Research map / Continual learning / Class-incremental
Happy: A Debiased Learning Framework for Continual 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 Happy, a rehearsal-free debiased framework for continual generalized category discovery that combines clustering-guided classifier initialization, group-wise soft entropy regularization, and hardness-aware prototype sampling with distillation.
Problem
Continual Generalized Category Discovery (C-GCD): incrementally discovering new classes from unlabeled data containing both old and new classes, without storing old samples, while avoiding prediction bias toward old classes and forgetting of hard old classes.
Approach
Treats old vs. new classes as a binary group-wise soft entropy regularization on marginal probabilities to counter prediction bias, initializes new heads with KMeans centroids least similar to old heads, and samples old-class prototype Gaussians with probabilities weighted by inter-prototype hardness to reduce forgetting, plus self-distillation and feature KD.
Key results
- On ImageNet-100 over 5 stages, Happy reaches 78.58 All ACC at stage 5 vs GM 71.08 (about 7.5% gain at the reported stage).
- On CIFAR100 with 10 stages, Happy achieves 57.81 All ACC at stage 10 vs 56.51 for MetaGCD.
- Ablation on CIFAR100: full model reaches 69.00 average All ACC vs 50.95 for the baseline without the proposed components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Happy instead of the strongest listed baseline | C-GCD with ViT-B/16 (DINO) · ImageNet-100, Stage-5 | All accuracy: 71.08 -> 78.58 | Helped |
| Add soft entropy regularization and clustering-guided initialization (ablation) | Happy · CUB, 5-stage average | New accuracy: 4.70 -> 32.91 | Helped |
| Add hardness-aware prototype sampling (ablation) | Happy with entropy regularization and clustering-guided initialization · CIFAR100, 5-stage average | Old accuracy: 67.39 -> 69.98 | Helped |
| Add knowledge distillation (ablation) | Happy with entropy regularization, clustering-guided initialization, and prototype sampling · CIFAR100, 5-stage average | Old accuracy: 69.98 -> 71.82 | Helped |
| Add knowledge distillation (ablation) | Happy with entropy regularization, clustering-guided initialization, and prototype sampling · CUB, 5-stage average | Old accuracy: 70.94 -> 71.29 | Helped |
| Add soft entropy regularization (ablation) | Happy category discovery · CIFAR100, after Stage-1 | New-sample misclassification rate to old classes (Δr): 63.25 -> 10.20 | Helped |
| Add hardness-aware modeling (ablation) | Happy forgetting mitigation · CIFAR100, after 5 stages | Hardest initial-class accuracy (Acc_h): 65.10 -> 70.23 | Helped |
| Use ground-truth old/new sample ratios as a prior in the regularizer (ablation) | Happy · Continual generalized category discovery | Accuracy | Hurt |
Surprising findings
- Adding the ground-truth new/old sample ratio as a prior to the regularization degrades performance because of the confidence gap between old and new classes.
- Baseline SimGCD nearly collapses on new classes in continual stages (e.g., 8.00% new ACC at stage 1 on CIFAR100).
Limitations
- Model confidence is not calibrated, leading to a confidence gap between old and new classes.
- Class number estimation for new classes is challenging in the unsupervised setting; silhouette-based estimation is used as a workaround.
- Evaluated only on classification tasks.
Experimental setup
Datasets: CIFAR100 ImageNet-100 TinyImageNet CUB Stanford Cars FGVC Aircraft CIFAR100-C
Benchmarks: CIFAR100 ImageNet-100 TinyImageNet CUB Stanford Cars FGVC Aircraft CIFAR100-C
Metrics: All accuracy Old accuracy New accuracy forgetting M_f discovery M_d silhouette-estimated class number accuracy
Baselines: KMeans VanillaGCD SimGCD SimGCD+ FRoST GM MetaGCD
Builds on: SimGCD GCD FRoST GM MetaGCD
Closest papers by meaning
- Continual Generalized Category Discovery: Learning and Forgetting from a Bayesian Perspective (ICML 2025 · similarity 0.92)
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.89)
- Solving the Catastrophic Forgetting Problem in Generalized Category Discovery (CVPR 2024 · similarity 0.89)
- DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (ICLR 2025 · similarity 0.88)
- Towards Understanding Parametric Generalized Category Discovery on Graphs (ICML 2025 · similarity 0.88)
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.88)
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.