Pith. sign in

REVIEW 4 major objections 6 minor 69 references

PM-MOE: Mixture of Experts on Private Model Parameters for Personalized Federated Learning

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 This paper proposes PM-MoE, a two-stage framework that lets each client blend the converged personalized model parameters of other clients through a gating network plus an energy-based denoiser, and reports consistent accuracy gains…

desk verdict A plausible plug-in personalization idea with broad experiments, but the main proof has a false step and the privacy claim contradicts the method's own Phase 2 design. read the letter →

arxiv 2502.00354 v1 pith:IJIX7OHD submitted 2025-02-01 cs.LG cs.AIcs.CR

Yu Feng , Yangli-ao Geng , Yifan Zhu , Zongfu Han , Xie Yu , Kaiwen Xue , Haoran Luo , Mengyang Sun
show 2 more authors
Guangwei Zhang Meina Song
This is my paper · ORCID
classification cs.LGcs.AIcs.CR
keywords personalizedfederatedlearningmixtureofexpertsenergy-baseddenoisingmodelsplittinggatenetworkparametersharingtop-kselection
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 aims to show that in personalized federated learning, a client's own locally trained personalized parameters are not the only useful resource: the converged personalized parameters of other clients contain domain knowledge that can help, provided they are selected and weighted per input. PM-MoE is a two-stage framework that first trains any model-split-based PFL method to convergence, then collects everyone's personalized parameters into a pool, lets each client learn a small gating network to mix the most relevant of those parameters with its own, and uses an energy-based denoising step to discard parameters that would hurt. If the central claim is right, then personalized federated learning can be improved without changing the base algorithm or adding much training cost, by reusing the private parameters that would otherwise sit unused.

What carries the argument

The two central mechanisms are the Mixture of Personalized Modules (MPM) and the Energy-based Denoising Method (EDM). MPM handles two cases: for local personalized parameters (PP), it computes a weighted sum of the top-k pooled parameters; for personalized experts (PE), it computes each expert's output feature and mixes them with top-k gate weights. EDM builds an energy function comparing each other client's projected representation to the local client's anchor representation, then uses negative Helmholtz free energy as a confidence score, removing the lowest gamma fraction of experts before the gate training, which prevents the gate from being optimized in the wrong direction by noisy modules. The gating network is the only trainable part in the second stage; the pooled modules are frozen.

What would settle it

Run a membership-inference attack (or a model-inversion attack) on a PM-MoE deployment where the server broadcasts the full personalized parameter pool to all clients. If an adversary can reliably determine whether a specific data point was in a victim client's training set from those parameters, the paper's central privacy claim is false. A second, accuracy-focused falsifier would be to construct a heterogeneity regime where the gate must route to an expert whose data distribution heavily overlaps the local one while the denoiser removes the best expert; if the bound's monotonicity in M fails there, the theoretical lower bound would be contradicted by experiment.

Watch

Extended reading notes

Core claim

The central claim is that exchanging converged personalized modules among clients is beneficial and can be done safely and cheaply, contrary to the usual PFL design where personalized parameters stay strictly local and are never shared. The paper introduces a generic wrapper, PM-MoE, that works on top of any model-split-based PFL method: after the base method converges, the server collects the personalized feature extractors and personalized parameters from all clients into two pools; each client then trains a small gating network to top-k select and weight these pooled modules given an input, and an energy-based denoising module computes per-module confidence scores using negative Helmholtz free energy against the client's own anchor representation, dropping a bottom fraction of experts as noise. The paper reports that applying this wrapper to nine PFL methods across six datasets gives consistent small accuracy improvements (average gains roughly from 0.02 to 0.9 points), and provides a theoretical lower bound (Theorem 3.1) stating that if each expert has accuracy p and the gate assigns samples to correct experts with probability ratio 1+alpha, then the mixed expert accuracy is at least ((1+alpha)p)/(1+alpha(p+(1-p)/M)), which is strictly above p.

Load-bearing premise

The framework's privacy guarantee rests on the assumption that broadcasting every client's converged personalized model parameters to every other client reveals no information about that client's training data; if model inversion or membership inference can extract that information, the privacy framing collapses even if the accuracy numbers still hold.

Editorial extensions

If this is right

  • PM-MoE can be dropped onto existing model-split PFL algorithms without changing their training, giving small but consistent accuracy gains at the cost of only a tiny fine-tuning phase.
  • The framework extends naturally to any number of clients because the gate only sees a fixed-size pooled set, and the theoretical bound says the per-client gain grows with the number of clients M.
  • If the gate is well trained and the pool is large, the bound approaches 100% accuracy, so the framework is especially promising in settings with many clients that share some data categories.
  • The energy-based denoising step is shown in the ablation to add a meaningful improvement over plain MoE mixing (e.g., average gain on three datasets rises from 0.19% to 0.53%).
  • Because only the gating network parameters are optimized in the fine-tuning stage, the extra computation is minimal (the paper reports roughly 0.4% of the pre-training cost for one case).

Reading between the lines

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

  • A natural extension the authors do not discuss is whether the same pooling and gating idea transfers to non-model-split PFL methods that have no explicit private parameters, since the gain depends on having a clear notion of a personalizable module to pool.
  • The energy-based denoising criterion uses the local client as a fixed anchor; an untested variant would be to let the anchor itself be learned or to use a second-order similarity metric, which might adapt better when the local model is itself noisy.
  • The theoretical bound treats each expert as independently correct with probability p, but in real FL settings the experts are correlated because they share the global backbone; a more realistic bound would likely be tighter and could predict when the improvements should shrink.
  • A direct test of the privacy premise would be to run a membership-inference or model-inversion attack against the broadcast personalized parameters; if such attacks succeed, the paper's claim that sharing these parameters is privacy-preserving would need qualification even though the accuracy gains may still hold.
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 PM-MoE, a two-stage post-hoc wrapper for model-split personalized federated learning. After a base PFL method converges, the server collects all clients' personalized parameters and expert models into a pool and broadcasts that pool to every client; each client then locally trains a small gating network to mix the top-k external experts and uses an energy-based score to drop supposedly noisy experts. The wrapper is applied to nine existing model-split PFL algorithms and evaluated on six datasets under two Dirichlet heterogeneity settings, with the reported result being consistent but very small accuracy gains. A theoretical section (Theorem 3.1) claims a lower bound on the accuracy of the mixture under an assumed gate-assignment ratio, and Appendix A.4 claims the protocol preserves privacy because clients cannot infer training data from model parameters.

Significance. If the empirical gains are reproducible and statistically meaningful, the idea of exchanging converged personalized modules through a lightweight second-stage gating network is simple and potentially useful as a plug-in for model-split PFL methods. The paper's strengths are breadth—nine baselines, six datasets, two heterogeneity settings—and the public code release. However, the current evidence is not sufficient: the supporting theorem contains a false algebraic identity as printed, the privacy premise is unsupported and internally contradicted by the protocol, and the headline improvements in the main tables are frequently below 0.1 percentage points with no error bars or significance tests. These issues bear directly on the paper's central claims, so the paper needs substantive revision before the claims can be accepted.

major comments (4)
  1. [Section 3.4, Eq. (16) and Appendix A.1, Eq. (23)] The algebraic identity used in the proof of Theorem 3.1 is false as printed. Eq. (16) asserts that (1+alpha)^s / ((1+alpha)^s + M - s) equals (1+alpha)^s / (M + alpha s), which would require (1+alpha)^s + M - s = M + alpha s; this fails for general s and alpha (e.g., s=2, alpha=2 gives 9 + M - 2 vs. M + 4). The intended expression is almost certainly s(1+alpha) / (s(1+alpha) + M - s) = s(1+alpha) / (M + alpha s), since the correct-expert/incorrect-expert assignment ratio of 1+alpha gives total weight s(1+alpha) on correct experts. With that correction, the subsequent summation leading to Eq. (24) goes through. As printed, however, the proof of the theorem is invalid, and because Theorem 3.1 is the only formal support for the claim that MPE outperforms a single client, this error is load-bearing and must be fixed.
  2. [Appendix A.4 (privacy analysis) and Section 3.2 (Phase 2)] The privacy claim is not supported and is internally contradicted by the protocol. Phase 1 (Section 3.1) says personalized parameters are 'computed locally and not uploaded,' but Phase 2 (Section 3.2) has the server collect all converged personalized parameters and experts into a pool and synchronize that pool with all clients. Thus every client receives the raw personalized modules of every other client. Appendix A.4's assertion that 'clients cannot infer the training data or other private information from the model parameters' is a bare claim with no threat model, no differential-privacy budget, no noise mechanism, and no discussion of model-inversion or membership-inference attacks, which are known to be able to leak information from model parameters. Since the paper frames its setting as privacy-preserving federated learning, this unsupported premise undermines the privacy contribution even if the accuracy numbers are correct.
  3. [Section 4, Tables 1-3] The empirical evidence for the central 'consistent improvement' claim is statistically weak. Table 2 reports no error bars, no number of seeds, and no significance tests, and many of the gains are tiny: FedPer +PM-MoE averages +0.0275 points, FedRep +PM-MoE averages +0.0258 points, and several individual cells improve by only 0.01-0.03 points. These differences are plausibly within run-to-run noise, so the claim of improvement across nine methods is not established. The paper should report standard deviations or confidence intervals over multiple runs, significance tests, and ideally per-client accuracy distributions, especially for the largest gains such as DBE (+0.9033) and GPFL (+0.5208), which may be driven by particular datasets.
  4. [Section 3.4, Theorem 3.1] The theorem's conclusion is largely baked into its assumption. The proof assumes a trained gate assigns samples to correct versus incorrect experts in ratio 1+alpha with alpha>0; under that assumption the lower bound necessarily exceeds p, so the theorem quantifies a conditional guarantee rather than proving that the proposed MPM/EDM training procedure produces such a gate. The paper does not show that the learned gating network satisfies this ratio condition, nor does it connect alpha to the top-k selection or the energy-based dropout ratio gamma. The authors should state this limitation explicitly and, ideally, provide a direct analysis of the gating network's assignment behavior.
minor comments (6)
  1. [Throughout] There are numerous typos and inconsistent notations, including 'leanring' in the introduction, 'which which' in Section 3.1, inconsistent 'PM-MOE' / 'PM-MoE' spellings, and 'denosing' in Section 3.3; the paper should be carefully copy-edited.
  2. [Section 3.2, around Eqs. (5)-(6)] The text refers to 'formulas (1) and (2)' when describing the gating weight computations, but those equations are numbered (5) and (6); the cross-reference should be corrected.
  3. [Section 3.3, Eq. (13)] The objective loss is written as min over theta_PP, theta_PE of min_K sum_j ...; the symbol K is not defined and the subscript on the sum is unclear. It should presumably be a sum over the M clients.
  4. [Section 4.2 and Figures 5-8] Figures 5, 6, 7, and 8 are referenced in the text but are not included in the manuscript; the figures need to be added or the references removed.
  5. [Table 3] The baseline labeled 'pFL' in the ablation table is not identified; the text should specify which underlying method (e.g., LG-FedAvg or FedCP) is being ablated.
  6. [Appendix A.9] The section 'Concerns about time cost' reads like a response to a reviewer rather than a normal experimental-details section; its content (run times and percentages) should be integrated into Sections 4 or A.5 in a more formal way.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the empirical gains are measured against external baselines and Theorem 3.1 is a conditional bound, not a fitted prediction; self-citations are non-load-bearing.

full rationale

The paper's central empirical claim is self-contained: PM-MoE is applied to nine external PFL baselines on six public datasets, and the reported improvements are measured outcomes rather than quantities fitted from the same test labels. Theorem 3.1 is a conditional mathematical statement: it assumes a gate with a positive correct-versus-incorrect assignment ratio 1+alpha and derives, by Jensen's inequality, P_MPE >= ((1+alpha)p)/(1+alpha(p+(1-p)/M)) > p. The conclusion follows from the stated assumption by explicit algebra, and alpha is an assumed gate property, not a parameter fitted from the test data and then renamed as a prediction. The energy-based denoiser essentially repackages cosine-similarity scoring and logsumexp as 'Helmholtz free energy,' which is a novelty-framing concern rather than a circular reduction. Self-citations (e.g., references [14], [15], [48], and [58]) appear only in general ML context and related-work discussions, and they do not carry the load-bearing argument. Appendix A.4's assertion that 'clients cannot infer the training data or other private information from the model parameters' is an unsupported and internally strained privacy premise given that Phase 2 distributes all clients' converged personalized modules to every client, but that is a correctness and security risk, not a circular derivation chain. No equation in the paper reduces to its own input by construction, and no fitted parameter is presented as a prediction.

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

The central result relies on hyperparameters tuned per benchmark (gamma, top-k, gate size, learning rate), an idealized binomial ensemble model, and an unsupported privacy assumption.

free parameters (5)
  • gamma (dropout ratio) = 0.2
    Set to 0.2 in Section 4.2 ablation; controls how many experts are removed by energy denoising.
  • top-k = half of clients for S=0, all clients for S=20
    Figure 6 shows k tuned per heterogeneity setting; the final k is not fixed in the paper.
  • eta_moe (gating learning rate) = 0.1 or 0.5 recommended
    Figure 7 shows higher learning rates are preferred; the value is chosen by experiment.
  • E_moe (fine-tuning iterations) = 50
    Figure 8 shows performance decreases with more iterations, so 50 is used.
  • gating network architecture = 4 layers, LeakyReLU, Orthogonal initialization
    Figure 5 selects these hyperparameters empirically.
assumptions (4)
  • domain assumption Each client's personalized expert independently predicts correctly with probability p, with independent errors across experts.
    Theorem 3.1 models expert successes as i.i.d. Binomial(M, p); real clients are not independent and errors are correlated, so the bound is on an idealized ensemble.
  • domain assumption The gating network achieves a correct-to-incorrect assignment ratio of 1+alpha with alpha > 0.
    Theorem 3.1 assumes the gate is better than random; the paper does not measure alpha or prove the trained gate achieves it.
  • ad hoc to paper The energy function E_k(h_j, h_k) = -v_k[I], using a single dimension of the cosine projection, measures expert similarity.
    Section 3.3 introduces this energy; no principled justification for using one dimension, and the temperature T is not specified in experiments.
  • domain assumption Sharing converged personalized model parameters does not leak private information.
    Appendix A.4 assumes parameters alone reveal nothing; this is contradicted by model inversion and membership inference literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PM-MOE: Mixture of Experts on Private Model Parameters for Personalized Federated Learning." pith.science (2026). https://pith.science/paper/IJIX7OHD

@misc{pith2026250200354,
  author       = {Pith},
  title        = {Pith review of: PM-MOE: Mixture of Experts on Private Model Parameters for Personalized Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJIX7OHD}},
  note         = {Machine review of arXiv:2502.00354}
}
read the original abstract

Federated learning (FL) has gained widespread attention for its privacy-preserving and collaborative learning capabilities. Due to significant statistical heterogeneity, traditional FL struggles to generalize a shared model across diverse data domains. Personalized federated learning addresses this issue by dividing the model into a globally shared part and a locally private part, with the local model correcting representation biases introduced by the global model. Nevertheless, locally converged parameters more accurately capture domain-specific knowledge, and current methods overlook the potential benefits of these parameters. To address these limitations, we propose PM-MoE architecture. This architecture integrates a mixture of personalized modules and an energy-based personalized modules denoising, enabling each client to select beneficial personalized parameters from other clients. We applied the PM-MoE architecture to nine recent model-split-based personalized federated learning algorithms, achieving performance improvements with minimal additional training. Extensive experiments on six widely adopted datasets and two heterogeneity settings validate the effectiveness of our approach. The source code is available at \url{https://github.com/dannis97500/PM-MOE}.

Figures

Figures reproduced from arXiv: 2502.00354 by the authors.

Figure 1
Figure 1. Motivation of our study. (A) t-SNE graph shows the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall Architecture of Personalized Model parameters with Mixture of Experts [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Diagram of Mixture of Personalized Parameters. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Diagram of Mixture of Personalized Experts. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Results of Gating Network Parameters [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Impact of Top k in PM-MOE. • Gating Network Learning Rate Analysis: Keeping other vari￾ables constant, we set the learning rate of the gating network 𝜂𝑚𝑜𝑒 to 0.05, 0.1, and 0.5. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Impact of Gating Network Learning Rate in PM [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Impact of Training Epochs in PM-MOE. 4.3 Analysis of the combination of personalized federated learning and MOE For integrating MoE in personalized federated learning, both PFL￾MoE and FedMoE use gated networks to adjust the weight balance between local personalized mo…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

69 extracted references · 43 canonical work pages

  1. [1]

    Mohammed Adnan, Shivam Kalra, Jesse C Cresswell, Graham W Taylor, and Hamid R Tizhoosh. 2022. Federated learning and differential privacy for medical image analysis. Scientific reports 12, 1 (2022), 1953

    work page 2022

  2. [2]

    Mohamad M Al Rahhal, Yakoub Bazi, Reham Al-Dayil, Bashair M Alwadei, Nassim Ammour, and Naif Alajlan. 2022. Energy-based learning for open-set classification in remote sensing imagery. International Journal of Remote Sensing 43, 15-16 (2022), 6027–6037

    work page 2022

  3. [3]

    Morris Chang

    Mohammad Al-Rubaie and J. Morris Chang. 2016. Reconstruction Attacks Against Mobile-Based Continuous Authentication Systems in the Cloud. IEEE Trans. Inf. Forensics Secur. 11, 12 (2016), 2648–2663. https://doi.org/10.1109/TIFS.2016. 2594132

    work page doi:10.1109/tifs.2016 2016

  4. [4]

    Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh, and Sunav Choudhary. 2019. Federated Learning with Personalization Layers. CoRR abs/1912.00818 (2019)

    arXiv 2019

  5. [5]

    Nuria Rodríguez Barroso, Goran Stipcich, Daniel Jiménez-López, José Antonio Ruiz-Millán, Eugenio Martínez-Cámara, Gerardo González-Seco, María Victoria Luzón, Miguel Angel Veganzones, and Francisco Herrera. 2020. Federated Learn- ing and Differential Privacy: Software tools analysis, the Sherpa.ai FL framework and methodological guidelines for preserving ...

    work page doi:10.1016/j.inffus.2020.07.009 2020

  6. [6]

    Hong-You Chen and Wei-Lun Chao. 2022. On Bridging Generic and Person- alized Federated Learning for Image Classification. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net

    work page 2022

  7. [7]

    Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. 2017. A Downsampled Variant of ImageNet as an Alternative to the CIFAR datasets.CoRR abs/1707.08819 (2017). arXiv:1707.08819 http://arxiv.org/abs/1707.08819

    arXiv 2017

  8. [8]

    Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. 2021. Exploiting Shared Representations for Personalized Federated Learning. In Pro- ceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event (Proceedings of Machine Learning Research, Vol. 139) , Marina Meila and Tong Zhang (Eds.). PMLR...

    work page 2021

Show all 69 references
  1. [9]

    Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Shuming Ma, Zhifang Sui, and Furu Wei. 2022. Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers. arXiv preprint arXiv:2212.10559 (2022)

    2022 arXiv

  2. [10]

    Lydia de la Torre. 2018. A guide to the california consumer privacy act of 2018. A vailable at SSRN 3275571(2018)

    2018

  3. [11]

    Dinh, Nguyen Hoang Tran, and Tuan Dung Nguyen

    Canh T. Dinh, Nguyen Hoang Tran, and Tuan Dung Nguyen. 2020. Personalized Federated Learning with Moreau Envelopes. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virt...

    2020

  4. [12]

    Yilun Du and Igor Mordatch. 2019. Implicit Generation and Modeling with Energy Based Models. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada , Han...

    2019

  5. [13]

    Ozdaglar

    Alireza Fallah, Aryan Mokhtari, and Asuman E. Ozdaglar. 2020. Personalized Fed- erated Learning with Theoretical Guarantees: A Model-Agnostic Meta-Learning Approach. In Advances in Neural Information Processing Systems 33: Annual Con- ference on Neural Information Processing S...

    2020

  6. [14]

    Jiawei Fan, Zhonghong Ou, Xie Yu, Junwei Yang, Shigeng Wang, Xiaoyang Kang, Hongxing Zhang, and Meina Song. 2022. Episodic projection network for out-of- distribution detection in few-shot learning. In 2022 26th International Conference on Pattern Recognition (ICPR) . IEEE, 3076–3082

    2022

  7. [15]

    Jiawei Fan, Yu Zhao, Xie Yu, Lihua Ma, Junqi Liu, Fangqiu Yi, and Boxun Li. 2022. DTR: An Information Bottleneck Based Regularization Framework for Video Action Recognition. In Proceedings of the 30th ACM International Conference on Multimedia. 3877–3885

    2022

  8. [16]

    Binbin Guo, Yuan Mei, Danyang Xiao, and Weigang Wu. 2021. PFL-MoE: Person- alized Federated Learning Based on Mixture of Experts. In Web and Big Data - 5th International Joint Conference, APWeb-W AIM 2021, Guangzhou, China, August 23-25, 2021, Proceedings, Part I (Lecture Note...

    2021

  9. [17]

    Andrew Hard, Kanishka Rao, Rajiv Mathews, Françoise Beaufays, Sean Augen- stein, Hubert Eichner, Chloé Kiddon, and Daniel Ramage. 2018. Federated Learn- ing for Mobile Keyboard Prediction.CoRR abs/1811.03604 (2018). arXiv:1811.03604 http://arxiv.org/abs/1811.03604

    2018 arXiv

  10. [18]

    Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al . 2012. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. I...

    2012

  11. [19]

    Muhammad Akbar Husnoo, Adnan Anwar, Nasser Hosseinzadeh, Shama Naz Islam, Abdun Naser Mahmood, and Robin Doss. 2022. Fedrep: Towards horizontal federated load forecasting for retail energy providers. In 2022 IEEE PES 14th Asia- Pacific Power and Energy Engineering Conference (...

    2022

  12. [20]

    Jacobs, Michael I

    Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton

  13. [21]

    Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomás Mikolov. 2017. Bag of Tricks for Efficient Text Classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2017, Valencia, Spain, April 3-7, 2017...

    2017

  14. [22]

    Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista A

    Peter Kairouz, H. Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista A. Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, Rafael G. L. D’Oliveira, Hubert Eichner, Salim El Rouayheb, David Evans, Josh Gardner, Zachary Garrett,...

    2021

  15. [23]

    Reddi, Se- bastian U

    Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J. Reddi, Se- bastian U. Stich, and Ananda Theertha Suresh. 2020. SCAFFOLD: Stochastic Controlled Averaging for Federated Learning. In Proceedings of the 37th Interna- tional Conference on Machine Learning, ICML 202...

    2020

  16. [24]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  17. [25]

    Alex Krizhevsky, Geoffrey Hinton, et al. 2009. Learning multiple layers of features from tiny images. (2009)

    2009

  18. [26]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet classifi- cation with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)

    2012

  19. [27]

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. nature 521, 7553 (2015), 436–444

    2015

  20. [28]

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient- based learning applied to document recognition. Proc. IEEE 86, 11 (1998), 2278– 2324

    1998

  21. [29]

    Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, Fujie Huang, et al. 2006. A tutorial on energy-based learning. Predicting structured data 1, 0 (2006)

    2006

  22. [30]

    Qinbin Li, Bingsheng He, and Dawn Song. 2021. Model-Contrastive Federated Learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. Computer Vision Foundation / IEEE, 10713–10722

    2021

  23. [31]

    Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. 2021. Ditto: Fair and Robust Federated Learning Through Personalization. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event (Proceedings of Machine Learnin...

    2021

  24. [32]

    Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. 2021. FedBN: Federated Learning on Non-IID Features via Local Batch Normalization. arXiv:2102.07623 [cs.LG] https://arxiv.org/abs/2102.07623

    2021 arXiv

  25. [33]

    Xin-Chun Li, De-Chuan Zhan, Yunfeng Shao, Bingshuai Li, and Shaoming Song

  26. [34]

    Paul Pu Liang, Terrance Liu, Ziyin Liu, Ruslan Salakhutdinov, and Louis-Philippe Morency. 2020. Think Locally, Act Globally: Federated Learning with Local and Global Representations. CoRR abs/2001.01523 (2020). arXiv:2001.01523 http://arxiv.org/abs/2001.01523

    2020 arXiv

  27. [35]

    Stich, and Martin Jaggi

    Tao Lin, Lingjing Kong, Sebastian U. Stich, and Martin Jaggi. 2020. Ensemble Distillation for Robust Model Fusion in Federated Learning. InAdvances in Neural Information Processing Systems 33: Annual Conference on Neural Information Pro- cessing Systems 2020, NeurIPS 2020, Dec...

    2020

  28. [36]

    Owens, and Yixuan Li

    Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. 2020. Energy-based Out-of-distribution Detection. In Advances in Neural Infor- mation Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , H...

    2020

  29. [37]

    Yang Liu, Anbu Huang, Yun Luo, He Huang, Youzhi Liu, Yuanyuan Chen, Lican Feng, Tianjian Chen, Han Yu, and Qiang Yang. 2020. FedVision: An Online Visual Object Detection Platform Powered by Federated Learning. In The Thirty- Fourth AAAI Conference on Artificial Intelligence, A...

    2020

  30. [38]

    Bing Luo, Pengchao Han, Peng Sun, Xiaomin Ouyang, Jianwei Huang, and Ningn- ing Ding. 2023. Optimization Design for Federated Learning in Heterogeneous 6G Networks. IEEE Netw. 37, 2 (2023), 38–43

    2023

  31. [39]

    Jun Luo and Shandong Wu. 2022. Adapt to Adaptation: Learning Personalization for Cross-Silo Federated Learning. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, Luc De Raedt (Ed.). ijcai...

    2022 doi

  32. [40]

    Saeed Masoudnia and Reza Ebrahimpour. 2014. Mixture of experts: a literature survey. Artificial Intelligence Review 42 (2014), 275–293

    2014

  33. [41]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. 2017. Communication-Efficient Learning of Deep Net- works from Decentralized Data. InProceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, ...

    2017

  34. [42]

    Jaehoon Oh, Sangmook Kim, and Se-Young Yun. 2022. FedBABU: Toward En- hanced Representation for Federated Image Classification. In The Tenth Inter- national Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net

    2022

  35. [43]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems 35...

    2022

  36. [44]

    Protection Regulation. 2016. Regulation (EU) 2016/679 of the European Parliament and of the Council. Regulation (eu) 679 (2016), 2016

    2016

  37. [45]

    Hyowoon Seo, Jihong Park, Seungeun Oh, Mehdi Bennis, and Seong-Lyun Kim. [n. d.]. 16 federated knowledge distillation. ([n. d.])

  38. [46]

    Tao Shen, Jie Zhang, Xinkang Jia, Fengda Zhang, Gang Huang, Pan Zhou, Fei Wu, and Chao Wu. 2020. Federated Mutual Learning. CoRR abs/2006.16765 (2020). arXiv:2006.16765 https://arxiv.org/abs/2006.16765

    2020 arXiv

  39. [47]

    Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. 2022. FedProto: Federated Prototype Learning across Heteroge- neous Clients. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative App...

    2022

  40. [48]

    Fengrui Tian, Jiawei Fan, Xie Yu, Shaoyi Du, Meina Song, and Yu Zhao. 2022. TCVM: Temporal Contrasting Video Montage Framework for Self-supervised Video Representation Learning. In Proceedings of the Asian Conference on Com- puter Vision. 1539–1555

    2022

  41. [49]

    Michalis K Titsias and Aristidis Likas. 2002. Mixture of experts classification using a hierarchical mixture model. Neural Computation 14, 9 (2002), 2221–2244

    2002

  42. [50]

    Praveen Venkateswaran, Vatche Isahagian, Vinod Muthusamy, and Nalini Venkatasubramanian. 2023. Fedgen: Generalizable federated learning for se- quential data. In 2023 IEEE 16th International Conference on Cloud Computing (CLOUD). IEEE, 308–318

    2023

  43. [51]

    Yabin Wang, Zhiheng Ma, Zhiwu Huang, Yaowei Wang, Zhou Su, and Xiaopeng Hong. 2023. Isolation and Impartial Aggregation: A Paradigm of Incremental Learning without Interference. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on...

    2023

  44. [52]

    Chuhan Wu, Fangzhao Wu, Lingjuan Lyu, Yongfeng Huang, and Xing Xie. 2022. Communication-efficient federated learning via knowledge distillation. Nature communications 13, 1 (2022), 2032

    2022

  45. [53]

    Han Xiao, Kashif Rasul, and Roland Vollgraf. 2017. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms. CoRR abs/1708.07747 (2017). arXiv:1708.07747 http://arxiv.org/abs/1708.07747

    2017 arXiv

  46. [54]

    Jian Xu, Xinyi Tong, and Shao-Lun Huang. 2023. Personalized Federated Learning with Feature Alignment and Classifier Collaboration. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net

    2023

  47. [55]

    Qiang Yang, Yang Liu, Yong Cheng, Yan Kang, Tianjian Chen, and Han Yu. 2019. Federated Learning. Morgan & Claypool Publishers. https://doi.org/10.2200/ S00960ED2V01Y201910AIM043

    2019

  48. [56]

    Liping Yi, Gang Wang, Xiaoguang Liu, Zhuan Shi, and Han Yu. 2023. FedGH: Heterogeneous Federated Learning with Generalized Global Header. In Proceed- ings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023- 3 November 2023 , Ab...

    2023

  49. [57]

    Liping Yi, Han Yu, Chao Ren, Heng Zhang, Gang Wang, Xiaoguang Liu, and Xiaoxiao Li. 2024. FedMoE: Data-Level Personalization with Mixture of Ex- perts for Model-Heterogeneous Personalized Federated Learning. arXiv preprint arXiv:2402.01350 (2024)

    2024 arXiv

  50. [58]

    Xie Yu and Wentao Zhang. 2024. Anchor-Based Masked Generative Distillation for Pixel-Level Prediction Tasks. In 35th British Machine Vision Conference 2024, BMVC 2024, Glasgow, UK, November 25-28, 2024. BMVA. https://papers.bmvc2024. org/0365.pdf

    2024

  51. [59]

    Jianqing Zhang, Yang Hua, Jian Cao, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. 2023. Eliminating Domain Bias for Federated Learning in Representation Space. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Proce...

    2023

  52. [60]

    Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, Jian Cao, and Haibing Guan. 2023. GPFL: Simultaneously Learning Global and Per- sonalized Feature Information for Personalized Federated Learning. In IEEE/CVF International Conference on Computer Vision, ICCV...

    2023

  53. [61]

    Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. 2023. FedALA: Adaptive Local Aggregation for Personalized Federated Learning. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative A...

    2023

  54. [62]

    Jianqing Zhang, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. 2023. FedCP: Separating Feature Information for Personalized Federated Learning via Conditional Policy. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining...

    2023

  55. [63]

    Jianqing Zhang, Yang Liu, Yang Hua, Hao Wang, Tao Song, Zhengui Xue, Ruhui Ma, and Jian Cao. 2023. PFLlib: Personalized Federated Learning Algorithm Library. CoRR abs/2312.04992 (2023). https://doi.org/10.48550/ARXIV.2312.04992 arXiv:2312.04992

    2023 doi arXiv

  56. [64]

    Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015. Character-level Con- volutional Networks for Text Classification. In Advances in Neural Informa- tion Processing Systems 28: Annual Conference on Neural Information Process- ing Systems 2015, December 7-12, 2015, Montreal, Qu...

    2015

  57. [65]

    Zhao, Andrew M

    Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Y. Zhao, Andrew M. Dai, Zhifeng Chen, Quoc V. Le, and James Laudon. 2022. Mixture-of- Experts with Expert Choice Routing. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Inform...

    2022

  58. [66]

    fine" label (its specific class) and a

    Zhuangdi Zhu, Junyuan Hong, and Jiayu Zhou. 2021. Data-Free Knowledge Distillation for Heterogeneous Federated Learning. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event (Proceedings of Machine Learning Research...

    2021

  59. [1991]

    Neural Comput

    Adaptive Mixtures of Local Experts. Neural Comput. 3, 1 (1991), 79–87. https://doi.org/10.1162/NECO.1991.3.1.79

    1991 doi

  60. [2021]

    In Ma- chine Learning and Knowledge Discovery in Databases

    FedPHP: Federated Personalization with Inherited Private Models. In Ma- chine Learning and Knowledge Discovery in Databases. Research Track - European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part I (Lecture Notes in Computer Science, Vol....

    2021

  61. [2023]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 4015–4026

Pith tools

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