acceptodds
Under review as a conference paper at ICLR 2027

NOMAD: Federated Generalized Category Discovery Across Unseen Domains

Abstract

Generalized Category Discovery (GCD) classifies known categories while discovering novel ones from unlabeled data. Practical deployments add two constraints rarely studied together: source data are distributed across institutions that cannot share raw samples, and the model must generalize to an unseen domain. Federated GCD rarely evaluates held-out-domain transfer, and domain-generalized GCD assumes centralized data. We formalize their intersection as **Federated Domain Generalization for Generalized Category Discovery (FedDG-GCD)**: decentralized clients train a model that must recognize known and discover novel categories in an unseen target domain without target-domain access. To our knowledge, this setting is unexplored, and no evaluation protocol exists for it. We propose **NOMAD** (**N**on-IID **O**pen-world **M**odel for **A**daptive **D**iscovery), combining a raw-data-free *Global Style Bank* for cross-client appearance diversification, *Adaptive Prototypical Alignment* for stable known-class anchors and confusable-prototype separation, and *perturbation-stability aggregation* to down-weight updates with unstable saliency under Style-Bank perturbations. A conditional target-risk decomposition relates these components to federated estimation error, residual style discrepancy, client heterogeneity, prototype error, and open-world routing and clustering. On PACS, Office-Home, and Mini-DomainNet under Dirichlet heterogeneity (), **NOMAD** improves overall accuracy by **+2.5 to +4.0 points** over the strongest compared method in each setting , while remaining competitive on both old and novel classes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generalized Category Discovery, Federated Learning, Domain Generalization, Novel Category Discovery, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1kkJOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FedDG-GCD, where decentralized clients train a model to recognize known categories and discover novel categories in an unseen target domain, without target-domain access during training. The proposed method, NOMAD, combines a Global Style Bank, adaptive prototype alignment, and saliency-stability-weighted aggregation, using the same style perturbations for local training and client weighting. At deployment, known samples are identified through prototype-based rejection, while rejected samples are clustered with an estimated category count. Experiments on 3 benchmarks verify its superioty.

Critical strengths
  1. Unexplored settings of FedDG-GCD. Decentralized data, unseen-domain deployment, and unknown categories can coexist in practice. The proposed protocol brings these constraints into a common evaluation framework.

  2. Coherent methodological design. Sharing a perturbation family between style diversification and client weighting provides a clear organizing principle. Prototype alignment also operates in the representation space used at inference.

Critical weaknesses
  1. The relationship between domain generalization, federated leanring, open world discovery and FedDG-GCD is not well discussed. I can understand the FedDG-GCD is a new setting, but I think its necessity should be more strongly supported. In this paper, it seems that it's just a intersection of above 3 settings. Firstly, the author should give some evidence that this setting is useful in real world. Secondly, I can not distinguish the unique challenges of FedDG-GCD compared with the above settings. In Table 1, it seems that the baselines like DG2CD-Net and HIDISC can also achive the same goals of FedDG-GCD with a slight performance drop even it does not consider this setting.

  2. Too many parameters included in the method may affect its practicality. NOMAD introduces many interacting hyperparameters, but how most are selected without target-domain access remains unclear. Sensitivity analysis is mainly limited to PACS and varies one parameter at a time, with some choices causing substantial performance drops. It therefore remains uncertain whether a fixed configuration can transfer reliably to new deployment conditions.

  3. Saliency stability is an insufficiently validated proxy for client usefulness. A model can attend consistently to an incorrect or uninformative region, so stable saliency does not necessarily imply transferable representations. Estimating a client’s weight from a fixed probe further risks dependence on its category and appearance, particularly under severe label skew. Moreover, min–max normalization can amplify small, noisy differences between clients. The aggregation comparisons demonstrate empirical utility, but do not establish that the score reliably measures contribution to unseen-domain recognition and discovery.

Questions for authors
  1. Which failure is specific to the joint FedDG-GCD setting rather than inherited from its component tasks? Can you provide controlled comparisons that isolate these interactions and compare shared versus independent perturbations for training and aggregation?

  2. Are the small novel-class improvements on Office-Home and Mini-DomainNet statistically reliable? Can you evaluate more class splits, novel-class proportions, and client counts, alongside wall-clock and communication costs?

Reviewer_V88XOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Federated Domain Generalization for Generalized Category Discovery, where decentralized clients train a model to recognize known classes and discover novel categories in an unseen domain. Its proposed method, NOMAD, combines a Global Style Bank for cross-client appearance augmentation, prototype alignment for known-class consistency and separation, and client aggregation based on representation stability under the same style perturbations. At inference, the model assigns confident samples to known classes and clusters rejected samples into novel categories. Validation on PACS, Office-Home, and Mini-DomainNet under two levels of client label heterogeneity reports overall accuracy gains of approximately 2.5–4.0 percentage points over the strongest compared methods, supported by ablations, sensitivity analyses, and a conditional target-risk decomposition.

Critical strengths
  1. The paper brings together three challenges that are usually studied separately—federated learning, category discovery, and generalization to unseen domains.
  2. Using the same style perturbations for local augmentation and client weighting gives NOMAD a clear rationale. It connects what clients are trained to withstand with how their updates are weighted. This makes the combination more compelling than simply adding separate augmentation and aggregation techniques.
Critical weaknesses
  1. The experiments do not directly isolate the paper’s central claim that augmentation and aggregation benefit from sharing the same perturbation family. Section 6 evaluates style augmentation and aggregation choices separately, but does not compare shared versus independently sampled perturbations while keeping the remaining components fixed. This matters because the proposed coupling is a main contribution, yet the reported gains could also come from combining individually useful components.
  2. The comparisons are difficult to interpret because several baselines require substantial changes to fit the proposed setting. Appendix E.1 explains these adaptations, including restricting DG$^2$CD-Net’s task-vector construction to within-client data and replacing whole-pool discovery with rejection followed by clustering. These choices are understandable, but can alter capabilities central to the original methods. Without a clearer assessment of their impact, it remains uncertain how much of NOMAD’s advantage reflects better representations versus better compatibility with the chosen protocol.
Questions for authors
  1. Could you clarify whether augmentation and aggregation reuse the exact same sampled perturbations or draw independently from the same Style Bank distribution? Do any existing ablations isolate the benefit of sharing this perturbation family while holding the local objective and aggregation rule fixed? This would help assess the central coupling claim.
  2. How sensitive are the baseline comparisons to the adaptations described in Appendix E.1, particularly DG$^2$CD-Net’s within-client task-vector construction and the common rejection-and-clustering protocol? Do you have existing checks showing that these adaptations preserve the baselines’ effectiveness? This would help establish how confidently the reported gains can be attributed to NOMAD.
  3. Equation (7) gives greater weight to clients with stable saliency maps under style perturbations. How do you distinguish useful semantic stability from consistently incorrect or uninformative responses? Does the existing analysis show whether these weights correlate with client usefulness, rather than primarily favoring clients with easier data or more common classes?
Reviewer_CiZsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces Federated Domain Generalization for Generalized Category Discovery (FedDG-GCD), which aims not only to recognize known classes but also to discover novel categories in unseen domains. To address this problem, the authors propose NOMAD, which leverages a global style bank to aggregate channel-wise feature statistics from multiple clients, thereby mitigating the influence of domain-specific features. Experimental results on the PACS, Office-Home, and Mini-DomainNet datasets demonstrate that the proposed method achieves superior performance compared to existing approaches.

Critical strengths
  • The proposed FedDG-GCD task is highly practical and relevant, as it addresses challenges that may arise in various real-world federated learning scenarios.
  • The paper provides detailed explanations of the proposed method and extensive additional analyses in the appendix.
Critical weaknesses
  • The main text frequently refers readers to the appendix for further explanations, which somewhat disrupts the overall flow and readability of the paper. The paper would benefit from further revision to make the main text as self-contained as possible, while reserving the appendix for supplementary details and additional analyses.
  • The global style bank may potentially harm domain generalization, as it tends to homogenize feature characteristics across different domains, potentially suppressing domain-specific diversity that could be beneficial for generalization.
  • The use of \citep{} and \citet{} is incorrect and should be corrected.
Questions for authors
  • In Tab. 1, the proposed method shows lower performance on the Old classes compared to that of PAL-GCD on the Office-Home and Mini-DomainNet datasets. However, the authors simply attribute this degradation to the known-novel balanced objective. If this performance trade-off is indeed caused by the balanced objective, additional experiments are needed to compare the New performance of the two methods when their Old performance is adjusted to comparable levels through hyperparameter tuning. Currently, the improvements in New performance appear relatively marginal compared to the degradation in Old performance. Therefore, a more detailed analysis is needed to explain why this phenomenon occurs on these datasets.
  • In Appendix H, the authors mention that the mixing strength parameter (ρ=2)(\rho=2) was empirically selected. However, to demonstrate the effectiveness of the adaptive mixing coefficient, it would be valuable to analyze how the average value of (qρ)(q^{\rho}) evolves during training and what value it converges to toward the end of training. Furthermore, the authors should compare the performance of the proposed adaptive strategy against a fixed mixing coefficient set to the observed converged value. Such an analysis would help clarify whether the adaptive mechanism provides meaningful benefits over a fixed coefficient.
  • L213–L215: Some notations are used before they are formally defined, which may cause confusion and hinder readability (e.g., those associated with Eqs. 7 and 9). The authors should ensure that all notations are clearly defined before their first use.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank you and final discussion update

We thank all three reviewers for the constructive discussion and valuable feedback, which has helped strengthen and clarify the paper.

Summary of Current Standing:

Reviewer 1kkJ (Score: 2 → 3): Confirmed that their concerns have been addressed and updated their score accordingly.

Reviewer V88X (Score: 3): Confirmed that their concerns are resolved and noted they are open to raising their score if the other reviewers' concerns are satisfied.

Reviewer CiZs (Score: 2): Confirmed that the majority of their initial concerns are resolved and indicated an intention to adjust their score accordingly.

All new experiments and clarifications from the discussion have been incorporated into the revised PDF. We remain available to address any final questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.