Pith. sign in

REVIEW 4 major objections 6 minor 39 references

Adaptive Prototype Knowledge Transfer for Federated Learning with Mixed Modalities and Heterogeneous Tasks

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A prototype-based federated learning framework coordinates image-only, text-only, and image-text clients on different tasks without a unified label space, and reports consistent gains over four baselines on three highly heterogeneous…

desk verdict A promising framework for prototype-based mixed-modality FL, but the paper's central premise — that client embeddings live in a shared space — is left ambiguous, and the experiments are not yet rigorous enough to confirm the method works as claimed. read the letter →

arxiv 2502.04400 v2 pith:VOCRC2FL submitted 2025-02-06 cs.LG cs.AIcs.CRcs.MM

Keke Gai , Mohan Wang , Jing Yu , Dongjue Wang , Qi Wu This is my paper · ORCID
classification cs.LGcs.AIcs.CRcs.MM
keywords federatedlearningmultimodalprototypeheterogeneoustasksknowledgetransfermixedmodalitiesclientrelationshipgraphmodalityalignment
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

Multimodal federated learning often assumes every client works on the same task with the same label space. This paper argues that the assumption is avoidable: it proposes AproMFL, a prototype-based framework in which clients with different modalities and different tasks can still exchange knowledge. Clients with labels build class-mean prototypes; clients without labels cluster their fused image-text embeddings to build pseudo-labeled prototype pairs. The server completes unimodal prototypes into multimodal ones, clusters them into global prototype pairs, and adaptively aggregates mapping modules instead of plain averaging. On three datasets with strong Non-IID data ($\alpha=0.1$), AproMFL reports consistent accuracy and recall gains over Local, FedIoT, MM-FedProx, and CreamFL.

What carries the argument

The key machinery is the prototype pipeline: class-mean embeddings when labels exist, K-means pseudo-labeled image-text pairs when they do not, and cosine-similarity-weighted semantic completion on the server. Semantic completion pairs a unimodal prototype with the top-$O$ most similar multimodal prototypes and weights them by similarity, turning image-only and text-only clients into contributors to a shared multimodal space. K-means on the completed prototypes creates global image-text prototype pairs, and a client relationship graph over mapping-module parameters determines aggregation weights. Two losses, $L_{GPT}$ and $L_{GMT}$, transfer global prototype knowledge and global model knowledge back into local models, aligning local and global representations.

What would settle it

Replace the shared CLIP encoder with different pretrained encoders for different clients, run the same AproMFL pipeline, and check whether the server's cosine-similarity semantic completion and K-means clustering still produce global prototypes that improve local accuracy and recall; if the gains over the Local baseline vanish, the shared-encoder premise is the load-bearing assumption.

Watch

Extended reading notes

Core claim

The paper's central discovery is that prototypes can serve as a task-agnostic and label-agnostic medium for knowledge transfer in multimodal federated learning. AproMFL's adaptive prototype construction lets labeled unimodal clients contribute class centroids while unlabeled multimodal clients contribute pseudo-labeled image-text prototype pairs. The server then converts unimodal prototypes into multimodal ones through similarity-weighted semantic completion, clusters all prototypes to form global pairs, and aggregates mapping modules using a client relationship graph instead of averaging whole models. Local training is regularized by a global prototype knowledge transfer loss and a global model knowledge transfer loss, which pull local representations toward the shared global knowledge. The reported experiments show gains over four baselines, with the best accuracy and recall 0.42% to 6.09% and 1.6% to 3.89% higher than FedIoT, respectively.

Load-bearing premise

All clients must use the same pretrained CLIP encoder, so their image and text embeddings already sit in a shared semantic space.

Editorial extensions

If this is right

  • A deployment can mix classification clients and retrieval clients without forcing them into the same label vocabulary.
  • No public dataset or shared label set is needed for cross-client knowledge transfer in the mixed-modality setting.
  • Unimodal clients inherit multimodal semantics through similarity-weighted prototype completion, which the paper shows improves their classification accuracy.
  • The client-relationship-graph aggregation avoids the model drift that plain averaging causes when local tasks differ.
  • The reported gains persist at high data heterogeneity ($\alpha=0.1$) and across three datasets.

Reading between the lines

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

  • If the shared-encoder premise is relaxed, the method would need an online alignment step; the paper does not test this, so a natural follow-up is to compare AproMFL against a variant with per-client encoder adaptation.
  • The same semantic-completion design could transfer to other modality pairs such as audio-video or sensor-text, provided the server has a similarity measure between the two modalities.
  • Because completed prototypes are transmitted to the server, a privacy analysis of what those prototypes leak about a client's class distribution remains an open extension beyond what the paper reports.
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 / 6 minor

Summary. The paper proposes AproMFL, a prototype-based multimodal federated learning framework for settings where clients have mixed modalities (image, text, multimodal) and heterogeneous tasks (classification and image-text retrieval) without a unified label space. Unimodal clients build label-guided prototypes, multimodal clients build clustering-based prototype pairs, and the server completes unimodal prototypes to multimodal pairs using cosine similarity, clusters them into global prototype pairs, and aggregates mapping modules via a client relationship graph. Local training is regularized by a global prototype transfer loss and a global model transfer loss. The authors report experiments on CIFAR-10, AG-NEWS, and Flickr30k comparing against Local, FedIoT, MM-FedProx, and CreamFL, and claim consistent gains in accuracy and recall.

Significance. The problem is timely and relevant: enabling collaboration among clients with genuinely different modalities and tasks, without unified labels, is an under-explored gap in multimodal federated learning. The adaptive prototype construction (label-guided for unimodal clients, clustering-based for multimodal clients) is a sensible design, and the client relationship graph aggregation is a plausible alternative to FedAvg under task heterogeneity. If the empirical claims were supported by rigorous experiments, this would be a useful contribution. The paper is readable and the method is mostly coherent. However, the current experimental protocol is not strong enough to establish the claimed gains, and one load-bearing assumption about the shared embedding space is left ambiguous. No code is provided.

major comments (4)
  1. [Implementation Details; Server-side Adaptive Aggregation] The entire cross-client prototype aggregation depends on comparing prototypes from different clients by cosine similarity. The manuscript never states whether the CLIP ViT-L/14 encoders are frozen during local training. In Table 5, the authors write that 'a projection head is appended to these backbone networks for fine-tuning,' which suggests that the backbones are updated. If each client fine-tunes its encoder on private data and a private task, prototypes from different clients live in incomparable feature spaces, and the semantic completion, multimodal clustering, and the L_GP_T alignment in Eqs. (7)-(8) lose their meaning. If the encoders are instead frozen, then the abstract's claim that prototype-based MFL offers 'no reliance on prior knowledge' is contradicted by the use of a pretrained CLIP model, and the cross-modal alignment is effectively outsourced to that model. Please state explicitly whether encoders are frozen or fine-tuned and, if they are fine-tuned, justify why cosine similarities between prototypes across clients remain commensurable.
  2. [Table 1; Experiments] The central empirical claim is supported only by point estimates. Table 1 reports no standard deviations, no number of repeated runs, and no significance tests. Many reported margins are small (for example, 0.42% on AG-NEWS Acc@1 between AproMFL and FedIoT at alpha=0.1, and 1.6% on Flickr30k R@1(t2i)), so without variance estimates these differences may be within run-to-run noise. Please report results over at least five independent seeds with means and standard deviations, and specify the number of clients, local epochs, and communication rounds used for the main comparison.
  3. [Tables 3 and 4 vs. Table 1] The main results in Table 1 appear to use K=10 and O=10, because the K=10 row in Table 3 and the O=10 row in Table 4 exactly reproduce the AproMFL numbers from Table 1. However, Tables 3 and 4 report test accuracy and recall across K and O, and the best values are not K=10/O=10: O=2 yields higher CIFAR-10 Acc@1 (60.98%) and Flickr30k R@1s (90.33%) than O=10, and K=40, 60, and 80 yield higher CIFAR-10 Acc@1 than K=10. If these sweeps were used to select K and O on the test set, the comparison in Table 1 is not a fair out-of-sample evaluation. Please either select hyperparameters on a held-out validation split, report a single default configuration, or justify why the particular K=10/O=10 configuration was used for the main comparison.
  4. [Baselines; Table 1] The configuration of the baselines is underspecified. CreamFL is a knowledge-distillation method that relies on a public dataset, but the Implementation Details do not state whether a public dataset was provided to CreamFL, whether all methods share the same CLIP encoders and mapping modules, or how many clients and rounds were used. CreamFL's very low numbers in Table 1 (e.g., 20.62% Acc@1 on CIFAR-10 at alpha=0.1) suggest that this baseline may not have been given its required public-data condition or was otherwise not configured in its intended operating regime. Please document the exact setup for every baseline, including any public data, optimizer, learning rate, and communication rounds, and rerun the comparisons under matched configurations.
minor comments (6)
  1. [Eq. (9); text after Eq. (9)] In the sentence defining the factor nu, both losses are written as Ll_task; the second one should be Lg_task, the global model's task loss.
  2. [Eq. (8)] The loss L_GP_T is defined for a single sample j but no summation over the minibatch is shown; please clarify whether the loss is averaged over all samples in the batch.
  3. [Abstract and Introduction] The phrase 'no reliance on prior knowledge' is inaccurate when the method uses pretrained CLIP encoders; the authors should either remove this claim or qualify it to mean no shared public dataset or unified labels.
  4. [Introduction and Experiments] There are typos such as 'multimdoal' in the Introduction and 'effecitvely' in the Experiments section; a careful proofread is needed.
  5. [Figure 2] Figure 2 is dense and the loss labels (e.g., L_GP_T and L_GM_T) are nearly illegible at normal print size; please enlarge the figure or use a more legible layout.
  6. [Table 2] The ablation study is not uniformly supportive of the design: AproMFL w/o GP achieves higher R@5s (150.75%) on Flickr30k than full AproMFL (150.61%), and the Acc@1 gains on CIFAR-10 are only 0.02% over w/o GP. The text should discuss this rather than claim that global prototypes always help.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: AproMFL is an empirical consistency-regularization framework whose global prototypes are aggregates of local prototypes, not a derivation that reduces to its inputs.

full rationale

The paper does not claim to derive a prediction from first principles. Its central mechanism is algorithmic: local prototypes are constructed by label-guided averaging or clustering, the server aggregates them via semantic completion and multimodal clustering, and the resulting global prototypes and global model are used as regularizers (L_GPT, L_GMT) during local training. This is a feedback/consistency loop, but not a definitional one: the global prototypes are not plugged back into the same equation that defines them in a way that makes the output identical to the input by construction. The claimed gains are empirical comparisons against baselines, not derived quantities. No fitted parameter is renamed as a prediction; hyperparameters K and O are swept and reported, although reporting the best configuration across tables carries some test-set selection risk, which is a methodological concern rather than circularity. The use of pretrained CLIP encoders is an external, pretrained component; it tensions with the abstract's 'no reliance on prior knowledge' phrasing, but that is a claims-consistency issue, not a circular-derivation issue. Self-citations in the references (e.g., PraVFed) are not load-bearing for the main results. No self-citation chain, uniqueness import, or ansatz-smuggling via citation appears. Under the hard rules, no specific equation or passage exhibits a reduction of a claimed prediction to an input by construction, so the appropriate finding is no significant circularity.

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

The method is a combination of known machine learning components (prototypes, clustering, knowledge distillation, cosine similarity) applied to a new setting. There are no new physical entities or mediators. The primary free parameters are the prototype count K and the completion parameter O, both tuned on the test set. The domain assumptions center on the shared CLIP embedding space and the validity of prototype semantics.

free parameters (4)
  • K (number of global prototype pairs) = 10 (chosen to match best results in Table 3)
    The number of global prototype pairs is varied in Table 3 and K=10 is used in the main results (Table 1). It is selected based on test performance.
  • O (modality completion parameter) = 10 (chosen to match best results in Table 4)
    The number of multimodal prototypes used for semantic completion is varied in Table 4 and O=10 is used in the main results. It is selected based on test performance.
  • lambda (balance in LLMR) = not specified
    The regularization loss LLMR in Equation (6) uses a weight lambda that balances local personalization and global knowledge, but its value is not reported.
  • tau (temperature in contrastive losses) = not specified
    The temperature tau appears in Equations (3), (4), and (7), but no value is given.
assumptions (4)
  • domain assumption All clients share the same pretrained CLIP encoder, so embeddings are already in a common cross-modal semantic space.
    Implementation Details state 'We use the CLIP (Radford et al. 2021) ViT-L/14 as encoders for different clients'. The semantic completion and prototype clustering rely on cosine similarity in this shared space.
  • domain assumption K-means clustering of fused image-text embeddings produces clusters that correspond to meaningful semantic prototypes.
    The clustering-based local prototype construction in the methodology section assumes that fused embeddings form separable clusters, and that averaging image and text embeddings within each cluster yields valid prototype pairs.
  • domain assumption Cosine similarity between mapping module parameters reflects compatibility for model aggregation.
    The client relationship graph-based model aggregation uses s(i,j) = theta_i dot theta_j / (||theta_i|| ||theta_j||) as similarity, implicitly assuming that parameter similarity indicates that modules can be usefully averaged.
  • domain assumption The top-O most similar multimodal prototypes provide valid cross-modal counterparts for a unimodal prototype.
    Semantic completion in the server uses cosine similarity to select and weight multimodal prototypes, assuming these form correct image-text pairs for the unimodal prototype. This is not theoretically justified and may add noise when O is large.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Prototype Knowledge Transfer for Federated Learning with Mixed Modalities and Heterogeneous Tasks." pith.science (2026). https://pith.science/paper/VOCRC2FL

@misc{pith2026250204400,
  author       = {Pith},
  title        = {Pith review of: Adaptive Prototype Knowledge Transfer for Federated Learning with Mixed Modalities and Heterogeneous Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VOCRC2FL}},
  note         = {Machine review of arXiv:2502.04400}
}
abstract

Multimodal Federated Learning (MFL) with mixed modalities enables unimodal and multimodal clients to collaboratively train models while ensuring clients' privacy. As a representative sample of local data, prototypes offer an approach with low resource consumption and no reliance on prior knowledge for MFL with mixed modalities. However, existing prototype-based MFL methods assume unified labels across clients and identical tasks per client, which is impractical in MFL with mixed modalities. In this work, we propose an Adaptive prototype-based Multimodal Federated Learning (AproMFL) framework for mixed modalities to address the aforementioned issues. Our AproMFL transfers knowledge through adaptively-constructed prototypes without unified labels. Clients adaptively select prototype construction methods in line with labels; server converts client prototypes into unified multimodal prototypes and cluster them to form global prototypes. To address model aggregation issues in task heterogeneity, we develop a client relationship graph-based scheme to dynamically adjust aggregation weights. Furthermore, we propose a global prototype knowledge transfer loss and a global model knowledge transfer loss to enable the transfer of global knowledge to local knowledge. Experimental results show that AproMFL outperforms four baselines on three highly heterogeneous datasets ($\alpha=0.1$) and two heterogeneous tasks, with the optimal results in accuracy and recall being 0.42%~6.09% and 1.6%~3.89% higher than those of FedIoT (FedAvg-based MFL), respectively.

Figures

Figures reproduced from arXiv: 2502.04400 by the authors.

Figure 1
Figure 1. (a) The framework of MFL with homogeneous [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of AproMFL. AproMFL consists of three modules: Adaptive local prototype construction builds [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Distribution of client representations under the Flickr30k dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Average accuracy and recall of the model under [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 31 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  4. [4]

    Bao, G.; Zhang, Q.; Miao, D.; Gong, Z.; Hu, L.; Liu, K.; Liu, Y.; and Shi, C. 2024. Multimodal federated learning with missing modality via prototype mask and contrast. In nternational Conference on Machine Learning. Vienna, Austria

    work page 2024

  5. [5]

    Chen, H.; Zhang, Y.; Krompass, D.; Gu, J.; and Tresp, V. 2024. Fed DAT : An approach for foundation model finetuning in multi-modal heterogeneous federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, 11285--11293. Vancouver, Canada

    work page 2024

  6. [6]

    Chen, J.; and Zhang, A. 2022. Fed MS plit: Correlation-adaptive federated multi-task learning across multimodal split networks. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, 87--96. Washington, DC, USA

    work page 2022

  7. [7]

    Dai, Y.; Chen, Z.; Li, J.; Heinecke, S.; Sun, L.; and Xu, R. 2023. Tackling data heterogeneity in federated learning with class prototypes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 7314--7322

    work page 2023

  8. [8]

    Feng, T.; Bose, D.; Zhang, T.; Hebbar, R.; Ramakrishna, A.; Gupta, R.; Zhang, M.; Avestimehr, S.; and Narayanan, S. 2023. Fedmultimodal: A benchmark for multimodal federated learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 4035--4045. Long Beach, CA, USA

    work page 2023

Show all 39 references
  1. [9]

    Healy, J.; and McInnes, L. 2024. Uniform manifold approximation and projection. Nature Reviews Methods Primers, 4(1): 82

    2024

  2. [10]

    H.; Qi, H.; and Brown, M

    Hsu, T.-M. H.; Qi, H.; and Brown, M. 2019. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, pp(99)

    2019 arXiv

  3. [11]

    Huang, W.; Ye, M.; Shi, Z.; Li, H.; and Du, B. 2023. Rethinking federated learning with domain shift: A prototype view. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 16312--16322. Vancouver, BC, Canada

    2023

  4. [12]

    Krizhevsky, A.; Hinton, G.; et al. 2009. Learning multiple layers of features from tiny images

    2009

  5. [13]

    Q.; Thwal, C

    Le, H. Q.; Thwal, C. M.; Qiao, Y.; Tun, Y. L.; Nguyen, M. N.; and Hong, C. S. 2024. Cross-Modal Prototype based Multimodal Federated Learning under Severely Missing Modality. arXiv preprint arXiv:2401.13898, pp(99): 1--12

    2024 arXiv

  6. [14]

    Li, J.; Li, F.; Zhu, L.; Cui, H.; and Li, J. 2023 a . Prototype-guided knowledge transfer for federated unsupervised cross-modal hashing. In Proceedings of the 31st ACM International Conference on Multimedia, 1013--1022. Ottawa, ON, Canada

    2023

  7. [15]

    Li, M.; Tang, X.; Chen, S.; Weng, Y.; Peng, L.; and Yang, W. 2023 b . Exploring the Impact of Non-IID on Federated Learning. In 2023 International Conference on Blockchain Technology and Information Security, 159--167. Xi'an, China

    2023

  8. [16]

    Li, Q.; He, B.; and Song, D. 2021. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10713--10722. virtual

    2021

  9. [17]

    K.; Zaheer, M.; Sanjabi, M.; Talwalkar, A.; and Smith, V

    Li, T.; Sahu, A. K.; Zaheer, M.; Sanjabi, M.; Talwalkar, A.; and Smith, V. 2020. Federated optimization in heterogeneous networks. In Proceedings of Machine learning and systems, 429--450. Austin, TX, USA

    2020

  10. [18]

    Li, Z.; Hou, Z.; Liu, H.; Li, T.; Yang, C.; Wang, Y.; Shi, C.; Xie, L.; Zhang, W.; Xu, L.; et al. 2024. Federated Learning in Large Model Era: Vision-Language Model for Smart City Safety Operation Management. In Companion Proceedings of the ACM on Web Conference 2024, 1578--15...

    2024

  11. [19]

    McMahan, B.; Moore, E.; Ramage, D.; Hampson, S.; and y Arcas, B. A. 2017. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, 1273--1282. Fort Lauderdale, FL, USA

    2017

  12. [20]

    Peng, Y.; Bian, J.; and Xu, J. 2024. Fed mm : Federated Multi-Modal Learning with Modality Heterogeneity in Computational Pathology. In ICASSP 2024-2024 IEEE International Conference on Acoustics, 1696--1700. Seoul, Republic of Korea

    2024

  13. [21]

    R.; Gyawali, P.; and Bhattarai, B

    Poudel, P.; Shrestha, P.; Amgain, S.; Shrestha, Y. R.; Gyawali, P.; and Bhattarai, B. 2024. CAR-MFL : Cross-Modal Augmentation by Retrieval for Multimodal Federated Learning with Missing Modalities. In International Conference on Medical Image Computing and Computer-Assisted I...

    2024

  14. [22]

    Qi, F.; and Li, S. 2024. Adaptive Hyper-graph Aggregation for Modality-Agnostic Federated Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12312--12321. Seattle, WA, USA

    2024

  15. [23]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. Virtual Event

    2021

  16. [24]

    Sun, G.; Mendieta, M.; Dutta, A.; Li, X.; and Chen, C. 2024. Towards Multi-modal Transformers in Federated Learning. In European Conference on Computer Vision, 229--246

    2024

  17. [25]

    Wang, H.; Yurochkin, M.; Sun, Y.; Papailiopoulos, D.; and Khazaeni, Y. 2020 a . Federated learning with matched averaging. In 8th International Conference on Learning Representations, 1. Addis Ababa, Ethiopia

    2020

  18. [26]

    Wang, J.; Liu, Q.; Liang, H.; Joshi, G.; and Poor, H. V. 2020 b . Tackling the objective inconsistency problem in heterogeneous federated optimization. In Advances in neural information processing systems, 7611--7623. virtual

    2020

  19. [27]

    Wang, S.; Gai, K.; Yu, J.; Zhang, Z.; and Zhu, L. 2025. PraVFed: Practical Heterogeneous Vertical Federated Learning via Representation Learning. IEEE Transactions on Information Forensics and Security

    2025

  20. [28]

    Wang, Y.; Fu, H.; Kanagavelu, R.; Wei, Q.; Liu, Y.; and Goh, R. S. M. 2024. An aggregation-free federated learning for tackling data heterogeneity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 26233--26242

    2024

  21. [29]

    Xiong, B.; Yang, X.; Song, Y.; Wang, Y.; and Xu, C. 2023. Client-Adaptive Cross-Model Reconstruction Network for Modality-Incomplete Multimodal Federated Learning. In Proceedings of the 31st ACM International Conference on Multimedia, 1241--1249

    2023

  22. [30]

    Yan, K.; Cui, S.; Wuerkaixi, A.; Zhang, J.; Han, B.; Niu, G.; Sugiyama, M.; and Zhang, C. 2024. Balancing Similarity and Complementarity for Federated Learning. In Proceedings of the 41st International Conference on Machine Learning, 55739--55758. Vienna, Austria

    2024

  23. [31]

    Yazdinejad, A.; Dehghantanha, A.; Karimipour, H.; Srivastava, G.; and Parizi, R. M. 2024. A robust privacy-preserving federated learning model against model poisoning attacks. IEEE Transactions on Information Forensics and Security, 19: 6693--6708

    2024

  24. [32]

    Young, P.; Lai, A.; Hodosh, M.; and Hockenmaier, J. 2014. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2: 67--78

    2014

  25. [33]

    Yu, Q.; Liu, Y.; Wang, Y.; Xu, K.; and Liu, J. 2023. Multimodal federated learning via contrastive representation ensemble. In The Eleventh International Conference on Learning Representations. Kigali, Rwanda

    2023

  26. [34]

    Zhang, J.; Liu, Y.; Hua, Y.; and Cao, J. 2024. Fed TGP : Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning. In Proceedings of the AAAI conference on artificial intelligence, volume 38, 16768--16776

    2024

  27. [35]

    Zhang, X.; Zhao, J.; and LeCun, Y. 2015. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems, 649--657. Montreal, Quebec, Canada

    2015

  28. [36]

    Zhao, Y.; Barnaghi, P.; and Haddadi, H. 2022. Multimodal federated learning on iot data. In 2022 IEEE/ACM Seventh International Conference on Internet-of-Things Design and Implementation, 43--54

    2022

  29. [37]

    Zhou, Y.; Qu, X.; You, C.; Zhou, J.; Tang, J.; Zheng, X.; Cai, C.; and Wu, Y. 2025. Fedsa: A unified representation learning via semantic anchors for prototype-based federated learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 23009--23017

    2025

  30. [38]

    Zong, L.; Xie, Q.; Zhou, J.; Wu, P.; Zhang, X.; and Xu, B. 2021. FedCMR : Federated cross-modal retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, 1672--1676. Virtual Event, Canada

    2021

  31. [39]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence '...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.