Pith. sign in

REVIEW 2 major objections 5 minor 22 references

FedProIn treats class prototypes as trainable parameters and stops client drift by separating feature drift from prototype drift, keeping federated medical accuracy nearly as high under non-IID data as under IID.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 21:16 UTC pith:FQXTKOYW

load-bearing objection Solid incremental FL methods paper: learnable multi-prototypes + FDL/PCL + NIA give measurable non-IID gains on two public medical sets, with ablation and code; not foundational, but clean enough to engage. the 2 major comments →

arxiv 2607.04158 v1 pith:FQXTKOYW submitted 2026-07-05 eess.IV cs.CV

FedProIn: Mitigating Client Drift for Learnable Prototypes in Federated Medical Imaging

Harsh Kumar , Tarun Kumar Garg , Vaanathi Sundaresan This is my paper
classification eess.IV cs.CV
keywords federated learningclient driftlearnable prototypesmedical imagingnon-IID dataprototype aggregationfeature divergence lossnormalized influence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Medical images from different hospitals differ in scanners, protocols and disease mix, so ordinary federated averaging lets each hospital’s model wander away from a common representation. FedProIn replaces the usual single mean-vector prototype with several learnable prototypes per class and trains them jointly with the feature encoder. Client drift is explicitly split into two pieces: feature drift (local embeddings leaving the global embedding space) and prototype drift (local prototypes being pulled toward hospital-specific data). Two regularizers—feature-divergence loss that anchors local features to a frozen global encoder, and prototype-contrastive loss that pulls current prototypes toward the server’s prototypes while pushing them away from the client’s previous ones—keep both pieces in check. On the server the prototypes are not averaged by sample count; they are weighted by how often each prototype was actually used as the nearest neighbour during local training (normalized influence aggregation). On two long-tailed medical benchmarks the resulting global model loses only a few points of accuracy when data become non-IID, while standard federated and prototype methods lose substantially more.

Core claim

Client drift in federated medical imaging can be decomposed into feature drift and prototype drift and mitigated by jointly optimizing multiple learnable class prototypes with a feature-divergence loss, a prototype-contrastive loss, and a usage-weighted (normalized-influence) aggregation rule, yielding global models whose accuracy remains close under non-IID partitions to their IID performance.

What carries the argument

Normalized Influence Aggregation (NIA): each client records how many times each of its learnable prototypes served as nearest neighbour; the server forms the global prototype as a weighted average of the local prototypes using those counts as weights, automatically down-weighting unused or noisy prototypes while preserving rare but useful ones.

Load-bearing premise

The previous-round global encoder and global prototypes are always a better, more general reference than the client’s own current features and prototypes, so pulling toward them is helpful rather than harmful under severe local shift.

What would settle it

On a medical dataset with extreme label or feature shift, replace the frozen global-encoder reference or the global-prototype attractor with a neutral or local-only reference and check whether accuracy under non-IID conditions still stays within a few points of the IID result; a large drop would falsify the claim that the two regularizers reliably correct client drift.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes FedProIn, a federated learning framework for medical imaging that treats multiple class prototypes as learnable parameters optimized jointly with a feature encoder. Client drift is decomposed into feature drift and prototype drift, addressed by a feature-divergence loss (FDL, Eq. 2) that aligns local embeddings to a frozen previous-round global encoder and a prototype-contrastive loss (PCL, Eq. 3) that pulls local prototypes toward global ones while pushing them away from the client's previous prototypes. Server-side aggregation uses FedAvg for the encoder and a Normalized Influence Aggregation (NIA, Eq. 7) that weights each prototype by its nearest-neighbor usage counts (influence matrix I_i, Eq. 6). On HAM10000 and Matek-19 under IID and Dirichlet-α=0.5 non-IID partitions (K=10 clients, ResNet-18), FedProIn reports higher accuracy, weighted F1 and MCC than FedAvg, FedProx, MOON, BalanceFL, FedProto, FedTGP, FedProc and FedPLVM (Table 1), with ablations (Fig. 2) showing statistically significant drops when FDL, PCL or NIA are removed. Code is released.

Significance. If the reported gains hold, FedProIn is a useful practical contribution for non-IID medical FL: it combines learnable multi-prototypes (extending CPL-style ideas into FL), an explicit drift decomposition with two regularizers, and an influence-aware aggregation that is better suited to multi-prototype models than plain FedAvg. Strengths include public datasets, a public code link, three-run means±std, and a Mann–Whitney ablation that supports the necessity of each component. The work is incremental relative to existing prototype FL methods but addresses a real medical-imaging pain point (scanner/protocol/population shift and long-tailed classes) with a clear, implementable design.

major comments (2)
  1. §2.1, Eq. (3) (PCL) and the surrounding claim that global prototypes are always a more generalized reference: under severe label or feature shift the previous-round global prototypes can themselves be biased, so the contrastive pull may be harmful rather than helpful. The paper should either (i) provide a controlled stress test with stronger heterogeneity (e.g., α≤0.1 or pathological label skew) showing when PCL remains beneficial, or (ii) clearly scope the claim to the evaluated regime (K=10, α=0.5) and discuss failure modes. Ablation on HAM10000 non-IID already shows PCL helps there, so this is a generalizability caveat rather than an internal contradiction, but it is load-bearing for the method's stated motivation.
  2. Table 1 and §3: several strong FL baselines that specifically target client drift (SCAFFOLD, FedDyn) are cited in the references but not compared. Given that the central claim is superior mitigation of client drift, at least one of these should be included (or a short justification given for exclusion) so that the gains over constraint-based methods are not limited to FedProx/MOON.
minor comments (5)
  1. §3 Implementation: λ_FDL/λ_PCL, M and margin m are chosen empirically; a short sensitivity table or plot (beyond the M sweep in Fig. 2) would strengthen reproducibility.
  2. Fig. 1 caption and §2.1: clarify that the frozen global encoder used for FDL is not updated locally and that influence counts are accumulated only over the client's local training samples of each class.
  3. Notation: P is used both for the full prototype tensor and for individual prototypes; a consistent subscript convention (e.g., p_k^c) would improve readability of Eqs. (6)–(7).
  4. Related work: briefly distinguish FedProIn from FPL and FedPLVM on how multi-prototypes are obtained (learnable parameters vs. clustering) so the novelty of NIA is clearer.
  5. Typos / formatting: missing spaces in several compound phrases (e.g., "non-IID data induces", "clientdrift"); "Title Suppressed Due to Excessive Length" headers should be removed in the camera-ready version.

Circularity Check

0 steps flagged

No circularity: empirical FL method with external held-out metrics and no self-referential derivation.

full rationale

FedProIn defines learnable prototypes, Feature Divergence Loss (Eq. 2), Prototype Contrastive Loss (Eq. 3), and Normalized Influence Aggregation (Eqs. 6–7) as design choices, then evaluates the resulting global model on held-out test accuracy, weighted F1 and MCC against independent baselines (FedAvg, FedProx, MOON, BalanceFL, FedProto, FedTGP, FedProc, FedPLVM) under both IID and Dirichlet non-IID partitions of public datasets. No quantity is fitted to data and later re-presented as a prediction; influence scores are simple nearest-prototype counts used only for weighted averaging, not for self-justification. Ablations (Fig. 2) and statistical tests are likewise empirical. Citations are to prior external methods; none supply a uniqueness theorem or ansatz that forces the claimed gains. The paper is therefore self-contained against external benchmarks and exhibits no circular reduction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard FL and prototype-learning machinery plus a handful of empirically chosen scalars and the modeling assumption that client drift cleanly factors into feature and prototype components that the two losses can correct. No new physical entities are postulated; the free parameters are ordinary ML hyper-parameters.

free parameters (4)
  • λ_FDL / λ_PCL = 1.0 / 1.0 or 0.1 / 0.01
    Weighting coefficients for the two regularizers; set to 1.0/1.0 (MCEL) or 0.1/0.01 (MCL) by empirical search on the target datasets.
  • M (prototypes per class) = 1 or 2
    Chosen by ablation; optimal value is dataset-dependent (M=2 for HAM10000, M=1 for Matek-19).
  • margin m in PCL
    Hyper-parameter controlling how much closer a local prototype must be to the global than to its previous state; set empirically.
  • Dirichlet α for non-IID simulation = 0.5
    Controls degree of label/feature heterogeneity; fixed at 0.5 without sensitivity sweep beyond the reported setting.
axioms (4)
  • ad hoc to paper Client drift decomposes additively into feature drift (local encoder maps same image to different regions) and prototype drift (local prototypes pulled toward client-specific statistics).
    Stated in §2.1 as the motivation for introducing separate FDL and PCL terms; not derived from a formal theorem.
  • domain assumption Global prototypes from the previous round are a more generalized representation than any client’s local prototypes, so contrastive attraction is beneficial.
    Explicitly used to justify the form of ℓ_PCL (Eq. 3).
  • ad hoc to paper Nearest-prototype counts (influence matrix I_i) are a faithful measure of a prototype’s contribution to the global representation.
    Defines the NIA weights (Eqs. 6–7); no independent validation that count equals semantic importance.
  • domain assumption Standard FedAvg weighted averaging is appropriate for the feature encoder parameters.
    Adopted without modification (Eq. 5).
invented entities (2)
  • Normalized Influence Aggregation (NIA) no independent evidence
    purpose: Replace data-size weighting of multi-prototypes by usage-count weighting so that under-utilized or client-specific prototypes do not dominate the global average.
    Defined by Eqs. 6–7; no external theoretical guarantee or independent measurement of “influence” is supplied.
  • Feature Divergence Loss (FDL) + Prototype Contrastive Loss (PCL) pair no independent evidence
    purpose: Explicitly regularize the two postulated components of client drift.
    New combination of MSE feature alignment and a margin-based prototype contrast; effectiveness shown only by ablation on the same two datasets.

pith-pipeline@v1.1.0-grok45 · 13698 in / 2865 out tokens · 25961 ms · 2026-07-11T21:16:12.190735+00:00 · methodology

0 comments
read the original abstract

Federated learning (FL) is severely hindered by statistical heterogeneity due to variations in scanners, acquisition protocols, and patient populations. Such non-IID data induces client drift during local optimization, leading to unstable convergence and suboptimal global models when parameter-based aggregation is applied. We propose a prototype-based, influence-aware federated learning framework (FedProIn) that uses multiple learnable class prototypes to capture shared semantic structures across heterogeneous clients. We introduce feature divergence loss and prototype contrastive loss to mitigate client drift by decomposing it into feature drift and prototype drift. In addition, we propose a normalized influence aggregation strategy that adaptively weights client prototypes according to their contribution to the global representation, reducing the impact of biased or low-quality updates. Experimental results on two publicly available medical datasets, HAM10000 and Matek-19, demonstrate that FedProIn achieves accuracies of (83.5% IID, 81.1% non-IID) on HAM10000 and (96.2% IID, 95.8% non-IID) on Matek-19, respectively, outperforming existing baselines in both conditions. Our code is available at https://github.com/harsh-kmr/FedProIn.

Figures

Figures reproduced from arXiv: 2607.04158 by Harsh Kumar, Tarun Kumar Garg, Vaanathi Sundaresan.

Figure 1
Figure 1. Figure 1: illustrates the overall framework of FedProIn. At the start of the communi￾cation round t, the server broadcasts the global state w t−1 to the participating clients. Clients perform local training and return the updated parameters w t i alongside an influence matrix Ii , which guides the server’s aggregation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ablation results. Bar plots showing weighted F1-Score with varying pro￾totype counts (M) on HAM10000. Aggregated across all runs, the full FedProIn method demonstrates statistically significant improvements over all ablated vari￾ants (Mann–Whitney U test; vs. No FDL p < 0.001, vs. No PCL p < 0.05, vs. No FDL & No PCL p < 0.001, vs. No NIA, FDL & PCL p < 0.001). NIA alone already yields noticeable gains, va… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 1 canonical work pages

  1. [1]

    Pattern Recognition151, 110424 (2024)

    Guan, H., Yap, P.-T., Bozoki, A., Liu, M.: Federated learning for med- ical image analysis: A survey. Pattern Recognition151, 110424 (2024). https://doi.org/10.1016/j.patcog.2024.110424

    work page doi:10.1016/j.patcog.2024.110424 2024

  2. [2]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Yang, H.-M., Zhang, X.-Y., Yin, F., Liu, C.-L.: Robust Classification With Convo- lutional Prototype Learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3474–3482. IEEE, Salt Lake City, UT, USA (2018). https://doi.org/10.1109/CVPR.2018.00366

    doi:10.1109/cvpr.2018.00366 2018

  3. [3]

    In: 2022 21st ACM/IEEE Inter- national Conference on Information Processing in Sensor Networks (IPSN), Milano, Italy, pp

    Shuai, X., Shen, Y., Jiang, S., Zhao, Z., Yan, Z., Xing, G.: BalanceFL: Addressing Class Imbalance in Long-Tail Federated Learning. In: 2022 21st ACM/IEEE Inter- national Conference on Information Processing in Sensor Networks (IPSN), Milano, Italy, pp. 271–284. IEEE (2022). https://doi.org/10.1109/IPSN54338.2022.00029

    doi:10.1109/ipsn54338.2022.00029 2022

  4. [4]

    HHS.gov: HIPAA Home | HHS.gov, https://www.hhs.gov/hipaa/index.html, last accessed 2025/01/12

    2025

  5. [5]

    https://gdpr-info.eu/, last accessed 2026/01/25

    GDPR Info: General Data Protection Regulation (GDPR). https://gdpr-info.eu/, last accessed 2026/01/25

    2026

  6. [6]

    In: Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), pp

    McMahan, B., Moore, E., Ramage, D., Hampson, S., Arcas, B.A.: Communication- efficient learning of deep networks from decentralized data. In: Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), pp. 1273–1282 (2017)

    2017

  7. [7]

    In: Proceedings of Machine Learning and Systems (MLSys), vol

    Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. In: Proceedings of Machine Learning and Systems (MLSys), vol. 2, pp. 429–450 (2020). https://arxiv.org/abs/1812.06127

    Pith/arXiv arXiv 2020

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Li, Q., He, B., Song, D.: Model-contrastive federated learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10709–10719 (2021). https://doi.org/10.1109/CVPR46437.2021.01057

    doi:10.1109/cvpr46437.2021.01057 2021

  9. [9]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: FedProto: Fed- erated Prototype Learning across Heterogeneous Clients. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 8, pp. 8432–8440. AAAI Press (2022). https://doi.org/10.1609/aaai.v36i8.20819

    doi:10.1609/aaai.v36i8.20819 2022

  10. [10]

    Future Generation Computer Systems 143, 93–104 (2023)

    Mu, X., Shen, Y., Cheng, K., Geng, J., Jia, J., Li, T.: Fedproc: Prototypical con- trastive federated learning on non-iid data. Future Generation Computer Systems 143, 93–104 (2023)

    2023

  11. [11]

    Zhang, J., Liu, Y., Hua, Y., Cao, J.: FedTGP: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federatedlearning.In:ProceedingsoftheAAAIConferenceonArtificialIntelligence, vol. 38, no. 15, pp. 16768–16776. AAAI Press (2024)

    2024

  12. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2023), pp

    Huang, W., Li, J., Chen, Y., Ding, Z., Zhou, Z.-H.: Rethinking federated learning with domain shift: A prototype view. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2023), pp. 16312–16321. IEEE (2023)

    2023

  13. [13]

    In: Advances in Neural Information Processing Systems, vol

    Wang, L., Zhang, Y., Li, X., Jin, D., Yang, Q.: Taming cross-domain represen- tation variance in federated prototype learning with heterogeneous data domains. In: Advances in Neural Information Processing Systems, vol. 37, pp. 88348–88372 (2024)

    2024

  14. [14]

    Scientific Data5(1), 1–9 (2018) 10 Harsh Kumar, Tarun Kumar Garg, and Vaanathi Sundaresan

    Tschandl, P., Rosendahl, C., Kittler, H.: The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data5(1), 1–9 (2018) 10 Harsh Kumar, Tarun Kumar Garg, and Vaanathi Sundaresan

    2018

  15. [15]

    Matek, C., Schwarz, S., Spiekermann, K., et al.: Human-level recognition of blast cells in acute myeloid leukaemia with convolutional neural networks. Nat. Mach. Intell.1, 538–544 (2019). https://doi.org/10.1038/s42256-019-0101-9

    doi:10.1038/s42256-019-0101-9 2019

  16. [16]

    In: III, H.D., Singh, A

    Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S., Stich, S.U., Suresh, A.T.: Scaf- fold: Stochastic controlled averaging for federated learning. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning (ICML 2020), Proceedings of Machine Learning Research, vol. 119, pp. 5132–5143. PMLR (2020)

    2020

  17. [17]

    In: International Confer- ence on Learning Representations (ICLR) (2021)

    Acar,D.A.E.,Zhao,Y.,Navarro,R.M.,Mattina,M.,Whatmough,P.N.,Saligrama, V.: Federated learning based on dynamic regularization. In: International Confer- ence on Learning Representations (ICLR) (2021)

    2021

  18. [18]

    In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A

    Tan, Y., Long, G., Ma, J., Liu, L., Zhou, T., Jiang, J.: Federated learning from pre-trained models: A contrastive learning approach. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems 35 (NeurIPS 2022), pp. 19332–19344 (2022)

    2022

  19. [19]

    Matthews, B.W.: Comparison of the predicted and observed secondary structure of T4 phage lysozyme. Biochim. Biophys. Acta405(2), 442–451 (1975)

    1975

  20. [20]

    In: 2009 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., Fei-Fei, L.: ImageNet: A large- scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 248–255. IEEE (2009)

    2009

  21. [21]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  22. [22]

    Advances in neural information processing systems32, 2019

    Paszke, A and Gross, S and Massa, F and Lerer, A., et al.: Pytorch: An impera- tive style, high-performance deep learning library. Advances in neural information processing systems32, 2019

    2019