Pith. sign in

REVIEW 4 major objections 5 minor 53 references

FedGSCA: Medical Federated Learning with Global Sample Selector and Client Adaptive Adjuster under Label Noise

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read FedGSCA claims a federated learning framework that pools per-client Gaussian mixture selectors to handle heterogeneous label noise, and reports top F1 scores on three medical datasets.

desk verdict A useful, narrowly scoped noisy-FL method with broad experiments, but the server-side GMM aggregation needs an identifiability fix and code release before the gains are fully credible. read the letter →

arxiv 2507.10611 v1 pith:UBIWEO2G submitted 2025-07-13 cs.LG cs.AIcs.CV

classification cs.LGcs.AIcs.CV
keywords FederatedlearningLabelnoiseMedicalimageclassificationGaussianmixturemodelheterogeneityPseudo-labelingCredalsetsRobustloss
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the main obstacle to federated learning under label noise is not the noise itself but the fact that each hospital sees only its own noisy slice of the data, so local sample selection is unstable. It proposes FedGSCA, which makes each client fit a two-component Gaussian mixture model to its training losses and sends only the mixture parameters to the server; the server averages them into a Global Sample Selector that every client then uses to separate clean from noisy samples. On top of this, a Client Adaptive Adjustment mechanism creates class-specific pseudo-label thresholds and trains with a Robust Credal Labeling loss that hedges among several plausible labels rather than committing to a possibly wrong one. On three medical image datasets under symmetric, asymmetric, extreme, and heterogeneous noise, FedGSCA reports the highest F1 score in nearly every configuration, with the largest gains over the strongest baseline in heterogeneous noise cases.

What carries the argument

The load-bearing object is the Global Sample Selector (GSS), a server-side weighted average of each client's Gaussian Mixture Model parameters—two means, two variances, and two mixing weights describing the clean and noisy loss distributions. The server computes this average, broadcasts it back, and clients initialize their local selectors from it; a selector coefficient derived from the local loss dispersion decides how strict the clean/noisy cutoff is. The second mechanism is the Client Adaptive Adjustment (CAA), which uses class-adaptive thresholds to assign pseudo-labels to noisy samples and then trains with the Robust Credal Labeling (RCL) loss, a credal-set loss that projects the model's prediction onto the boundary of a set of plausible labels instead of a single target.

What would settle it

Run FedGSCA on a synthetic two-client setup with known clean/noisy splits and, before one aggregation round, permute the two GMM components of one client before upload; if the resulting global selector starts classifying that client's clean samples as noisy, or the F1 score drops sharply, the method depends on accidental component alignment, while no change would indicate the ordering concern is not fatal.

Watch

Extended reading notes

Core claim

On its own terms, FedGSCA claims that noise-handling knowledge can be pooled across privacy-separated clients by aggregating the parameters of per-client Gaussian mixture models, and that doing so stabilizes the global model more than filtering noise locally. The paper presents results on Kvasir-Capsule, OIA-ODIR, and a real-world colon slides dataset where FedGSCA beats FedAvg, FedProx, FedRobust, FedCorr, FedGP, FedFixer, and FedNoRo, and the largest margins appear under extreme symmetric noise and under heterogeneous noise types, e.g., 85.71 versus 81.20 F1 for the hardest mixed-type configuration on Kvasir-Capsule. The ablation studies attribute these gains to all three pieces: GSS, adaptive threshold pseudo-labeling, and the Robust Credal Labeling loss, with GSS removal causing the largest drop on Kvasir-Capsule.

Load-bearing premise

The server-side average in Equation (4) assumes that when it averages each client's two Gaussian components, the first component always means clean and the second always means noisy across every client; the paper never states how this ordering is enforced, and a swapped component would corrupt the shared clean/noisy selector.

Editorial extensions

If this is right

  • On Kvasir-Capsule with heterogeneous noise types, FedGSCA reaches 85.71 F1 for the 0-S40%-P20%-C20% configuration, outperforming FedGP by 4.51 points.
  • On OIA-ODIR under 0-20%-20%-40% symmetric noise, FedGSCA achieves 85.16 F1, the best reported in that comparison.
  • On the real-world Chaoyang dataset, FedGSCA achieves F1 76.92, beating FedGP's 75.98 and reducing confusion for the Adenocarcinoma and Adenoma classes.
  • The Global Sample Selector reduces training instability, measured as the average proximal distance between local and global model weights.
  • FedGSCA reaches FedAvg's peak F1 in about 96.78 minutes versus FedAvg's 178.34 minutes, despite a higher per-round time cost.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The GSS aggregation assumes a fixed ordering of GMM components so that the first component always means clean and the second means noisy across every client; the paper never states how this ordering is enforced, so ordering by mean loss or constrained initialization is a natural extension to test.
  • The same global-selector idea could be applied to per-client confidence distributions rather than losses, which might help under noise types that do not inflate loss, but this is not evaluated in the paper.
  • The Robust Credal Labeling loss may also help with genuine label ambiguity rather than only synthetic noise, which could be tested on datasets with expert disagreement, though the paper only covers noisy-label benchmarks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes FedGSCA, a federated learning framework for medical image classification under label noise. It combines a Global Sample Selector (GSS), which aggregates per-client Gaussian mixture model parameters over sample losses to form a global clean/noisy selector, with a Client Adaptive Adjuster (CAA) that applies class-adaptive pseudo-labeling and a Robust Credal Labeling (RCL) loss. The method is evaluated on Kvasir-Capsule, OIA-ODIR, and a real-world Chaoyang colon-slides dataset under symmetric, pair-flip, clinical-knowledge-based asymmetric, and heterogeneous noise settings, and is compared with seven FL baselines. The central claim is that FedGSCA outperforms these baselines, especially under extreme symmetric and heterogeneous noise, and that both GSS and CAA contribute to the gains.

Significance. The label-noise and class-imbalance problems addressed here are relevant for practical medical FL, and the paper's empirical scope is a genuine strength: three datasets, multiple noise types including heterogeneous rate/type configurations, repeated trials with reported standard deviations, and ablations isolating RCL, ATP, and GSS. The reported gains, e.g., about one F1 point over FedGP under 50% symmetric noise on Kvasir-Capsule and up to 4.5 F1 points under heterogeneous-type noise, are plausible if the mechanism works as described. However, the central novelty, the global aggregation of client GMM parameters, rests on an unspecified component-alignment assumption, and the absence of code or a validation-based hyperparameter selection protocol weakens the reproducibility of the numerical claims. The paper is not a derivation paper, so no mathematical circularity is alleged, but the self-training loop through pseudo-labels deserves the same scrutiny as sample-selection methods in the centralized literature.

major comments (4)
  1. [Section III-B, Eqs. (4) and (7)] Equation (4) aggregates each client's GMM parameters component-wise, and Eq. (7) treats component 1 as 'clean' by computing the posterior P(z=1|x,y). GMM components are identifiable only up to permutation, so unless an explicit alignment rule is imposed, a weighted average of parameters can mix the clean component of one client with the noisy component of another. Since the aggregated S(t+1) is broadcast to all clients and used through Eq. (7) to construct Dc_k and then Dpseudo_k in Eq. (12), a component swap on any client would propagate incorrect clean/noisy assignments into the pseudo-labels and the final model. Initializing each CSS from the current GSS, as stated in Section III-B1, may encourage consistent ordering, but S(0) is not specified and EM can still swap components when local loss distributions are multimodal or noise is extreme. The paper should specify a canonical ordering rule (e.g., ordering components by mean loss) and provide a diagnostic that tracks component identity across clients and rounds, since this assumption is load-bearing for the main novelty and for the claimed performance in extreme and heterogeneous noise scenarios.
  2. [Section IV-A4 and Section IV-C3, Table XI] The hyperparameters ζ0 and β1 are selected using the same benchmark configuration (0-S10%-P10%-C10%) that also appears in the main comparison tables, and no separate validation split is described. Table XI reports the resulting optimal values and the sensitivity around them, but the selection is performed on the same datasets and configurations used to report final results. Without code or a held-out validation protocol, the reported margins over state-of-the-art methods may partly reflect selection on the test benchmarks. Please report hyperparameter selection on a validation split or over multiple seeds/configurations, and make the code available, so that the empirical comparisons are reproducible.
  3. [Section V and Algorithm 1] The stability and efficiency claims are not fully supported by the reported settings. Algorithm 1 and Section IV-A4 specify T=100 training rounds, but Fig. 5 plots 'Training Rounds' up to 500 for the stability metric. In addition, Section V states that FedGSCA reaches FedAvg's peak performance level with F1=77.78, yet Table IV reports FedAvg's F1 as 56.81 at 50% symmetric noise and FedGSCA's own F1 as 77.39 in the same setting; the 77.78 value appears to be FedGSCA's recall, not FedAvg's peak, and the experimental configuration is unclear. Please reconcile the number of rounds and identify exactly which configuration and metric support the efficiency comparison.
  4. [Section III-D1, Eq. (14)] Equation (14) is applied to all samples in the updated training set \(\hat{D}_k\), which includes the clean subset Dc_k, but the text motivates RCL as handling noisy samples. Under Eq. (14), any class whose predicted probability is at least \(\beta\) receives full possibility \(\pi_i(y)=1\), even when the observed label is clean. This means the credal set can hedge away a known clean label, and the projection in Eq. (16) will then redistribute probability mass away from the true label. The paper does not quantify how much clean-label information is discarded, nor does any ablation isolate RCL's effect on clean versus noisy subsets. Please clarify whether the clean subset is intentionally treated with the same relaxed loss, or add an experiment that separates RCL applied only to noisy samples from RCL applied to all samples.
minor comments (5)
  1. [Abstract and Contributions] The framework is called FedGSCA in the title and abstract, but the contribution bullet at the end of Section I and the implementation details in Section IV-A4 refer to it as FedCAGS; please use the name consistently throughout.
  2. [Table XI] The caption of Table XI says the hyperparameter analysis is conducted on the OIA-ODIR dataset, while the surrounding text says it is on the Kvasir-Capsule dataset under 0-S10%-P10%-C10% noise; please correct the mismatch.
  3. [Section III-B2, Eq. (7)] Equation (7) uses the global model parameters \(\theta^{(t)}\), but the text above it says the client downloads the updated global model \(\theta^{(t+1)}\) and GSS at the start of round t+1; please make the time indices consistent.
  4. [Algorithm 1, Line 7] The server aggregation step writes 'Update S with {S_k | k in M} using (4)', but Eq. (4) uses symbols \(\mu_k\), \(\sigma_k\), and \(\pi_k\) without round superscripts; please align the notation, e.g., \(\S_k^{(t)}\).
  5. [References] Reference [50], the Chaoyang dataset, includes a DOI '10.1038/s41597-021-00920-z' that appears to belong to a different article; please verify the citation metadata.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the performance claims are supported by external baselines and ablations, and the self-citations to prior work are non-load-bearing design/test details.

full rationale

This is an empirical method paper rather than a formal derivation, so the headline result—FedGSCA outperforming baselines on medical FL benchmarks—is established by experiments against FedAvg, FedProx, FedCorr, FedGP, FedFixer, FedNoRo, and others, not by construction from its own equations. The GMM sample selector uses model losses to split clean and noisy samples, and the pseudo-labeling loop is standard self-training practice rather than a case where a predicted quantity is defined by the fitted input. The self-citations to the authors' previous work [46] supply only the CK-Asymm noise generation recipe and the cosine decay schedule for beta; these are experimental inputs and hyperparameter details, not evidence for the central claim, so they are not load-bearing. The reviewer's concern about GMM component permutation in Eq. (4) is a real identifiability and robustness limitation, but it is not circularity: no evaluation metric is defined in terms of the selector parameters, and no reported result is forced by the aggregation formula. The paper is self-contained enough that its empirical comparisons and ablations stand independently of any self-citation chain.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several hand-chosen thresholds and on an unproven construction (global averaging of GMM parameters). No new physical entities, particles, forces, or data sources are introduced; the method is a composition of existing learning components.

free parameters (5)
  • ζ0 (adaptive pseudo-label initial threshold) = 0.8
    Set in Section IV-A4 and tuned in Table XI; values 0.7 and 0.9 reduce F1 by roughly 0.04 on the benchmark.
  • β0, β1 (cosine decay endpoints for RCL possibility threshold) = 0.75, 0.55
    Chosen from Table XI; β1=0.35 degrades F1 to 86.24 on OIA-ODIR, showing sensitivity to this choice.
  • α (label relaxation parameter) = 0.05
    Fixed in the implementation details; no sensitivity analysis is reported for this parameter.
  • τ_k coefficient 0.5 and cap 0.8 (selector threshold formula) = 0.5, 0.8
    Eq (5); described as empirical to prevent overly strict selection, but the values are not swept or justified beyond intuition.
  • δ noise-level split threshold = 0.1
    Step 3 of Algorithm 1; decides whether pseudo-labeling is applied; no sensitivity analysis is provided.
assumptions (4)
  • domain assumption Loss values of training samples follow a two-component Gaussian mixture separating clean and noisy samples
    Eq (1) in Section III-B1; the entire GSS rests on this fit, inherited from the small-loss heuristic of [38].
  • domain assumption Noisy samples tend to have larger loss than clean samples
    Stated around Eq (5) and in Section III-B1; used to justify the GMM and the dispersion-based threshold τ_k.
  • ad hoc to paper A weighted average of client GMM parameters is a meaningful global noise selector
    Eq (4) in Section III-B2; no component-alignment or permutation-symmetry handling is specified, so this is an unproven construction.
  • domain assumption The credal-set projection in Eqs (14)-(16) hedges label noise better than hard labels
    Taken from the authors' previous SSP-RACL work [46]; effects are only demonstrated empirically, without theoretical support.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedGSCA: Medical Federated Learning with Global Sample Selector and Client Adaptive Adjuster under Label Noise." pith.science (2026). https://pith.science/paper/UBIWEO2G

@misc{pith2026250710611,
  author       = {Pith},
  title        = {Pith review of: FedGSCA: Medical Federated Learning with Global Sample Selector and Client Adaptive Adjuster under Label Noise},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UBIWEO2G}},
  note         = {Machine review of arXiv:2507.10611}
}
read the original abstract

Federated Learning (FL) emerged as a solution for collaborative medical image classification while preserving data privacy. However, label noise, which arises from inter-institutional data variability, can cause training instability and degrade model performance. Existing FL methods struggle with noise heterogeneity and the imbalance in medical data. Motivated by these challenges, we propose FedGSCA, a novel framework for enhancing robustness in noisy medical FL. FedGSCA introduces a Global Sample Selector that aggregates noise knowledge from all clients, effectively addressing noise heterogeneity and improving global model stability. Furthermore, we develop a Client Adaptive Adjustment (CAA) mechanism that combines adaptive threshold pseudo-label generation and Robust Credal Labeling Loss. CAA dynamically adjusts to class distributions, ensuring the inclusion of minority samples and carefully managing noisy labels by considering multiple plausible labels. This dual approach mitigates the impact of noisy data and prevents overfitting during local training, which improves the generalizability of the model. We evaluate FedGSCA on one real-world colon slides dataset and two synthetic medical datasets under various noise conditions, including symmetric, asymmetric, extreme, and heterogeneous types. The results show that FedGSCA outperforms the state-of-the-art methods, excelling in extreme and heterogeneous noise scenarios. Moreover, FedGSCA demonstrates significant advantages in improving model stability and handling complex noise, making it well-suited for real-world medical federated learning scenarios.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 46 canonical work pages

  1. [1]

    Vucinich and Q

    S. Vucinich and Q. Zhu, ”The Current State and Challenges of Fairness in Federated Learning,” IEEE Access , vol. 11, pp. 80903-80914, 2023

  2. [2]

    Lancelot: Towards Efficient and Privacy-Preserving Byzantine-Robust Federated Learning within Fully Homomorphic Encryption

    S. Jiang, H. Y ang, Q. Xie, C. Ma, S. Wang, and G. Xing, ”Lanc elot: Towards Efficient and Privacy-Preserving Byzantine-Robus t Feder- ated Learning within Fully Homomorphic Encryption,” arXiv preprint arXiv:2408.06197, 2024

  3. [3]

    Z. Sun, P . Kairouz, A. T. Suresh, and H. B. McMahan, ”Can yo u really backdoor federated learning?,” arXiv preprint arXiv:1911.07963, 2019

  4. [4]

    Hard sample a ware noise robust learning for histopathology image classificat ion,

    C. Zhu, W. Chen, T. Peng, Y . Wang, and M. Jin, “Hard sample a ware noise robust learning for histopathology image classificat ion,” IEEE Trans. Med. Imaging , vol. 41, no. 4, pp. 881–894, Apr. 2021

  5. [5]

    Mendieta, T

    M. Mendieta, T. Y ang, P . Wang, M. Lee, Z. Ding, and C. Chen, ”Local learning matters: Rethinking data heterogeneity in federa ted learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision a nd Pattern Recognition, pp. 8397-8406, 2022

  6. [6]

    Communication-efficient learning of deep networks from de central- ized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arc as, “Communication-efficient learning of deep networks from de central- ized data,” in Proc. Artificial Intelligence and Statistics (AISTATS) , pp. 1273–1282, 2017

  7. [7]

    Federated learning with extremely noisy clients via negative distillation,

    Y . Lu, et al , “Federated learning with extremely noisy clients via negative distillation,” in Proc. AAAI Conf. Artificial Intelligence , vol. 38, no. 13, pp. 14184–14192, 2024

  8. [8]

    Tackling noisy clients in federated learning with end- to-end label correction,

    X. Jiang, et al , “Tackling noisy clients in federated learning with end- to-end label correction,” arXiv preprint arXiv:2408.04301 , 2024. YE et al.: FEDGSCA: MEDICAL FEDERA TED LEARNING WITH GLOBAL SAMPLE SE LECTOR AND CLIENT ADAPTIVE ADJUSTER UNDER LABEL NOISE 11

Show all 53 references
  1. [9]

    FedDiv: Collabo rative noise filtering for federated learning with noisy labels,

    J. Li, G. Li, H. Cheng, Z. Liao, and Y . Y u, “FedDiv: Collabo rative noise filtering for federated learning with noisy labels,” i n Proc. AAAI Conf. Artificial Intelligence , vol. 38, no. 4, pp. 3118–3126, 2024

  2. [10]

    A systematic stu dy of the class imbalance problem in convolutional neural networ ks,

    M. Buda, A. Maki, and M. A. Mazurowski, “A systematic stu dy of the class imbalance problem in convolutional neural networ ks,” Neural Networks, vol. 106, pp. 249–259, 2018

  3. [11]

    M. Li, Q. Li, and Y . Wang, ”Class balanced adaptive pseud o labeling for federated semi-supervised learning,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit , pp. 16292–16301, 2023

  4. [12]

    Understand- ing deep learning (still) requires rethinking generalizat ion,

    C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals, “ Understand- ing deep learning (still) requires rethinking generalizat ion,” Commun. ACM, vol. 64, no. 3, pp. 107–115, 2021

  5. [13]

    On the robustnes s of decision tree learning under label noise,

    A. Ghosh, N. Manwani, and P . S. Sastry, “On the robustnes s of decision tree learning under label noise,” in Adv. Knowl. Discov. Data Min.: Proc. 21st Pacific-Asia Conf., PAKDD , Jeju, South Korea, pp. 685– 697, 2017

  6. [14]

    Loss fa ctorization, weakly supervised learning and label noise robustness,

    G. Patrini, F. Nielsen, R. Nock, and M. Carioni, “Loss fa ctorization, weakly supervised learning and label noise robustness,” in Proc. Int. Conf. Machine Learning (ICML) , pp. 708–717, 2016

  7. [15]

    A survey of label-noise representation learning: Past, pres ent and future,

    B. Han, Q. Y ao, T. Liu, G. Niu, I. W. Tsang, J. T. Kwok, et al. , “A survey of label-noise representation learning: Past, pres ent and future,” arXiv preprint arXiv:2011.04406 , 2020

  8. [16]

    Learnin g from noisy labels with deep neural networks: A survey,

    H. Song, M. Kim, D. Park, Y . Shin, and J.-G. Lee, “Learnin g from noisy labels with deep neural networks: A survey,” IEEE Trans. Neural Netw. Learn. Syst. , vol. 34, no. 11, pp. 8135–8153, 2022

  9. [17]

    Weakly supervised learning with side information for noisy labeled images,

    L. Cheng, et al., “Weakly supervised learning with side information for noisy labeled images,” in Proc. 16th Eur . Conf. Comput. Vis. (ECCV) , ,vol. 16, pp. 306–321, 2020

  10. [18]

    Lea rning with noisy labels via sparse regularization,

    X. Zhou, X. Liu, C. Wang, D. Zhai, J. Jiang, and X. Ji, “Lea rning with noisy labels via sparse regularization,” in Proc. IEEE/CVF Int. Conf. Computer Vision (ICCV) , pp. 72–81, 2021

  11. [19]

    Fine- grained classification with noisy labels,

    Q. Wei, L. Feng, H. Sun, R. Wang, C. Guo, and Y . Yin, “Fine- grained classification with noisy labels,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR) , pp. 11651–11660, 2023

  12. [20]

    Meta label cor rection for noisy label learning,

    G. Zheng, A. H. Awadallah, and S. Dumais, “Meta label cor rection for noisy label learning,” in Proc. AAAI Conf. Artif. Intell. , vol. 35, no. 12, pp. 11053–11061, 2021

  13. [21]

    NoiseBox: To wards more efficient and effective learning with noisy labels,

    C. Feng, G. Tzimiropoulos, and I. Patras, “NoiseBox: To wards more efficient and effective learning with noisy labels,” IEEE Trans. Circuits Syst. Video Technol., 2024

  14. [22]

    Robust classification from noisy labels: Integrating additional knowledge for chest radiography abnormality as sessment,

    S. G¨ undel, et al. , “Robust classification from noisy labels: Integrating additional knowledge for chest radiography abnormality as sessment,” Med. Image Anal. , vol. 72, Art. no. 102087, 2021

  15. [23]

    Improving medical images classification with label noise using dual-uncertainty estimation,

    L. Ju, et al. , “Improving medical images classification with label noise using dual-uncertainty estimation,” IEEE Trans. Med. Imaging , vol. 41, no. 6, pp. 1533–1546, 2022

  16. [24]

    A fundus image classification framework for learning with noisy labels,

    T. Hu, et al. , “A fundus image classification framework for learning with noisy labels,” Comput. Med. Imaging Graph. , vol. 108, Art. no. 102278, 2023

  17. [25]

    Robust stocha stic neural ensemble learning with noisy labels for thoracic disease cl assification,

    H. Wang, J. He, H. Cui, B. Y uan, and Y . Xia, “Robust stocha stic neural ensemble learning with noisy labels for thoracic disease cl assification,” IEEE Trans. Med. Imaging , 2024

  18. [26]

    Federated optimization in heterogeneous networks ,

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, an d V . Smith, “Federated optimization in heterogeneous networks ,” Proc. Mach. Learn. Syst. , vol. 2, pp. 429–450, 2020

  19. [27]

    Robust federated learning: The case of affine distribution shifts,

    A. Reisizadeh, F. Farnia, R. Pedarsani, and A. Jadbabai e, “Robust federated learning: The case of affine distribution shifts, ” Adv. Neural Inf. Process. Syst. , vol. 33, pp. 21554–21565, 2020

  20. [28]

    Fed-DR-Filte r: Using global data representation to reduce the impact of noisy lab els on the performance of federated learning,

    S. Duan, C. Liu, Z. Cao, X. Jin, and P . Han, “Fed-DR-Filte r: Using global data representation to reduce the impact of noisy lab els on the performance of federated learning,” Future Gener . Comput. Syst. , vol. 137, pp. 336–348, 2022

  21. [29]

    Fedcorr: Multi- stage federated learning for label noise correction,

    J. Xu, Z. Chen, T. Q. S. Quek, and K. F. E. Chong, “Fedcorr: Multi- stage federated learning for label noise correction,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR) , pp. 10184– 10193, 2022

  22. [30]

    Robust federated learning with noisy and heterogeneous clients,

    X. Fang and M. Y e, “Robust federated learning with noisy and heterogeneous clients,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , pp. 10072–10081, 2022

  23. [31]

    Towards federated learning against noisy labels via local self-regularization,

    X. Jiang, S. Sun, Y . Wang, and M. Liu, “Towards federated learning against noisy labels via local self-regularization,” in Proc. 31st ACM Int. Conf. Information & Knowledge Management (CIKM) , pp. 862– 873, 2022

  24. [32]

    Curriculum- Based Federated Learning for Machine Fault Diagnosis With Noisy L abels,

    W. Sun, R. Y an, R. Jin, R. Zhao, and Z. Chen, “Curriculum- Based Federated Learning for Machine Fault Diagnosis With Noisy L abels,” IEEE Trans. Ind. Inform. , 2024

  25. [33]

    Federated data quality assessment approach: robust learn ing with mixed label noise,

    B. Zeng, X. Y ang, Y . Chen, H. Y u, C. Hu, and Y . Zhang, “Federated data quality assessment approach: robust learn ing with mixed label noise,” IEEE Trans. Neural Netw. Learn. Syst. , doi: 10.1109/TNNLS.2023.3306874, 2023

  26. [34]

    F ed- Noisy: Federated noisy label learning benchmark,

    S. Liang, J. Huang, J. Hong, D. Zeng, J. Zhou, and Z. Xu, “F ed- Noisy: Federated noisy label learning benchmark,” arXiv preprint arXiv:2306.11650, 2023

  27. [35]

    Fednoil: A simple two-level sampling method for federated learning with nois y labels,

    Z. Wang, T. Zhou, G. Long, B. Han, and J. Jiang, “Fednoil: A simple two-level sampling method for federated learning with nois y labels,” arXiv preprint arXiv:2205.10110 , 2022

  28. [36]

    Medical federated l earning with joint graph purification for noisy label learning,

    Z. Chen, W. Li, X. Xing, and Y . Y uan, “Medical federated l earning with joint graph purification for noisy label learning,” Med. Image Anal., vol. 90, Art. no. 102976, 2023

  29. [37]

    Intelligent hand ling of noise in federated learning with co-training for enhanced diagnost ic precision,

    F. F. Babar, F. Jamil, and F. F. Babar, “Intelligent hand ling of noise in federated learning with co-training for enhanced diagnost ic precision,” in Proc. Int. Conf. Computational Collective Intelligence , pp. 279–291, 2024

  30. [38]

    Improving speaker verifi cation with noise-aware label ensembling and sample selection: Le arning and correcting noisy speaker labels,

    Z. Fang, L. He, L. Li, and Y . Hu, “Improving speaker verifi cation with noise-aware label ensembling and sample selection: Le arning and correcting noisy speaker labels,” IEEE/ACM Trans. Audio, Speech, Lang. Process. , vol. 32, pp. 2988–3001, 2024

  31. [39]

    Permuter, J

    H. Permuter, J. Francos, and I. Jermyn, ”A study of Gauss ian mixture models of color and texture features for image classificatio n and segmentation,” Pattern Recognit., vol. 39, no. 4, pp. 695–706, 2006

  32. [40]

    Z. Wang, J. Jiang, B. Han, et al., ”Seminll: A framework o f noisy-label learning by semi-supervised learning,” arXiv preprint arXiv:2012.00925, 2020

  33. [41]

    N. Pu, W. Li, X. Ji, et al., ”Federated generalized categ ory discovery,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. , pp. 28741– 28750, 2024

  34. [42]

    Lienen, C

    J. Lienen, C. Demir, and E. H¨ ullermeier, ”Conformal cr edal self- supervised learning,” in Proc. Conformal and Probabilistic Prediction with Applications , pp. 214–233, 2023

  35. [43]

    Lienen and E

    J. Lienen and E. H¨ ullermeier, ”Mitigating label noise through data ambiguation,” in Proc. AAAI Conf. Artif. Intell. , vol. 38, no. 12, pp. 13799–13807, 2024

  36. [44]

    K. Wang, K. Shariatmadar, S. K. Manchingal, F. Cuzzolin , D. Moens, and H. Hallez, ”CreINNs: Credal-Set Interval Neural Networks for Uncertainty Estimation in Classification Tasks,” arXiv preprint arXiv:2401.05043, 2024

  37. [45]

    Possibility theory, probabili ty theory and multiple-valued logics: A clarification,

    D. Dubois and H. Prade, “Possibility theory, probabili ty theory and multiple-valued logics: A clarification,” Annals of Mathematics and Artificial Intelligence, vol. 32, pp. 35–66, 2001

  38. [46]

    SSP-RACL: Classifica tion of noisy fundus images with self-supervised pretraining an d robust adaptive credal loss,

    M. Y e, Y . Huangfu, Y . Li, and Z. Y u, “SSP-RACL: Classifica tion of noisy fundus images with self-supervised pretraining an d robust adaptive credal loss,” arXiv preprint arXiv:2409.18147 , 2024

  39. [47]

    H¨ ullermeier, ”Credal Learning Theory,” arXiv preprint arXiv:2402.00957, 2024

    Michele Caprio, Maryam Sultana, Eleni Elia, Fabio Cuzz olin. H¨ ullermeier, ”Credal Learning Theory,” arXiv preprint arXiv:2402.00957, 2024

  40. [48]

    Kvasir-Capsule, a video capsule endoscopy dataset,

    P . H. Smedsrud, et al. , “Kvasir-Capsule, a video capsule endoscopy dataset,” Scientific Data , vol. 8, no. 1, p. 142, 2021

  41. [49]

    A benchmark of oc ular disease intelligent recognition: One shot for multi-disea se detection,

    N. Li, T. Li, C. Hu, K. Wang, and H. Kang, “A benchmark of oc ular disease intelligent recognition: One shot for multi-disea se detection,” in Benchmarking, Measuring, and Optimizing , F. Wolf and W. Gao, Eds., Bench 2020, Lecture Notes in Computer Science, vol. 12 614, Cham: ...

  42. [50]

    Hard sample aware noise robust learning for histopathology image class ification,

    C. Zhu, W. Chen, T. Peng, Y . Wang, and M. Jin, “Hard sample aware noise robust learning for histopathology image class ification,” IEEE Trans. Med. Imaging , vol. 41, no. 4, pp. 881–894, 2021, doi: 10.1038/s41597-021-00920-z

  43. [51]

    L VM-Med: Learning large-scale self-supervised vision models for medical imaging via second-order graph ma tching,

    M. H. Nguyen, et al., “L VM-Med: Learning large-scale self-supervised vision models for medical imaging via second-order graph ma tching,” Adv. Neural Inf. Process. Syst. , vol. 36, 2024

  44. [52]

    Ji, et al

    X. Ji, et al. , ”FedFixer: Mitigating Heterogeneous Label Noise in Federated Learning,” in Proc. AAAI Conf. Artif. Intell. , vol. 38, no. 11, pp. 12830-12838, 2024

  45. [53]

    FedNoRo: Towards noise-robust fe derated learning by addressing class imbalance and label noise hete rogeneity,

    Wu, Nannan and Y u, Li and Jiang, Xuefeng and Cheng, Kwang - Ting and Y an, Zengqiang, “FedNoRo: Towards noise-robust fe derated learning by addressing class imbalance and label noise hete rogeneity,” arXiv preprint arXiv:2305.05230 , 2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.