Research map / Label-efficient and robust learning / Recognition and long-tail
Active Generalized Category Discovery
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 Active Generalized Category Discovery (AGCD), which actively queries labels for novel-class samples during GCD training, using an Adaptive-Novel sampling strategy and a stable EMA-based label mapping.
Problem
Generalized category discovery (GCD) suffers from imbalanced accuracy and inconsistent confidence between old and new classes, especially in low-label regimes, and labeling all novel classes is costly.
Approach
Extends GCD with pool-based active learning over unlabeled data containing novel classes; the Adaptive-Novel strategy selects confident novel samples in early rounds and informative novel samples later (switching based on label-mapping stability), with class-wise diversity, and a stable label mapping via EMA model predictions aligns queried ground-truth labels with the classifier's label indices.
Key results
- Adaptive-Novel reaches 71.25% All accuracy on CIFAR100 after 5 rounds versus 67.28% for Random and 65.08%/66.62% Margin vs. ours on CUB (66.62% vs. 62.74% Random).
- On CUB with ~2.5 queries per class, new-class accuracy improves by 25.52% over the GCD baseline without AGCD.
- On ImageNet-100 the old-new accuracy gap is reduced from about 28% to 13.72%, and the method remains robust with estimated class numbers (82.46% All vs. 77.86% Random).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Adaptive-Novel instead of Margin | SimGCD active category discovery · CIFAR10 | All accuracy: 92.34 -> 93.15 | Helped |
| Use Adaptive-Novel instead of BADGE | SimGCD active category discovery · CIFAR100 | All accuracy: 67.22 -> 71.25 | Helped |
| Use Adaptive-Novel instead of BADGE | SimGCD active category discovery · ImageNet-100 | All accuracy: 81.48 -> 83.34 | Helped |
| Use Adaptive-Novel instead of BADGE | SimGCD active category discovery · CUB | All accuracy: 65.84 -> 66.62 | Helped |
| Use Adaptive-Novel instead of Margin | SimGCD active category discovery · Stanford Cars | All accuracy: 46.03 -> 48.36 | Helped |
| Use Adaptive-Novel instead of BADGE | SimGCD active category discovery · FGVC-Aircraft | All accuracy: 52.03 -> 53.74 | Helped |
| Add novelty-aware sampling to random selection (ablation) | SimGCD active category discovery · CIFAR100 | All accuracy: 67.28 -> 69.33 | Helped |
| Add informativeness to novelty-only selection (ablation) | SimGCD active category discovery · CIFAR100 | All accuracy: 69.33 -> 69.65 | Helped |
| Add class-wise diversity to novelty and informativeness selection (ablation) | SimGCD active category discovery · CIFAR100 | All accuracy: 69.65 -> 71.25 | Helped |
| Add novelty-aware sampling to random selection (ablation) | SimGCD active category discovery · CUB | All accuracy: 62.74 -> 63.58 | Helped |
| Add informativeness to novelty-only selection (ablation) | SimGCD active category discovery · CUB | All accuracy: 63.58 -> 64.28 | Helped |
| Add class-wise diversity to novelty and informativeness selection (ablation) | SimGCD active category discovery · CUB | All accuracy: 64.28 -> 66.62 | Helped |
Surprising findings
- Sampling more novel samples does not necessarily help: BADGE selects more novel samples than Margin on Stanford Cars but performs worse.
- Conventional entropy-based active learning can underperform random sampling in some settings (e.g., Entropy 76.25% vs Random 82.74% on CIFAR10 All accuracy).
Limitations
- Relies on the stability of label mapping computed from very few labeled samples, especially for new classes.
- Evaluated only on image classification benchmarks with a single ViT-B/16 DINO backbone.
- Sampling-strategy switch relies on a heuristic stability criterion.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet-100 CUB Stanford Cars FGVC-Aircraft
Benchmarks: CIFAR10 CIFAR100 ImageNet-100 CUB Stanford Cars FGVC-Aircraft
Metrics: accuracy (All/Old/New, Hungarian-matched) Novelty Coverage Novelty Ratio Novelty Uniformity Novelty Information
Baselines: Random Entropy LeastConf Margin KMeans CoreSet BADGE
Builds on: SimGCD GCD Margin sampling DINO
Closest papers by meaning
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.90)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.90)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.89)
- Happy: A Debiased Learning Framework for Continual Generalized Category Discovery (NeurIPS 2024 · similarity 0.89)
- Solving the Catastrophic Forgetting Problem in Generalized Category Discovery (CVPR 2024 · similarity 0.89)
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.89)
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.