Research map / Label-efficient and robust learning / Recognition and long-tail
Transfer and Alignment Network 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 TAN, a BERT-based generalized category discovery method that transfers prototypes from labeled to unlabeled data (P2P and P2I transfer) and aligns instance features with augmentations and calibrated prototypes (I2P and I2I alignment) to improve novel-category accuracy.
Problem
Generalized category discovery in text: recognizing known and novel categories in unlabeled data using labeled data with only known categories; existing methods perform poorly on novel categories due to biased knowledge transfer and noisy representation learning.
Approach
Model categories as prototypes; transfer labeled (known-category) prototypes to correct the biased unlabeled prototypes via similarity-weighted top-k calibration (P2P Trans), and pull known-category instances toward labeled prototypes (P2I Trans); then align instances with calibrated prototypes and with their augmentations via contrastive learning.
Key results
- TAN reaches H-score 66.70 on BANKING, 86.64 on StackOverflow and 87.02 on CLINC, outperforming DPN (60.73, 83.13, 84.56).
- Novel-category accuracy improves on average by 4.76% over prior SOTA, per the paper.
- Ablation on BANKING: removing I2P Align drops Novel accuracy from 56.23 to 45.44 and H-score from 66.70 to 59.10.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TAN instead of DPN | GCD · BANKING | H-score: +5.97 | Helped |
| Use TAN instead of DPN | GCD · StackOverflow | H-score: +3.51 | Helped |
| Use TAN instead of DPN | GCD · CLINC | H-score: +2.46 | Helped |
| Use TAN instead of DPN | GCD · BANKING | Novel accuracy: +6.23 | Helped |
| Use TAN instead of DPN | GCD · StackOverflow | Novel accuracy: +4.13 | Helped |
| Use TAN instead of DPN | GCD · CLINC | Novel accuracy: +3.92 | Helped |
| Add P2I Trans (ablation) | TAN · BANKING | H-score: 65.46 -> 66.70 | Helped |
| Add P2P Trans (ablation) | TAN · BANKING | H-score: 63.45 -> 66.70 | Helped |
| Add labeled cross-entropy loss (L_ce) (ablation) | TAN · BANKING | H-score: 62.70 -> 66.70 | Helped |
| Add I2I Align (ablation) | TAN · BANKING | H-score: 61.26 -> 66.70 | Helped |
| Add unlabeled cross-entropy loss (L_u) (ablation) | TAN · BANKING | H-score: 60.11 -> 66.70 | Helped |
| Add I2P Align (ablation) | TAN · BANKING | H-score: 59.10 -> 66.70 | Helped |
Surprising findings
- Removing P2P Trans, I2P Align or L_u increases known-category accuracy while degrading novel-category accuracy and H-score.
Limitations
- Requires a labeled set containing known categories and assumes the number of categories is known (estimated via filtering algorithm or over-clustering).
- Evaluated only on three text intent/question classification benchmarks.
Experimental setup
Datasets: BANKING StackOverflow CLINC
Benchmarks: BANKING StackOverflow CLINC
Metrics: H-score Known accuracy Novel accuracy clustering accuracy (Hungarian) pseudo-label accuracy
Baselines: DeepCluster DCN DEC KM-BERT KM-GloVe AG-GloVe SAE Simple Semi-DC Self-Labeling CDAC+ DTC
Builds on: DPN DAC GCD SimCSE masked language modeling
Closest papers by meaning
- A Unified Knowledge Transfer Network for Generalized Category Discovery (AAAI 2024 · similarity 0.95)
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.89)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.89)
- Consistent Supervised-Unsupervised Alignment for Generalized Category Discovery (NeurIPS 2025 · similarity 0.89)
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.89)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 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.