Pith. sign in

REVIEW 4 major objections 3 minor 46 references

Measuring and Predicting Where and When Pathologists Focus their Visual Attention while Grading Whole Slide Images of Cancer

T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A two-stage transformer predicts where and when pathologists look while grading prostate cancer whole-slide images, the paper claims.

desk verdict The abstract announces a pathology attention scanpath model, but the submitted body is an unrelated federated learning paper, so the claimed contribution is entirely absent. read the letter →

arxiv 2508.01668 v1 pith:O6GN2GNH submitted 2025-08-03 eess.IV cs.CV

Souradeep Chakraborty , Ruoyu Xue , Rajarsi Gupta , Oksana Yaskiv , Constantin Friedman , Natallia Sheuka , Dana Perez , Paul Friedman
show 6 more authors
Won-Tak Choi Waqas Mahmud Beatrice Knudsen Gregory Zelinsky Joel Saltz Dimitris Samaras
This is my paper
classification eess.IVcs.CV
keywords visualattentionpredictionscanpathwholeslideimagesprostatecancergradingdigitalpathologyfixationextractiontransformerpathologisttraining
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

The paper tries to establish that a pathologist's visual attention during prostate cancer grading can be predicted as a dynamic scanpath, not just as a static map of where experts look. It defines attention as the trajectory of viewport centers and magnification levels, compresses those trajectories into fixations with a new extraction algorithm, and trains a two-stage transformer model on data from 43 pathologists viewing 123 whole-slide images. The first stage predicts attention heatmaps at different magnifications, and the second stage autoregressively predicts the next fixation points starting from the WSI center. If the claim holds, expert-like attention becomes a learnable output that could support tools for training pathology residents.

What carries the argument

The central object is the attention scanpath, represented as a temporal sequence of viewport positions (x, y, m), where m is the magnification level at which the pathologist is viewing. A fixation extraction algorithm compresses raw viewport trajectories into fixations while preserving semantic information, and this preprocessed data is what the model learns from. The predictive machinery is a two-stage transformer: the first sub-network predicts attention heatmaps as static attention across different magnifications, and the second sub-network uses these heatmaps as multi-magnification feature representations to autoregressively predict the next fixation points, starting from the WSI center.

What would settle it

Record eye-tracking data from pathologists while they grade the same prostate cancer WSIs and compare gaze fixations with the fixations extracted from viewport-center trajectories. If many gaze fixations occur at positions that differ from the viewport center while the viewport is stationary, or if predicted scanpaths match viewport-derived fixations but not gaze fixations, then the model is predicting a self-defined proxy rather than expert attention.

Watch

Extended reading notes

Core claim

The central claim is that expert attention in whole-slide image reading is predictable at the scanpath level. The paper defines an attention trajectory as the sequence of viewport center coordinates (x, y) together with magnification m, collects this from 43 pathologists across 123 prostate cancer WSIs, and introduces a fixation extraction algorithm that simplifies each trajectory while preserving semantic information. A two-stage transformer then models the scanpath: the first stage produces attention heatmaps across magnifications, and the second stage consumes multi-magnification features from the first stage to autoregressively predict successive fixation points, beginning at the WSI center. The abstract reports that the resulting scanpath predictions outperform chance and baseline models, with the intended payoff being a training tool that helps pathology trainees allocate attention the way experts do.

Load-bearing premise

The approach assumes that where a pathologist moves the viewport center, plus the magnification setting, is a faithful record of where their attention is, even though eyes can move within a stationary viewport and not every viewport center is a true fixation.

Editorial extensions

If this is right

  • Digital pathology training platforms could show trainees where an expert would look next as they scan a slide.
  • The first-stage attention heatmaps could serve as visual summaries of diagnostically relevant regions, independent of the scanpath.
  • The two-stage autoregressive design could be applied to other image-heavy domains where experts navigate large images, such as radiology.
  • Reliable scanpath prediction could generate synthetic expert-viewing demonstrations for education when real expert time is scarce.

Reading between the lines

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

  • If viewport trajectories are genuinely predictable, they likely encode diagnostic strategy, so predicted scanpaths could be used as additional features for automated WSI grading models.
  • The same approach could be tested on other cancer types or on screening tasks where search strategy, not just final diagnosis, is the object of interest.
  • A caution specific to the received manuscript: the full text supplied is about a different topic, heterogeneous federated learning, so the abstract's experimental claims could not be checked against the body in this version; verification requires the actual attention-prediction paper.
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 / 3 minor

Summary. The manuscript announces a study that measures and predicts where and when pathologists focus their visual attention while grading whole slide images (WSIs) of prostate cancer. The abstract describes a dataset of x, y, and magnification viewport trajectories from 43 pathologists over 123 WSIs, a fixation extraction algorithm, a two-stage transformer model for static attention heatmap prediction followed by autoregressive scanpath generation, and experimental results claimed to outperform chance and baseline models. However, the supplied full text is the complete text of a different paper, arXiv:2508.01669 (FedVTC), which is about model-heterogeneous federated learning and contains no mention of WSIs, pathologists, fixations, scanpaths, attention heatmaps, or magnification. The announced method, experiments, and results are therefore entirely absent from the manuscript.

Significance. If the claimed contribution existed as described, it could be valuable for pathology training and for understanding expert visual search in gigapixel images, and the reported dataset of 43 pathologists and 123 WSIs would be a useful resource. However, the current submission provides no method description, experimental protocol, baseline definitions, evaluation metrics, or results for the announced task. The full text appears to be a complete, self-contained paper on a different topic; that work may have its own merits, but it does not support or even address the abstract's claims. No code, data, or machine-checked proofs are provided for the attention-prediction contribution. I concur with the stress-test concern: the central claim is unverifiable because the manuscript does not contain its supporting evidence.

major comments (4)
  1. [Abstract / Full Text] The central claim that the proposed scanpath prediction model outperforms chance and baseline models is unsupported because the supplied full text is the complete text of a different paper, arXiv:2508.01669 (FedVTC), on model-heterogeneous federated learning; it contains no mention of whole slide images, pathologists, fixations, scanpaths, attention heatmaps, magnification, or the proposed two-stage architecture. Consequently, no method, dataset description, experimental protocol, baseline comparison, or result table for the announced attention-prediction task exists in this manuscript.
  2. [Abstract (fixation extraction)] The abstract states that the fixation extraction algorithm 'preserv[es] semantic information' and that the scanpath targets are constructed from viewport centers, but the manuscript never defines the fixation extraction thresholds, windowing parameters, or any validation that extracted fixations correspond to actual attentional fixations; this is load-bearing because the model is trained and evaluated against these self-defined targets.
  3. [Full Text, Section 4] The only experiments reported in the full text are FedVTC's generalization accuracy, communication cost, and memory consumption on MNIST, CIFAR10, CIFAR100, and Tiny-ImageNet (Tables 1-4); these are unrelated to where-and-when attention prediction on prostate whole slide images, so the abstract's empirical claim cannot be checked.
  4. [Full Text, Section 3] The two-stage transformer architecture is only named in the abstract; the body's methodology section describes variational transposed convolution for federated learning, not attention heatmap prediction or autoregressive next-fixation generation, so the proposed model's inputs, outputs, loss functions, and hyperparameters are entirely absent.
minor comments (3)
  1. [Appendix C] The appendix states that large language models were used only to correct spelling and grammatical mistakes, but this statement appears to refer to the FedVTC text rather than the announced attention-prediction manuscript; a resubmission would need a statement covering the actual manuscript.
  2. [Abstract (magnification levels)] The abstract refers to 'different magnifications' but does not define the magnification levels or their relationship to WSI pyramid levels; this should be specified in any revised version.
  3. [General] The manuscript lacks a data availability statement for the viewport trajectories and any code, and it does not cite prior work on eye tracking in pathology or scanpath prediction; such details would be needed for reproducibility and context.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular step is demonstrable: the abstract's scanpath claims have no supporting body text, so there is no derivation chain to reduce to its inputs.

full rationale

Under the strict requirement that circularity be shown as Eq. X = Eq. Y by construction or a fitted parameter renamed as a prediction, no such reduction can be quoted from this submission. The abstract describes a pathology attention-scanpath system, but the full text is the unrelated FedVTC federated-learning paper (arXiv:2508.01669): it contains no whole-slide images, pathologists, viewport traces, fixation extraction, attention heatmaps, two-stage transformer, or scanpath experiments. The abstract's claim that 'our scanpath prediction model outperforms chance and baseline models' is therefore unsupported and unverifiable, and the described derivation chain is entirely missing. That is a completeness/integrity failure, not a circularity failure. The abstract's operationalization of attention as viewport x,y,m movement is an explicit labeling choice; even if one doubts its construct validity, the abstract does not secretly fit the target into the model, and no equation-level reduction is available to exhibit. Accordingly, the circularity score is 0, with the caveat that the submission's central claim cannot be checked because its methods and results are absent.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the validity of the viewport trajectory as attention, on the fixation extraction preserving meaning, and on the sample representing expert grading. None of these are established in the abstract, and the submitted body is unrelated. Model hyperparameters and fixation thresholds are free parameters that are not reported.

free parameters (3)
  • fixation extraction thresholds and windowing parameters = not specified
    The proposed fixation extraction algorithm converts viewport trajectories into fixation scanpaths; its thresholds determine the target labels used for training and evaluation and are not stated in the abstract.
  • transformer depth, heads, and latent dimensions = not specified
    Both stages are described only as transformer-based; architecture choices affect capacity and are absent from the submission.
  • scanpath length and stopping criterion = not specified
    The autoregressive stage predicts next fixation points starting at the WSI center; the number of steps and when generation stops shape both the predicted trajectories and how they are scored.
assumptions (3)
  • domain assumption Viewport center coordinates plus magnification are a faithful proxy for a pathologist's visual attention.
    The abstract defines attention as x, y, m viewport movements; gaze can move within a stationary viewport, so this is a measurement assumption rather than a demonstrated fact. Entire study is built on it.
  • ad hoc to paper Fixation extraction preserves the semantic content of the attention trajectory.
    The abstract states the algorithm simplifies trajectories while preserving semantic information but gives no proof or ablation in the provided text; this defines the very target the model predicts.
  • domain assumption The 43 pathologists and 123 WSIs are representative of expert prostate cancer grading behavior.
    Generalization of the predictor depends on this sample; no sampling or diversity information is in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Measuring and Predicting Where and When Pathologists Focus their Visual Attention while Grading Whole Slide Images of Cancer." pith.science (2026). https://pith.science/paper/O6GN2GNH

@misc{pith2026250801668,
  author       = {Pith},
  title        = {Pith review of: Measuring and Predicting Where and When Pathologists Focus their Visual Attention while Grading Whole Slide Images of Cancer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O6GN2GNH}},
  note         = {Machine review of arXiv:2508.01668}
}
read the original abstract

The ability to predict the attention of expert pathologists could lead to decision support systems for better pathology training. We developed methods to predict the spatio-temporal (where and when) movements of pathologists' attention as they grade whole slide images (WSIs) of prostate cancer. We characterize a pathologist's attention trajectory by their x, y, and m (magnification) movements of a viewport as they navigate WSIs using a digital microscope. This information was obtained from 43 pathologists across 123 WSIs, and we consider the task of predicting the pathologist attention scanpaths constructed from the viewport centers. We introduce a fixation extraction algorithm that simplifies an attention trajectory by extracting fixations in the pathologist's viewing while preserving semantic information, and we use these pre-processed data to train and test a two-stage model to predict the dynamic (scanpath) allocation of attention during WSI reading via intermediate attention heatmap prediction. In the first stage, a transformer-based sub-network predicts the attention heatmaps (static attention) across different magnifications. In the second stage, we predict the attention scanpath by sequentially modeling the next fixation points in an autoregressive manner using a transformer-based approach, starting at the WSI center and leveraging multi-magnification feature representations from the first stage. Experimental results show that our scanpath prediction model outperforms chance and baseline models. Tools developed from this model could assist pathology trainees in learning to allocate their attention during WSI reading like an expert.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 35 canonical work pages

  1. [5]

    �������� �������� ���������������� ���������� �������� �������� ������������ ���������������� ���������� ���������������������� ���������� ���������������� ���������� ������������ ���������������� ���������� ���������� Figure 5: The sub-model aggregation scheme for sub-model-based model-heterogeneous FL (”W A” stands for weighted average). A.2 COMPLETEMOD...

    work page 2023

  2. [6]

    Alternatively, clients communicate some Parameter-Independent Messages (PIM) with the server for knowledge sharing, such as prototypes and feature covariates

    15 Published as a conference paper at ICLR 2026 ������������ ���������������� ���������������� ���������������� ������ ������ ���������������������������� Figure 6: In complete model-heterogeneous FL, the model architectures between clients are inde- pendent, irrelevant, and cannot be aggregated altogether. Alternatively, clients communicate some Paramete...

    work page 2026

  3. [7]

    Enmao Diao, Jie Ding, and Vahid Tarokh

    doi: 10.1109/TPAMI.2023.3261988. Enmao Diao, Jie Ding, and Vahid Tarokh. Hetero�fl�: Computation and communication efficient federated learning for heterogeneous clients. InProceedings of the 2021 International Conference on Learning Representations (ICLR 2021),

    arXiv 2023

  4. [8]

    Dinh, Nguyen H

    Canh T. Dinh, Nguyen H. Tran, and Tuan Dung Nguyen. Personalized federated learning with moreau envelopes. InProceedings of the 34th Conference on Neural Information Processing Systems (NeurIPS 2020),

    work page 2020

  5. [9]

    Vincent Dumoulin and Francesco Visin

    ISBN 9781713829546. Vincent Dumoulin and Francesco Visin. A guide to convolution arithmetic for deep learning.arXiv preprint arXiv:1603.07285,

    arXiv

  6. [11]

    Deep residual learning for image recog- nition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. InProceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2016), pp. 770–778,

    work page 2016

  7. [13]

    Overcoming data and model heterogeneities in decentralized federated learning via synthetic anchors.arXiv preprint arXiv:2405.11525,

    Chun-Yin Huang, Kartik Srinivas, Xin Zhang, and Xiaoxiao Li. Overcoming data and model heterogeneities in decentralized federated learning via synthetic anchors.arXiv preprint arXiv:2405.11525,

    arXiv

  8. [14]

    Towards understanding biased client selection in federated learning

    Yae Jee Cho, Jianyu Wang, and Gauri Joshi. Towards understanding biased client selection in federated learning. InProceedings of the 25th International Conference on Artificial Intelligence and Statistics (AISTATS 2022), volume 151, pp. 10351–10375,

    work page 2022

Show all 46 references
  1. [15]

    Model pruning enables efficient federated learning on edge devices.IEEE Transactions on Neural Networks and Learning Systems,

    11 Published as a conference paper at ICLR 2026 Yuang Jiang, Shiqiang Wang, Victor Valls, Bong Jun Ko, Wei-Han Lee, Kin K Leung, and Leandros Tassiulas. Model pruning enables efficient federated learning on edge devices.IEEE Transactions on Neural Networks and Learning Systems,

    2026

  2. [16]

    SCAFFOLD: Stochastic controlled averaging for federated learning

    Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for federated learning. InProceedings of the 37th International Conference on Machine Learning (ICML 2020), volume 119, p...

    2020

  3. [17]

    DepthFL : Depthwise federated learning for heterogeneous clients

    Minjae Kim, Sangyoon Yu, Suhyun Kim, and Soo-Mook Moon. DepthFL : Depthwise federated learning for heterogeneous clients. InProceedings of the Eleventh International Conference on Learning Representations (ICLR 2023),

    2023

  4. [18]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    arXiv

  5. [21]

    Daliang Li and Junpu Wang

    ISBN 9781450391818. Daliang Li and Junpu Wang. Fedmd: Heterogenous federated learning via model distillation.arXiv preprint arXiv:1910.03581,

    1910 arXiv

  6. [23]

    No fear of heterogene- ity: Classifier calibration for federated learning with non-IID data

    12 Published as a conference paper at ICLR 2026 Mi Luo, Fei Chen, Dapeng Hu, Yifan Zhang, Jian Liang, and Jiashi Feng. No fear of heterogene- ity: Classifier calibration for federated learning with non-IID data. InProceedings of the 35th Conference on Neural Information Proces...

    2026

  7. [24]

    Personalized federated learning via feature distribution adaptation

    Connor Mclaughlin and Lili Su. Personalized federated learning via feature distribution adaptation. Proceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS 2024), 37:77038–77059,

    2024

  8. [25]

    Duy Phuong Nguyen, Sixing Yu, J

    doi: 10.1109/TDSC.2024.3372634. Duy Phuong Nguyen, Sixing Yu, J. Pablo Mu˜noz, and Ali Jannesari. Enhancing heterogeneous fed- erated learning with knowledge extraction and multi-model fusion. InProceedings of the SC ’23 Workshops of the International Conference on High Perfor...

    2024

  9. [26]

    doi: 10.1145/3624062.3626325

    ISBN 9798400707858. doi: 10.1145/3624062.3626325. Ziru Niu, Hai Dong, A. K. Qin, and Tao Gu. Flrce: Resource-efficient federated learning with early- stopping strategy.IEEE Transactions on Mobile Computing, 23(12):14514–14529,

  10. [27]

    doi: 10.1109/TMC.2024.3447000. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K ¨opf, Ed- ward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank C...

    2024

  11. [29]

    Nalavala Partheswar Reddy, Shaik Khaja Hassain, P Baraneedharan, and Shaik Abdul Rahman

    doi: 10.1109/TSP.2022.3153135. Nalavala Partheswar Reddy, Shaik Khaja Hassain, P Baraneedharan, and Shaik Abdul Rahman. In- ternet of things-based remote weather forecasting system using legacy sensors. InProceedings of the 9th International Conference on Advanced Computing an...

    2022

  12. [30]

    Himadri Nath Saha, Supratim Auddy, Subrata Pal, Shubham Kumar, Shivesh Pandey, Rocky Singh, Amrendra Kumar Singh, Priyanshu Sharan, Debmalya Ghosh, and Sanhita Saha

    doi: 10.1109/ICACCS57279.2023.10113052. Himadri Nath Saha, Supratim Auddy, Subrata Pal, Shubham Kumar, Shivesh Pandey, Rocky Singh, Amrendra Kumar Singh, Priyanshu Sharan, Debmalya Ghosh, and Sanhita Saha. Health mon- itoring using internet of things (iot). InProceedings of th...

    2023

  13. [31]

    Felix Sattler, Tim Korjakow, Roman Rischke, and Wojciech Samek

    doi: 10.1109/IEMECON.2017.8079564. Felix Sattler, Tim Korjakow, Roman Rischke, and Wojciech Samek. Fedaux: Leveraging unla- beled auxiliary data in federated learning.IEEE Transactions on Neural Networks and Learning Systems, 34(9):5531–5543,

    2017

  14. [32]

    13 Published as a conference paper at ICLR 2026 Virginia Smith, Chao-Kai Chiang, Maziar Sanjabi, and Ameet S Talwalkar

    ISBN 978-1-57735-887-9. 13 Published as a conference paper at ICLR 2026 Virginia Smith, Chao-Kai Chiang, Maziar Sanjabi, and Ameet S Talwalkar. Federated multi-task learning.Proceedings of the 31st Conference on Neural Information Processing Systems (NIPS 2017), 30,

    2026

  15. [33]

    Phoenix: A federated generative diffusion model

    Fiona Victoria Stanley Jothiraj and Afra Mashhadi. Phoenix: A federated generative diffusion model. InCompanion Proceedings of the ACM Web Conference 2024, WWW ’24, pp. 1568–1577, New York, NY , USA,

    2024

  16. [34]

    Robustfed: A truth inference approach for robust federated learning

    Farnaz Tahmasebian, Jian Lou, and Li Xiong. Robustfed: A truth inference approach for robust federated learning. InProceedings of the 2022 International Conference on Information and Knowledge Management (CIKM 2022), pp. 1868–1877,

    2022

  17. [35]

    Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang

    ISBN 9781450392365. Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fed- proto: Federated prototype learning across heterogeneous clients. InProceedings of the 2022 AAAI Conference on Artificial Intelligence (AAAI 2022), volume 36, pp. 8432–8440,

    2022

  18. [36]

    Bridging model heterogeneity in federated learning via uncertainty-based asymmetrical reciprocity learn- ing

    Jiaqi Wang, Chenxu Zhao, Lingjuan Lyu, Quanzeng You, Mengdi Huai, and Fenglong Ma. Bridging model heterogeneity in federated learning via uncertainty-based asymmetrical reciprocity learn- ing. InProceedings of the 41st International Conference on Machine Learning (ICML 2024),

    2024

  19. [37]

    Liping Yi, Han Yu, Gang Wang, Xiaoguang Liu, and Xiaoxiao Li

    ISSN 0360-0300. Liping Yi, Han Yu, Gang Wang, Xiaoguang Liu, and Xiaoxiao Li. pfedafm: Adaptive feature mix- ture for data-level personalization in heterogeneous federated learning on mobile edge devices. In 2025 IEEE 41st International Conference on Data Engineering (ICDE), p...

    2025

  20. [38]

    Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning

    Jianqing Zhang, Yang Liu, Yang Hua, and Jian Cao. Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning. InProceedings of the 2024 AAAI Conference on Artificial Intelligence (AAAI 2024), vo...

    2024

  21. [39]

    Fed-CBS: A heterogeneity-aware client sampling mechanism for federated learning via class-imbalance reduction

    Jianyi Zhang, Ang Li, Minxue Tang, Jingwei Sun, Xiang Chen, Fan Zhang, Changyou Chen, Yiran Chen, and Hai Li. Fed-CBS: A heterogeneity-aware client sampling mechanism for federated learning via class-imbalance reduction. InProceedings of the 40th International Conference on Ma...

    2023

  22. [40]

    Improved distribution matching for dataset condensation

    Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. InProceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2023), pp. 7856–7865,

    2023

  23. [41]

    Data-free knowledge distillation for heterogeneous federated learning

    Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. Data-free knowledge distillation for heterogeneous federated learning. InProceedings of the 2021 International Conference on Machine Learning (ICML 2021), pp. 12878–12889,

    2021

  24. [42]

    A DEFINITION OFMODELHETEROGENEITY To our best knowledge, the definitions of ”model-heterogeneous federated learning” in the state-of- the-art can be divided into the following two categories. 14 Published as a conference paper at ICLR 2026 A.1 SUB-MODEL-BASEDMODELHETEROGENEITY...

    2026

  25. [43]

    (Jiang et al., 2023; Li et al., 2021a; Jiang et al.,

    dynamically extracts sub-models using a sliding window across the neurons in the global model. (Jiang et al., 2023; Li et al., 2021a; Jiang et al.,

    2023

  26. [46]

    Despite this, we additionally tune�on the MNIST dataset to study the impact of different�values

    for fairness. Despite this, we additionally tune�on the MNIST dataset to study the impact of different�values. As shown in Figure 8a, when�is too small, the generalization accuracy is low as the effect of fine-tuning is weak. When�is too large, many redundant synthetic samples...

    2026

  27. [1998]

    Ang Li, Jingwei Sun, Pengcheng Li, Yu Pu, Hai Li, and Yiran Chen

    URLhttp://yann.lecun.com/exdb/mnist/. Ang Li, Jingwei Sun, Pengcheng Li, Yu Pu, Hai Li, and Yiran Chen. Hermes: An efficient federated learning framework for heterogeneous mobile clients. InProceedings of the 27th Annual Interna- tional Conference on Mobile Computing and Netwo...

    2021

  28. [2009]

    Madhyastha, and Mosharaf Chowdhury

    Fan Lai, Xiangfeng Zhu, Harsha V . Madhyastha, and Mosharaf Chowdhury. Oort: Efficient feder- ated learning via guided participant selection. InProceedings of the 15th USENIX Symposium on Operating Systems Design and Implementation (OSDI 2021), pp. 19–35,

    2021

  29. [2016]

    Feddc: Federated learning with non-iid data via local drift decoupling and correction

    Liang Gao, Huazhu Fu, Li Li, Yingwen Chen, Ming Xu, and Cheng-Zhong Xu. Feddc: Federated learning with non-iid data via local drift decoupling and correction. InProceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2022), pp. 10112– 10121, June

    2022

  30. [2017]

    ISBN 9781510860964

    Curran Associates Inc. ISBN 9781510860964. Samuel Horv ´ath, Stefanos Laskaridis, Mario Almeida, Ilias Leontiadis, Stylianos Venieris, and Nicholas Lane. Fjord: Fair and accurate federated learning under heterogeneous targets with ordered dropout. InProceedings of the 35th Con...

    2021

  31. [2018]

    Fltrust: Byzantine-robust feder- ated learning via trust bootstrapping.arXiv preprint arXiv:2012.13995,

    Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. Fltrust: Byzantine-robust feder- ated learning via trust bootstrapping.arXiv preprint arXiv:2012.13995,

    2012 arXiv

  32. [2019]

    Learning to detect malicious clients for robust federated learning.arXiv preprint arXiv:2002.00211, 2020a

    Suyi Li, Yong Cheng, Wei Wang, Yang Liu, and Tianjian Chen. Learning to detect malicious clients for robust federated learning.arXiv preprint arXiv:2002.00211, 2020a. Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimi...

    2002 arXiv

  33. [2020]

    Ex- planatory object part aggregation for zero-shot learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(2):851–868,

    10 Published as a conference paper at ICLR 2026 Xin Chen, Xiaoling Deng, Yubin Lan, Yongbing Long, Jian Weng, Zhiquan Liu, and Qi Tian. Ex- planatory object part aggregation for zero-shot learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(2):851–868,

    2026

  34. [2021]

    Fedrolex: model-heterogeneous federated learning with rolling sub-model extraction

    Samiul Alam, Luyang Liu, Ming Yan, and Mi Zhang. Fedrolex: model-heterogeneous federated learning with rolling sub-model extraction. InProceedings of 31st Conference on Neural Infor- mation Processing Systems (NeurIPS 2022),

    2022

  35. [2022]

    Main Track

    doi: 10.24963/ijcai.2022/399. Main Track. Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the cifar datasets.arXiv preprint arXiv:1707.08819,

    2022 arXiv

  36. [2023]

    Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama

    doi: 10.1109/TSC.2021.3121879. Durmus Alp Emre Acar, Yue Zhao, Ramon Matas, Matthew Mattina, Paul Whatmough, and Venkatesh Saligrama. Federated learning based on dynamic regularization. InProceedings of the 2021 International Conference on Learning Representations (ICLR 2021),

    2021

  37. [2024]

    Sijie Cheng, Jingwen Wu, Yanghua Xiao, Yang Liu, and Yang Liu

    doi: 10.1109/TPAMI.2023.3325533. Sijie Cheng, Jingwen Wu, Yanghua Xiao, Yang Liu, and Yang Liu. Fedgems: Federated learning of larger server models via selective knowledge fusion.arXiv preprint arXiv:2110.11027,

    2023

  38. [2025]

    Kilian Pfeiffer, Ramin Khalili, and Joerg Henkel

    doi: 10.1109/TCCN.2025.3550359. Kilian Pfeiffer, Ramin Khalili, and Joerg Henkel. Aggregating capacity in FL through successive layer training for computationally-constrained devices. InProceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS 2023),

    2025

Pith tools

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