Research map / Label-efficient and robust learning / Recognition and long-tail
Consistent Supervised-Unsupervised Alignment 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 NC-GCD, which pre-assigns fixed Simplex ETF prototypes and aligns known and novel category features to them via unsupervised and supervised ETF losses, with a Semantic Consistency Matcher to stabilize cluster-label assignments across iterations.
Problem
Generalized category discovery suffers from inconsistent optimization objectives for known vs. novel categories and category confusion/feature overlap, as prior methods dynamically learn prototypes or classifier weights.
Approach
Pre-assign fixed Simplex ETF prototypes inspired by Neural Collapse and align both labeled features (supervised ETF alignment) and confident clustered unlabeled features (unsupervised ETF alignment) to them via a consistent ETF alignment loss, with a Hungarian-style Semantic Consistency Matcher enforcing one-to-one cluster-to-label/prototype matching across clustering iterations.
Key results
- On six GCD benchmarks with ground-truth K, NC-GCD reaches 68.7% average All accuracy, +3.0 over SPT, and +4.1 on New accuracy.
- Without ground-truth K, NC-GCD averages 64.7% All accuracy, +3.1 over CMS.
- Ablation shows the full model (supervised + unsupervised ETF) improves six-dataset average All accuracy by 5.4 points over the baseline (63.3% to 68.7%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NC-GCD instead of SPT | GCD with DINOv1 and ground-truth K · Six datasets average | All accuracy: +3.0 | Helped |
| Use NC-GCD instead of SPT | GCD with DINOv1 and ground-truth K · Six datasets average | Novel accuracy: +4.1 | Helped |
| Use NC-GCD instead of CMS | GCD with estimated K · Six datasets average | All accuracy: +3.1 | Helped |
| Use NC-GCD instead of CMS | GCD with estimated K · Six datasets average | Novel accuracy: +4.4 | Helped |
| Add supervised ETF alignment to the no-alignment baseline (ablation) | NC-GCD · Six datasets average | All accuracy: 63.3 -> 64.6 | Helped |
| Add unsupervised ETF alignment to the no-alignment baseline (ablation) | NC-GCD · Six datasets average | All accuracy: 63.3 -> 68.0 | Helped |
| Combine supervised and unsupervised ETF alignment (ablation) | NC-GCD · Six datasets average | All accuracy: 63.3 -> 68.7 (+5.4) | Helped |
| Combine supervised and unsupervised ETF alignment (ablation) | NC-GCD · Six datasets average | Novel accuracy: +7.4 | Helped |
| Set ETF loss coefficient β to 0.5 instead of β=1 (ablation) | NC-GCD · CUB-200 | Final ACC (All): 76.33 -> 71.44 | Hurt |
| Set ETF loss coefficient β to 2 instead of β=1 (ablation) | NC-GCD · CUB-200 | Final ACC (All): 76.33 -> 71.42 | Hurt |
| Set ETF loss coefficient β to 2 instead of β=1 (ablation) | NC-GCD · CUB-200 | Best ACC (Novel): 77.31 -> 78.75 | Helped |
| Underestimate category count K by 20% instead of using the true K (ablation) | NC-GCD · CUB-200 | All accuracy: 74.8 -> 70.3 | Hurt |
Limitations
- Performance drops when the number of categories K is underestimated, especially for novel categories (e.g., -20% bias on CUB-200 lowers All accuracy to 70.3%).
- Relies on estimating K to construct the ETF prototype set when ground-truth K is unavailable.
Experimental setup
Datasets: CIFAR-100 ImageNet-100 CUB-200-2011 Stanford Cars FGVC Aircraft Herbarium 19
Benchmarks: CIFAR-100 ImageNet-100 CUB-200-2011 Stanford Cars FGVC Aircraft Herbarium 19
Metrics: clustering accuracy (All) Old accuracy New accuracy
Baselines: GCD DCCL PromptCAL GPC SimGCD PIM TRAILER SelEx CMS SPT ProtoGCD RLCD
Builds on: SimGCD CMS Neural Collapse DINO
Closest papers by meaning
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.89)
- SEAL: Semantic-Aware Hierarchical Learning for Generalized Category Discovery (NeurIPS 2025 · similarity 0.89)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.89)
- Continual Generalized Category Discovery: Learning and Forgetting from a Bayesian Perspective (ICML 2025 · similarity 0.88)
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.88)
- Solving the Catastrophic Forgetting Problem in Generalized Category Discovery (CVPR 2024 · 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.