Research map / Label-efficient and robust learning / Recognition and long-tail
DebGCD: Debiased Learning with Distribution Guidance for 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 DebGCD, which co-trains an auxiliary debiased classifier with one-hot labels, a semantic distribution (OOD) detector in a separate feature space, and a certainty-based curriculum to improve generalized category discovery.
Problem
Generalized category discovery suffers from label bias between known and unknown classes, uniform treatment of unlabelled samples with varying certainty, and unexplicit modeling of semantic shift between known and novel classes.
Approach
An auxiliary debiased classifier trained with one-hot pseudo-labels in the same feature space as the GCD classifier, guided by a semantic distribution detector (one-vs-all binary classifiers on a separate l2-normalized projection) whose certainty score drives a curriculum weighting of unlabelled samples; trained end-to-end in one stage with no inference overhead.
Key results
- On SSB (CUB, Stanford Cars, FGVC-Aircraft), DebGCD reaches average All ACC 64.4, about 3 points above the second-best SPTNet (61.4).
- On ImageNet-1K, DebGCD reaches All ACC 65.0, surpassing the previous best by 7.9 points.
- Ablation on Stanford Cars: the full model improves All ACC from 53.8 (SimGCD baseline) to 65.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply debiased learning directly to the original GCD classifier instead of using an auxiliary classifier (ablation) | DebGCD · Stanford Cars | ACC All: 53.8 -> 51.3 | Hurt |
| Add an auxiliary classifier with debiased learning (ablation) | DebGCD · Stanford Cars | ACC All: 53.8 -> 58.5 | Helped |
| Add semantic distribution learning alone (ablation) | DebGCD · Stanford Cars | ACC All: 53.8 -> 56.5 | Helped |
| Add distribution guidance to debiased learning and semantic distribution learning (ablation) | DebGCD · Stanford Cars | ACC All: 60.7 -> 65.3 | Helped |
| Use only the labelled-sample debiased loss L_adl^s (ablation) | DebGCD · FGVC-Aircraft | ACC All: 54.2 -> 53.1 | Hurt |
| Use only the unlabelled-sample debiased loss L_adl^u (ablation) | DebGCD · FGVC-Aircraft | ACC All: 54.2 -> 57.9 | Helped |
| Use both labelled- and unlabelled-sample debiased losses (ablation) | DebGCD · FGVC-Aircraft | ACC All: 57.9 -> 61.7 | Helped |
| Set the debiasing threshold τ to 0.85 rather than 0.90 (ablation) | DebGCD · FGVC-Aircraft unlabelled set | ACC All: 59.4 -> 61.7 | Helped |
| Use a 1-layer rather than no projection network ρ_s (ablation) | DebGCD · SSB average | ACC All: 61.8 -> 63.1 | Helped |
| Use a 5-layer rather than 7-layer projection network ρ_s (ablation) | DebGCD · SSB average | ACC All: 64.1 -> 64.4 | No clear effect |
| Use DebGCD rather than InfoSieve | DINO ViT-B/16 · CUB | ACC All: 69.4 -> 66.3 | Hurt |
| Use DebGCD rather than the strongest reported baseline | DINO ViT-B/16 · CIFAR-10 | ACC All: 97.9 -> 97.2 | Hurt |
Surprising findings
- Directly applying debiased learning to the GCD classifier itself decreases performance (e.g., Stanford Cars All ACC 53.8 to 51.3), so an auxiliary classifier is needed.
- Using an l2-normalized feature space for the OOD detector worked better than the magnitude-aware space common in other OOD tasks.
Limitations
- Number of tuned transformer blocks (last two) was chosen as a compromise; tuning more blocks reportedly causes training instability.
- Relies on the known number of categories K; performance with estimated K is reported separately and is lower.
- Debiasing threshold tau (0.85) is calibrated on a labelled validation set.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-100 ImageNet-1K CUB Stanford Cars FGVC-Aircraft
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-100 ImageNet-1K CUB Stanford Cars FGVC-Aircraft Oxford-Pet Herbarium19
Metrics: ACC (All) ACC (Old) ACC (New) AUROC
Baselines: k-means RankStats+ UNO+ ORCA GCD XCon OpenCon PromptCAL DCCL GPC SimGCD μGCD
Builds on: SimGCD GCD DINO OpenCon
Closest papers by meaning
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.91)
- Solving the Catastrophic Forgetting Problem in Generalized Category Discovery (CVPR 2024 · similarity 0.90)
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.89)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.88)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.88)
- HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery (NeurIPS 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.