResNet18 architecture is the main setting, with ResNet50 as a transfer check.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 TinyImageNet MNIST Places365 ImageNetC-800
Benchmarks: CIFAR-10 CIFAR-100 TinyImageNet
Metrics: accuracy on target classes number of discovered classes selection precision
Baselines: MQNet CCAL Pal LfOSA EIOAL Monte Carlo dropout entropy LLoss BADGE Random
Builds on: SISOMe Outlier Exposure Energy-based OOD detection Loss Learning for active learning
Closest papers by meaning
- OOD-Chameleon: Is Algorithm Selection for OOD Generalization Learnable? (ICML 2025 · similarity 0.86)
- AHA: Human-Assisted Out-of-Distribution Generalization and Detection (NeurIPS 2024 · similarity 0.85)
- Inconsistency-Based Data-Centric Active Open-Set Annotation (AAAI 2024 · similarity 0.85)
- ProHOC: Probabilistic Hierarchical Out-of-Distribution Classification via Multi-Depth Networks (CVPR 2025 · similarity 0.85)
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.84)
- Image Background Serves as Good Proxy for Out-of-distribution Data (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.