Research map / Label-efficient and robust learning / Domain adaptation and generalization
HIDISC: A Hyperbolic Framework for Domain Generalization with 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 HiDISC, a non-episodic hyperbolic (Poincaré ball) framework for domain generalization with generalized category discovery that uses GPT-4o-guided diffusion augmentation, Tangent CutMix, and a unified Busemann/contrastive/outlier loss, achieving state-of-the-art on PACS, Office-Home, and DomainNet.
Problem
Generalized category discovery under domain shift: classify unlabeled target-domain samples into seen or novel categories when only labeled source-domain data is available for training (DG-GCD).
Approach
Learns hyperbolic (Poincaré ball) embeddings with penalized Busemann prototype alignment, a hybrid geodesic/angular contrastive loss, and an adaptive outlier loss on Tangent CutMix pseudo-novel samples, using only 1–2 diffusion-generated synthetic domains instead of episodic multi-domain simulation.
Key results
- HiDISC achieves 75.07% All on PACS, 56.78% on Office-Home, and 30.51% on DomainNet, versus 73.30%, 53.86%, 29.01% for DG2CD-Net.
- Average All accuracy 54.12% vs 52.06% for DG2CD-Net (+2.06), using 2 synthetic domains vs 9.
- Reports over 96x lower training FLOPs than DG2CD-Net.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HiDISC instead of DG²CD-Net | DG-GCD · PACS | All accuracy: +1.00 | Helped |
| Use HiDISC instead of DG²CD-Net | DG-GCD · Office-Home | All accuracy: +2.92 | Helped |
| Use HiDISC instead of DG²CD-Net | DG-GCD · DomainNet | All accuracy: +1.21 | Helped |
| Add Busemann alignment loss (ablation) | HiDISC · Office-Home | All accuracy: 26.17 -> 56.32 | Helped |
| Add hybrid hyperbolic contrastive loss (ablation) | HiDISC · Office-Home | All accuracy: 26.17 -> 50.95 | Helped |
| Add outlier repulsion to Busemann and hybrid contrastive losses (ablation) | HiDISC · Office-Home | All accuracy: 56.29 -> 56.78 | Helped |
| Replace synthetic domain augmentation with manual augmentations (ablation) | HiDISC · Office-Home | All accuracy: 56.78 -> 50.80 | Hurt |
| Remove synthetic domain augmentation (ablation) | HiDISC · Office-Home | All accuracy: 56.07 -> 56.78 | Hurt |
| Replace learnable curvature with fixed curvature c=0.01 (ablation) | HiDISC · Office-Home | All accuracy: 56.23 -> 56.78 | Hurt |
| Replace learnable curvature with fixed curvature c=0.03 (ablation) | HiDISC · Office-Home | All accuracy: 55.67 -> 56.78 | Hurt |
| Replace Tangent CutMix with Euclidean CutMix (ablation) | HiDISC · Office-Home | All accuracy: 56.78 -> 53.46 | Hurt |
| Set Busemann slope φ to 0.10 instead of 0.75 (ablation) | HiDISC · Office-Home | New accuracy: 53.21 -> 47.07 | Hurt |
Surprising findings
- Training on source data alone (no synthetic domains) reaches 56.07% on Office-Home, close to the full model, and performance of DG2CD-Net degrades sharply with fewer synthetic domains.
- Excessive synthetic augmentation degrades novel-class accuracy, with New accuracy peaking at 2 synthetic domains.
Limitations
- Reliance on diffusion-generated synthetic augmentations may propagate generative artifacts or bias.
- Synthetic domain augmentation shows diminishing or negative returns beyond 1-2 domains, causing seen-class bias.
- Evaluated only on three benchmarks (PACS, Office-Home, DomainNet) with ViT backbone.
Experimental setup
Datasets: PACS Office-Home DomainNet
Benchmarks: PACS Office-Home DomainNet
Metrics: clustering accuracy (All) Old accuracy New accuracy
Baselines: ViT GCD SimGCD CMS SelfEx CDAD-Net Hyp-GCD Hyp-SelfEx DG2CD-Net Hyp-DG2CD-Net
Builds on: DG2CD-Net HypCD Busemann learning CutMix SimGCD
Closest papers by meaning
- Hyperbolic Category Discovery (CVPR 2025 · similarity 0.93)
- When Domain Generalization meets Generalized Category Discovery: An Adaptive Task-Arithmetic Driven Approach (CVPR 2025 · similarity 0.92)
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.91)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.91)
- DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (ICLR 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.