Research map / Label-efficient and robust learning / Domain adaptation and generalization
When Domain Generalization meets Generalized Category Discovery: An Adaptive Task-Arithmetic Driven Approach
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.
Introduces domain-generalized generalized category discovery (DG-GCD) and DG2CD-Net, which trains a ViT with episodic cross-domain GCD tasks on source plus diffusion-synthesized domains and merges episode task vectors with validation-performance-based softmax weights.
Problem
Generalized category discovery under domain shift when only labeled source data (with base classes only) is available during training and the target domain, containing both known and novel classes, is unseen until inference.
Approach
Episodic training on cross-domain GCD tasks built from source data and InstructPix2Pix-synthesized pseudo-target domains, combined with a margin loss for open-set domain alignment and a task-arithmetic update whose task vectors are weighted by softmax over GCD (All) scores on a held-out validation distribution.
Key results
- DG2CD-Net achieves 52.06% average All accuracy across PACS, Office-Home and DomainNet, versus 48.99% for the best DG-trained CD-GCD baseline (CDAD-Net).
- DG2CD-Net reaches 73.30% All on PACS, the top result among DG settings in Table 1.
- Removing synthetic domains reduces PACS All accuracy by 22.39 points; replacing softmax weighting with min-max normalization reduces it by 11.32 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace generated synthetic domains with manual augmentations (ablation) | DG2CD-Net · PACS | All: 63.92 vs 73.30; decreases it by 9.38% | Hurt |
| Remove synthetic domains (ablation) | DG2CD-Net · PACS | All: 50.91 vs 73.30; reduces the All metric by 22.39% | Hurt |
| Use a single global update instead of iterative updates (ablation) | DG2CD-Net · PACS | All: 66.70 vs 73.30; lowers the All metric by 6.60% | Hurt |
| Fix known/novel class splits across episodes (ablation) | DG2CD-Net · PACS | All: 68.60 vs 73.30; 4.70% decrease | Hurt |
| Replace softmax task-vector weighting with min-max normalization (ablation) | DG2CD-Net · PACS | All: 61.98 vs 73.30; reduces All by 11.32% | Hurt |
| Use episode-specific validation data instead of a separate held-out Dvalid (ablation) | DG2CD-Net · PACS | All: 69.26 vs 73.30; decreases it by 4.04% | Hurt |
| Replace proposed CD-GCD losses with CDAD-Net losses (ablation) | DG2CD-Net · PACS | All: 69.82 vs 73.30; 3.48% drop | Hurt |
| Replace adaptive task-vector aggregation with Fisher merging (ablation) | DG2CD-Net · PACS | All: 71.67 vs 73.30; 1.63% decline | Hurt |
| Add supervised contrastive loss Ls_con (ablation) | DG2CD-Net · PACS | All: 41.98 -> 60.63 | Helped |
| Add unsupervised contrastive loss Lu_con (ablation) | DG2CD-Net · PACS | All: 41.98 -> 67.84 | Helped |
| Combine supervised and unsupervised contrastive losses (ablation) | DG2CD-Net · PACS | All: 41.98 -> 68.70 | Helped |
| Combine supervised contrastive loss and open-set adversarial loss (ablation) | DG2CD-Net · PACS | All: 41.98 -> 64.16 | Helped |
Surprising findings
- Vanilla GCD trained on source plus synthetic domains performs best among baselines on some settings but suffers gradient conflicts, and the mixed distributions reduce stability.
- Increasing episodes beyond six yields diminishing returns, attributed to rising sign conflicts between task vectors.
Limitations
- Gains diminish beyond six episodes due to sign conflicts among task vectors.
- Relies on a separate held-out mixed validation distribution, which the paper argues is needed; per-episode validation performs worse.
- Evaluated only on image classification benchmarks (PACS, Office-Home, DomainNet).
Experimental setup
Datasets: PACS Office-Home DomainNet
Benchmarks: PACS Office-Home DomainNet
Metrics: All accuracy Old accuracy New accuracy cluster number estimation
Baselines: ViT GCD SimGCD CMS CDAD-Net SODG-Net DAML TIES-Merging Fisher Merging
Builds on: Task Arithmetic CDAD-Net GCD InstructPix2Pix DINO
Closest papers by meaning
- HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery (NeurIPS 2025 · similarity 0.92)
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.90)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.90)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.88)
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.88)
- Active 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.