Research map / Continual learning / Class-incremental
Continual Generalized Category Discovery: Learning and Forgetting from a Bayesian Perspective
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 VB-CGCD, a Bayesian continual generalized category discovery method that fits per-class Gaussians via variational inference with a Mahalanobis nearest-class-mean classifier, covariance-based early stopping, and self-correcting re-labeling on a frozen DINO ViT-B/16 backbone.
Problem
Continual generalized category discovery (C-GCD) must learn new classes from mixed known/novel unlabeled streams without catastrophic forgetting of old classes, under exemplar-free online settings.
Approach
Model each class as a multivariate Gaussian fitted by stochastic variational inference, classify with a Mahalanobis nearest-class-mean rule, re-label pseudo-labels using the stored old-class distributions, and stop training early when the new-class covariance determinant matches the old-class average (a covariance-alignment criterion motivated by Bhattacharyya distance).
Key results
- On the standard 5-session setting, VB-CGCD reaches final overall accuracy of 81.23% on CIFAR100 and 66.44% on CUB200 (ablation table), with the paper reporting an average +15.21% gain over prior art.
- On the label-limited setting (10% classes labeled, 9 sessions) VB-CGCD reaches 67.86% final accuracy vs. 38.55% for state of the art, according to the abstract.
- Ablation shows SVI has the largest effect; removing it drops CIFAR100 final accuracy from 81.23% to 68.90%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add stochastic variational inference instead of point-estimating class distributions (ablation) | VB-CGCD · C100, TinyImageNet, IN100, and CUB, session 5 | All accuracy: 68.90 -> 81.23; 61.14 -> 74.52; 79.90 -> 87.32; 27.06 -> 66.44 | Helped |
| add offline fine-tuning (ablation) | VB-CGCD · C100, TinyImageNet, IN100, and CUB, session 5 | All accuracy: 73.48 -> 81.23; 68.69 -> 74.52; 84.0 -> 87.32; 51.10 -> 66.44 | Helped |
| add covariance-driven early stopping (ablation) | VB-CGCD · C100, TinyImageNet, IN100, and CUB, session 5 | All accuracy: 77.71 -> 81.23; 71.23 -> 74.52; 84.96 -> 87.32; 56.43 -> 66.44 | Helped |
| add self-corrective re-labeling (ablation) | VB-CGCD · C100, TinyImageNet, IN100, and CUB, session 5 | All accuracy: 80.20 -> 81.23; 73.67 -> 74.52; 86.54 -> 87.32; 63.32 -> 66.44 | Helped |
| use Mahalanobis distance rather than Euclidean distance (ablation) | VB-CGCD · C100, TinyImageNet, IN100, and CUB, session 5 | All accuracy: 76.06 -> 81.23; 70.33 -> 74.52; 85.06 -> 87.32; 55.79 -> 66.44 | Helped |
| use covariance-driven early stopping rather than training without it (ablation) | VB-CGCD · C100, TinyImageNet, IN100, and CUB, novelty learning rate | novelty learning rate: 81.62 -> 79.14; 70.60 -> 68.12; 89.04 -> 87.04; 60.77 -> 56.46 | Mixed |
| use covariance-driven early stopping rather than training without it (ablation) | VB-CGCD · C100, TinyImageNet, IN100, and CUB, forgetting rate | forgetting rate: 15.8 -> 6.94; 14.24 -> 5.72; 11.92 -> 5.68; 29.22 -> 6.72 | Helped |
| vary the covariance structure from full to diagonal (ablation) | VB-CGCD · C100, TinyImageNet, IN100, and CUB, session 5 | All accuracy: 81.23 vs 74.11; 74.52 vs 70.04 | Hurt |
| vary feature dimensionality to 100, 200, or 384 (ablation) | VB-CGCD · C100, session 5 | All accuracy: 76.41; 77.73; 81.23 | Mixed |
| vary the early-stopping ratio threshold from 0 to other reported values (ablation) | VB-CGCD · C100, final reported session | overall accuracy: 75.61; 76.13; 81.11; 81.23; 77.91; 77.71 | Mixed |
| use VB-CGCD instead of Happy | C-GCD, 10% labeled categories and 9 online sessions · C100, TinyImageNet, IN100, and CUB | All accuracy: 36.16 -> 74.72; 22.11 -> 66.30; 59.12 -> 82.88; 36.80 -> 47.46 | Helped |
| use VB-CGCD instead of Happy | C-GCD · C100, TinyImageNet, IN100, and CUB, session 5 | All accuracy: 60.77 -> 81.23; 54.93 -> 74.52; 78.58 -> 87.32; 54.37 -> 66.44 | Helped |
Surprising findings
- Euclidean-distance variant is reported to lose ~5.56 points on average, but on CUB the Euclidean variant has far lower forgetting (0.82) at the cost of novelty learning (28.15), showing the forgetting-plasticity trade-off.
Limitations
- Relies on quality of initial clustering, especially in high-dimensional spaces.
- Small-sample covariance estimation is weak on CUB200 (about 25 samples per class), reducing new-class accuracy and re-labeling accuracy (~65%).
Experimental setup
Datasets: CIFAR100 TinyImageNet ImageNet100 CUB200
Benchmarks: CIFAR100 TinyImageNet ImageNet100 CUB200
Metrics: All accuracy Old accuracy New accuracy forgetting rate novelty learning rate
Baselines: GM MetaGCD PromptCCD Happy
Builds on: GEM Happy PromptCCD BOFT
Closest papers by meaning
- Happy: A Debiased Learning Framework for Continual Generalized Category Discovery (NeurIPS 2024 · similarity 0.92)
- Solving the Catastrophic Forgetting Problem in Generalized Category Discovery (CVPR 2024 · similarity 0.90)
- Consistent Supervised-Unsupervised Alignment for Generalized Category Discovery (NeurIPS 2025 · similarity 0.88)
- Adaptive Discovering and Merging for Incremental Novel Class Discovery (AAAI 2024 · similarity 0.88)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.87)
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.87)
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.