HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery

AI summary

Overview

• Research area: Computer vision — open-world recognition, specifically the combined problem of Domain Generalization and Generalized Category Discovery (DG-GCD), using hyperbolic (Poincaré ball) representation learning.
• Technical level: Advanced. The paper assumes familiarity with Riemannian geometry (Poincaré ball, tangent space, exponential/logarithmic maps, Busemann functions), contrastive learning, diffusion-based image synthesis, and Rademacher-complexity generalization bounds.
• Scope: The paper proposes HiDISC, a non-episodic hyperbolic framework for DG-GCD that trains only on labeled source data plus one to two GPT-4o-guided synthetic domains, and reports state-of-the-art clustering accuracy on PACS, Office-Home, and DomainNet.

What This Paper Is About

Standard Generalized Category Discovery (GCD) methods assume that labeled and unlabeled data come from the same domain and are available together during training. In the real world, a model often has to handle an unseen target domain that contains both familiar ("old") categories and entirely new ("novel") ones, with no target data available at training time — the setting the authors call Domain Generalization with GCD (DG-GCD). The only prior DG-GCD method, DG²CD-Net, handles this with computationally expensive episodic training over many synthetic domains and task-vector aggregation, which accumulates errors. HiDISC instead asks whether hyperbolic geometry can serve as a unified foundation for both domain shift and novel-class discovery.

Key Contributions

1. HiDISC, the first hyperbolic DG-GCD framework. It jointly handles domain shift and category shift without target-domain supervision and without episodic simulation.
2. A unified loss formulation combining (i) a penalized Busemann alignment loss that anchors seen classes to fixed prototypes on the Poincaré boundary, (ii) a hybrid hyperbolic contrastive loss that blends geodesic distance and angular similarity, and (iii) an adaptive outlier repulsion loss. A learnable curvature parameter adapts the geometry per dataset.
3. Tangent CutMix, described as the first open-set augmentation designed specifically for hyperbolic geometry — it interpolates labeled features in the tangent space at the origin and maps the result back to the Poincaré ball, producing geometrically valid pseudo-novel samples.
4. A lightweight synthetic domain strategy: a GPT-4o-guided diffusion pipeline generates only 1–2 synthetic domains per image, ranked by a FID-based domain-diversity score, replacing the many episodic synthetic domains used by DG²CD-Net.

Main Findings

• State-of-the-art accuracy across all three benchmarks. HiDISC reaches All/Old/New = 75.07/75.54/74.52 on PACS, 56.78/59.23/53.21 on Office-Home, and 30.51/31.40/28.41 on DomainNet, averaging 54.12/55.39/52.05.
• Improvements over the prior DG-GCD method. Against DG²CD-Net (9 synthetic domains), HiDISC improves average overall accuracy by +2.06%, known-class accuracy by +2.38%, and novel-class accuracy by +1.27%. Per-dataset gains are +1.00/+0.26/+0.57 on PACS, +2.92/+2.16/–1.12 on Office-Home (novel-class accuracy is the one metric that drops), and +1.21/+0.95/+2.04 on DomainNet. On DomainNet, the most diverse benchmark, HiDISC beats the best previous method by +1.21%.
• Hyperbolic variants of baselines do not match HiDISC. A hyperbolic re-implementation of DG²CD-Net (Hyp-DG²CD-Net) reaches only 48.59/48.74/48.76 average, below the Euclidean DG²CD-Net; Hyp-GCD reaches 46.12/46.74/46.13 and Hyp-SelfEx 51.55/52.60/50.18. Euclidean synthetic-augmentation baselines are far behind (e.g., SimGCD+6 Synth at 26.01/31.44/17.67).
• Much lower compute. HiDISC (reported in Figure 5 as "Hyp-Busemann") requires 16.53 PFLOPs over 50 epochs at batch size 128×2 — roughly 2× less than GCD (33.06 PFLOPs), ~12× less than GCD+Synthetic (198.36 PFLOPs), and ~96× less than DG²CD-Net (1,586 PFLOPs).
• Cleaner, more separated embeddings. On a Poincaré-disk UMAP of PACS target ("Photo") clusters with "Sketch" as source, silhouette scores are –0.52 (Hyp-GCD), –0.42 (Hyp-SelfEx), –0.29 (Hyp-DG²CD-Net), and –0.14 for HiDISC.
• Better cluster-count estimation. HiDISC estimates 7 clusters on PACS (ground truth 7), 66 on Office-Home (ground truth 65), and 351 on DomainNet (ground truth 345), closer to truth than DG²CD-Net (7/67/355) and CDAD-Net (DG) (12/60/362).
• Learnable curvature converges to dataset-specific values: 0.041 for Office-Home, 0.059 for PACS, and 0.38 for DomainNet.
• Loss ablations on Office-Home show complementarity. Vanilla training yields 26.17% overall; Busemann alone 56.32%; hybrid hyperbolic contrastive alone 50.95%; Busemann + contrastive 56.29%; Busemann + outlier 51.04%; the full combination 56.78/59.23/53.21.
• Component ablations on Office-Home. Manual augmentations 50.80%; no synthetic domain 56.07%; fixed curvature c=0.01 56.23%; fixed curvature c=0.03 55.67%; Euclidean CutMix 53.46% (a 3.96% drop versus full HiDISC); full HiDISC 56.78%.
• Hyperbolic model and hyperparameters. The Poincaré ball (56.78/59.23/53.21) outperforms the Lorentz model (54.28/56.01/51.41). For the Busemann slope, φ=0.10 gives 58.84/65.77/47.07, φ=0.75 gives 56.78/59.23/53.21, φ=0.90 gives 57.76/62.82/49.18; φ=0.75 is chosen for balance. For the ℓ2 radius before exponential mapping, 1.5 gives 56.78/59.23/53.21, 1.0 gives 57.33/61.14/51.76, and 2.3 gives 57.31/60.96/52.04.
• Loss-weight sensitivity. The chosen weights (λ1=0.60, λ2=0.25, λ3=0.15) give 56.78/59.23/53.21, while alternative configurations give 52.12/53.33/50.07 and 51.37/52.17/50.01.
• More synthetic domains are not better. The paper reports that excessive augmentation causes overfitting to seen classes: old-class accuracy rises while new-class performance slowly degrades with additional synthetic domains.
• Ablation on the hybrid similarity balance. The paper states that a balanced combination of angular and geodesic components achieves the highest accuracy, with Figure 3 reporting 56.78 for the combination versus 56.09 and 55.95 for the individual settings.
• Upper bound context. CDAD-Net under joint source-and-target access (the DA upper bound) reaches 73.69/78.82/68.66 average, far above all DG-GCD methods.
• Theoretical argument. The authors extend a Rademacher-based bound to the Poincaré ball, giving L_T(f) ≤ L_S′(f) + Δ_H(S′,T) + R_H(H) + ε, and argue that under the same augmentation budget the hyperbolic discrepancy Δ_H(S′,T) is smaller than the Euclidean one, Δ_E(S′,T), yielding a tighter bound.
• Not reported in the provided content: the raw dataset sizes (number of images) for PACS, Office-Home, and DomainNet are not stated; they are deferred to the supplementary material, which is not included here. Full hyperparameter tables and the formal proof are likewise deferred to the supplementary material.

Methodology in Plain English

The model is trained only on labeled source-domain images. To avoid the cost of simulating many domains episodically, the authors generate one or two extra synthetic domains per image using a diffusion model steered by prompts written by GPT-4o (for example, "underwater" or "night-time" versions of class instances). Candidate synthetic domains are ranked by a diversity score that averages FID between the source domain and the synthetic domain and between pairs of synthetic domains, which rewards both divergence from the source and complementarity between synthetic sets; only the top one or two are used.

Features come from a frozen DINO-pretrained ViT followed by a 3-layer MLP. The resulting Euclidean vector is pushed into the Poincaré ball through an exponential map at the origin, with the curvature treated as a learnable parameter rather than fixed.

To create fake "novel" samples, the method takes two embeddings with different labels, maps them into the tangent space at the origin, linearly mixes them with a uniform mixing weight, and maps the mixture back into the ball. Because this mixing happens in tangent space, the result stays geometrically valid — unlike ordinary Euclidean CutMix, which can violate hyperbolic constraints.

Three losses shape the space. The penalized Busemann loss pulls each feature directionally toward a fixed class prototype placed uniformly on the boundary of the Poincaré ball; a second term penalizes embeddings that crowd the boundary too aggressively. The hybrid contrastive loss compares augmented views of the same input, using a similarity that linearly combines negative hyperbolic distance with cosine similarity in tangent space. The outlier loss pushes the mixed pseudo-novel embeddings away from all known-class prototypes by at least an adaptive, quantile-based margin. The three are combined with weights that sum to one. At test time, hyperbolic features of target-domain samples are clustered with K-Means, and Hungarian matching aligns predicted clusters to ground-truth labels. Results are averaged over three runs and all source–target domain combinations.

Why This Matters

• Research impact: The paper opens a new combination of problems — hyperbolic geometry applied to DG-GCD — and reports that simply moving existing methods (Hyp-GCD, Hyp-SelfEx, Hyp-DG²CD-Net) into hyperbolic space is not enough; the geometric machinery has to be designed for the task. It also argues, with a generalization bound, that hyperbolic geometry can shrink the source–target discrepancy under a fixed augmentation budget, which makes expensive episodic domain simulation less necessary.
• Real-world applications:
  • Autonomous driving, where a perception model trained in one city or weather condition must recognize both known and previously unseen object categories in a new environment.
  • Medical diagnostics, where a model trained on one scanner, hospital, or population must handle novel pathology categories at a different site without target labels.
  • Wildlife monitoring or biodiversity surveys, where new species appear at deployment time and camera-trap imagery shifts across locations and seasons.
  • Industrial visual inspection, where defect types evolve and imaging conditions differ between production lines.
• Industry relevance: The claimed ~96× reduction in training FLOPs relative to DG²CD-Net and the removal of episodic training loops matter for teams that cannot afford large-scale domain-simulation pipelines. The requirement for only one or two synthetic domains also lowers the compute and prompt-engineering burden, although the dependence on a GPT-4o-guided diffusion model still implies access to generative infrastructure.
• Caveats flagged by the authors: The broader-impact statement notes that reliance on diffusion-generated augmentations may limit applicability in resource-constrained or safety-critical environments, where generative artifacts could propagate bias.

Future Directions

• Continual DG-GCD: extending HiDISC to streaming settings where domains and categories arrive over time, as the authors explicitly propose.
• Integration with large-scale vision-language models: the paper lists combining HiDISC with VLMs as a future direction.
• Reducing dependence on generative augmentation: given that training solely on the source domain already gives competitive results and that more synthetic domains degrade novel-class performance, finding ways to get diversity without diffusion artifacts — or with safer generators — is a natural follow-up.
• Closing the gap to the upper bound: CDAD-Net with joint source–target access averages 73.69/78.82/68.66 versus HiDISC's 54.12/55.39/52.05, so a large margin remains between the DG-GCD setting and the domain-adaptation upper bound.

Target Audience

Researchers and graduate students working on domain generalization, novel class discovery, open-set recognition, and geometric deep learning. It is also relevant to practitioners building recognition systems that must deploy into unseen environments with evolving label spaces, and to anyone specifically interested in whether hyperbolic embeddings help beyond the single-domain GCD setting where HypCD operates. Readers without a background in Riemannian geometry will need to consult the supplementary material for the manifold operations, the formal generalization proof, and the algorithmic details the main text defers.

Authors’ abstract

Generalized Category Discovery (GCD) aims to classify test-time samples into either seen categories** -- available during training -- or novel ones, without relying on label supervision. Most existing GCD methods assume simultaneous access to labeled and unlabeled data during training and arising from the same domain, limiting applicability in open-world scenarios involving distribution shifts. Domain Generalization with GCD (DG-GCD) lifts this constraint by requiring models to generalize to unseen domains containing novel categories, without accessing targetdomain data during training. The only prior DG-GCD method, DG2CD-Net, relies on episodic training with multiple synthetic domains and task vector aggregation, incurring high computational cost and error accumulation. We propose HIDISC, a hyperbolic representation learning framework that achieves domain and category-level generalization without episodic simulation. To expose the model to minimal but diverse domain variations, we augment the source domain using GPT-guided diffusion, avoiding overfitting while maintaining efficiency. To structure the representation space, we introduce Tangent CutMix, a curvature-aware interpolation that synthesizes pseudo-novel samples in tangent space, preserving manifold consistency. A unified loss -- combining penalized Busemann alignment, hybrid hyperbolic contrastive regularization, and adaptive outlier repulsion -- **facilitates compact, semantically structured embeddings. A learnable curvature parameter further adapts the geometry to dataset complexity. HIDISC achieves state-of-the-art results on PACS , Office-Home , and DomainNet, consistently outperforming the existing Euclidean and hyperbolic (DG)-GCD baselines.