Pith. sign in

REVIEW 4 major objections 5 minor 25 references

Federated Learning Enhanced by Feature Reconstruction for Semantic Communication Module Updates of Agents

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

Pith's one-line read FedSFR's central claim is that letting weak-channel clients upload compact encoder features instead of model updates, and passing those features through decoder then encoder at the server, recovers compression loss and stabilizes…

desk verdict Useful extension of the authors' own FR idea to digital VQ-based JSCC, with a clean DP variant, but the convergence advantage is loaded into Assumption 1, which the paper's own schedule makes implausible. read the letter →

arxiv 2508.03248 v3 pith:OI4YBQWQ submitted 2025-08-05 eess.SP

classification eess.SP
keywords semanticcommunicationfederatedlearningvectorquantizationjointsource-channelcodingfeaturereconstructioncompressionerrorfeedbackdifferentialprivacyimagetransmission
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 proposes FedSFR, a federated-learning protocol for vector-quantization (VQ)-based image semantic communication in which clients under poor channel conditions transmit compact encoder output features instead of full compressed model updates. At the parameter server, those features are passed through the decoder and then the encoder in reverse processing order, defining a feature-reconstruction (FR) step that refines the aggregated model. The central claim is that this server-side FR step cancels a large part of the compression error introduced by sparsification and quantization, making capacity-limited federated training stable and communication-efficient. The paper backs the claim with a VQ-aware FR loss, an $O(1/\sqrt{T})$ convergence analysis, and a differential-privacy variant in which perturbing only the encoder halves the privacy budget. If correct, semantic-communication agents can refresh their JSCC models under tight uplink budgets without sacrificing reconstruction quality.

What carries the argument

The carrying object is the server-side feature-reconstruction (FR) step, which minimizes the VQ-aware loss $l_s(w; Y_1) = \mathrm{MSE}(Y_2, Y_1) + \alpha_s \mathrm{MSE}(\hat{Y}_2, \mathrm{sg}(Y_2)) + \beta_s \mathrm{MSE}(Y_2, \mathrm{sg}(\hat{Y}_2))$, where $Y_2 = f_\theta(f^{-1}_\phi(\mathrm{sg}(\hat{Y}_1)))$ and both $\hat{Y}_1,\hat{Y}_2$ pass through a simulated AWGN channel and the global VQ codebook. The argument hinges on Assumption 1, which bounds $\|a-b\|^2/(\|a\|^2+\|b\|^2) \le \varepsilon$ for the aggregate compression error $a = \sum_k p_k m_k^{(t+1)}$ and the FR server update $b = \eta_s^{(t)} \sum_e \nabla F_s$, so the convergence bound's $\varepsilon$ terms control how much FR cancels compression error. The matching client-side mechanism is two-mode upload: compressed gradient with error feedback for good channels, compact encoder feature vectors for bad channels.

What would settle it

Run FedSFR on CIFAR-10 under the paper's settings ($K=50$, $K_m=K_o=10$, $S_m/D=0.2$, 4-bit QSGD) and record, across global iterations, the aggregate compression error $a=\sum_k p_k m_k^{(t+1)}$ and the server FR update $b=\eta_s^{(t)}\sum_e \nabla F_s$; compute $\|a-b\|^2/(\|a\|^2+\|b\|^2)$. If typical ratios are near 2 rather than well below 1, Assumption 1 fails and FedSFR's proved advantage over the compressed no-FR baseline should not appear in a matched-bit-budget PSNR comparison.

Watch

Extended reading notes

Core claim

FedSFR's core discovery is that the JSCC autoencoder's inverse structure can be reused at the server as a compression-error corrector. In each global round, clients with good channels upload top-$S$ sparsified, quantized gradient updates with error feedback, while clients with poor channels upload compact feature vectors produced by their locally updated encoders from a shared public dataset. The server aggregates the compressed updates and then runs server-side SGD steps that feed the received features through the decoder $f^{-1}_\phi$ and then the encoder $f_\theta$, in reverse of the client-side processing order, using a loss with three terms: feature-reconstruction MSE plus two noisy codebook-alignment terms evaluated after simulated AWGN transmissions through the global VQ codebook. The paper argues that this FR update is aligned with the aggregate compression error, so it recovers information lost by the compressed uploads; under that alignment assumption it proves convergence at $O(1/\sqrt{T})$, and by privatizing the encoder before feature extraction it obtains a local differential-privacy guarantee about twice as strong as privatizing the full compressed update.

Load-bearing premise

The load-bearing premise is that the server's feature-reconstruction update points nearly in the same direction as the aggregate compression error left by quantized client uploads, so the reconstruction step cancels that error rather than adding drift.

Editorial extensions

If this is right

  • Under capacity limits, some clients can replace model-update uploads with compact feature vectors, cutting uplink bits whenever the model size $D$ is much larger than the feature size $Nd$, while the server-side FR keeps the global model aligned.
  • FedSFR converges at $O(1/\sqrt{T})$ when client and server learning rates are set to $\eta_c^{(t)}=\alpha(t)/\sqrt{T}$ and $\eta_s^{(t)}=\alpha(t)/T^{3/4}$ with $\eta_s<\eta_c$, and smaller $\varepsilon$ in Assumption 1 yields faster convergence.
  • The FR loss trains through two simulated AWGN realizations and the global VQ codebook, so the server update adapts the decoder, encoder, and codebook as a single joint pipeline rather than updating only one side.
  • Encoder-only perturbation under oneshot plus Laplace mechanisms gives a local differential-privacy budget roughly half that of perturbing the full compressed update, so the same noise yields stronger privacy or matching privacy with less noise.
  • Experiments on CIFAR-10 and CelebA show stable PSNR growth where FedAvg-style weighting is unstable and knowledge-distillation baselines lag, with larger gaps at higher resolution.

Reading between the lines

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

  • Beyond the paper, the features-instead-of-gradients plus server reconstruction pattern should transfer to other autoencoder-based JSCC modalities such as audio, video, or point clouds whenever the decoder approximately inverts the encoder and the model size dwarfs the feature size.
  • A natural adaptive extension is to measure the Assumption 1 alignment online and route clients to feature-upload mode only when the measured $\varepsilon$ is small; the paper fixes the routing by channel quality instead.
  • The factor-two privacy gain is structural because only the encoder is privatized, so it should persist under other differentially private mechanisms whose sensitivity scales with parameter count, not just oneshot plus Laplace.
  • Because the FR loss simulates AWGN twice at the server, the quality of the server update is tied to the chosen training SNR rather than the feature-uploading client's actual uplink channel; a mismatched-SNR experiment would test whether that decoupling holds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes FedSFR, a federated learning framework for vector-quantization-based image semantic communication. In each global round, clients with favorable channels upload compressed local model updates, while clients with poor channels upload compact encoder-output feature vectors. The parameter server first aggregates the compressed updates and then performs a server-side feature-reconstruction (FR) step, passing the received feature vectors through the decoder and then the encoder in reverse order. The paper introduces an FR loss function, argues that FR loss is a surrogate for image reconstruction error, proves an O(1/sqrt(T)) convergence rate under a new alignment assumption (Assumption 1), derives a local differential-privacy guarantee that is factor-2 stronger for encoder-only perturbation, and reports experiments on CIFAR-10 and CelebA.

Significance. If the claims were fully established, FedSFR would be a practically relevant contribution to communication-efficient federated training of VQ-based JSCC systems. The algorithmic idea is clean, the DP analysis is a genuinely nice structural observation, and the experiments cover two datasets and several baselines. The paper also provides formal statements and proofs for convergence and privacy, which is a strength. However, the central convergence theorem rests on an unverified and, under the paper's own learning-rate schedules, implausible alignment assumption, and the experimental comparison is not iso-budget. These issues need to be resolved before the advertised theoretical and practical claims can be accepted.

major comments (4)
  1. [IV.A, Theorem 1, Appendix B] Assumption 1 is the load-bearing component of the convergence analysis, but it is not derived from the algorithm and appears inconsistent with the paper's own schedules. The assumption postulates ||a-b||^2/(||a||^2+||b||^2) <= eps for a = sum_k p_k m_k^(t+1) and b = eta_s^(t) sum_e grad F_s, i.e., that the server FR update is nearly aligned with the compression error. In Appendix B, this assumption is used exactly once, to bound E||u-v||^2 in (a.2.1), replacing it with eps(E||u||^2+E||v||^2). Without a small eps, the terms B and C in Theorem 1 do not separate FedSFR from the compressed no-FR baseline. Yet a is an aggregated error-feedback memory from compressed-update clients, while b is an FR gradient computed on public feature vectors from the disjoint subset A_o^(t); nothing in the algorithm couples these quantities. Moreover, under the schedules of Theorem 1, eta_s^(t)/eta_c^(t) = O(T^{-1/4}) -> 0, so b/a -> 0 and the ratio in Assumption 1 tends to 1, not to a small eps. The paper also invokes Assumption 1 at a time index shift that is not justified: the assumption is stated for m^(t+1) and eta_s^(t), but in Appendix B it is applied to u = sum p_k m_k^(t) and v = eta_s^(t-1) sum grad. The authors should either prove a bound on eps for the actual algorithm, or measure eps/cos(a,b) in experiments, or reformulate the theorem as a conditional statement that does not advertise an improved convergence rate.
  2. [III.C.2] The argument that minimizing FR loss improves image reconstruction is conditional on an uncontrolled constant. The bound ||Xhat - X||^2 <= 2L^2||Y - Yhat||^2 + cˇ is obtained with cˇ = 2L^2||Yhat - Ycheck||^2, but no upper bound, dependence on problem parameters, or uniformity argument is given for cˇ. Also, the first-order Taylor analysis in III.C.1 assumes the Jacobian elements of the encoder are i.i.d. standard Gaussian, an assumption that is neither justified nor tested. These issues do not by themselves invalidate the algorithm, but they mean the paper's stated 'validity as a surrogate for image reconstruction error' is not established. The claims in Sections I and III-C should be softened or the missing control on cˇ should be supplied.
  3. [V, 'Comparison with Baselines'] The experimental comparison is not iso-budget, so the communication-efficiency claim is not directly supported. FedSFR clients in A_o^(t) transmit 50 feature vectors with 4-bit uniform scalar quantization under So/D=0.1, while FedSFD transmits 50 pairs of encoder output features and intermediate decoder features, which the paper itself describes as a 'significantly higher' uplink communication cost; SKBS uses different sparsification levels (Sm/D=0.8, So/D=0.4) and a smaller server model. The per-round uplink bits of each method are never reported. To substantiate the central claim of 'more efficient utilization of communication resources', the authors should report the exact per-client per-round communication budget for every method and either equalize the budgets or present PSNR as a function of total transmitted bits.
  4. [V, 'Comparison with Baselines'] The experimental results are presented without error bars, confidence intervals, or information about the number of random seeds. The paper's claims emphasize 'training stability' and 'robustness', but a single run cannot establish these properties, especially when the compared baselines show visibly unstable PSNR trajectories. The authors should report mean and standard deviation over at least several seeds, or otherwise justify why the reported curves are representative.
minor comments (5)
  1. [III.A] In the paragraph after Eq. (11), the text says 'the PS aggregates the compressed local updates {bar g}_k for k in A_o^(t)', but the intended set is A_m^(t); the index is inconsistent with the notation defined in Table I.
  2. [III.B, Eq. (14)] The notation Y1 is used both for a single received feature vector and, in Eq. (15), for the set of received feature vectors; this should be disambiguated.
  3. [IV.B, proof of Theorem 2] In the proof of Theorem 2, the sentence 'the sensitivity on the encoder of Mos is given by Delta_1^(ii) = ||g_{k,theta} - g'_{k,theta}||_1 <= Q' is immediately followed by 'assuming a symmetric encoder-decoder architecture, ||g_{k,theta}||_1 <= Q/2 holds'. The role of the Q/2 bound is unclear, since the displayed sensitivity is Q, not Q/2. Please clarify the argument.
  4. [V, Figs. 2(b) and 3(b)] The legend labels in Figs. 2(b) and 3(b) give percentages whose meaning is only explained in the text; please define the 'improvement ratio' directly in the figure caption.
  5. [IV.B, Definition 1] The adjacency condition '||D - D'||_1 = 1' is nonstandard for datasets; the usual definition is that D and D' differ in exactly one record. Please rephrase to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 1 is conditional on an explicit assumption, and the FR-surrogate argument is a stated approximation, not a renamed input or fitted prediction.

full rationale

The central claims of FedSFR are not circular. The server-side FR loss is explicitly defined (Eq. 16), and the surrogate argument in Section III.C is a first-order Taylor/Lipschitz derivation with stated auxiliary assumptions (i.i.d. Gaussian Jacobian, Lipschitz decoder, uncontrolled constant č); it is an approximate justification, not a definitional equivalence. Theorem 1 is conditional on standard FL assumptions plus Assumption 1, which indeed postulates that the server FR term b approximately equals the aggregated compression error a. This is the weakest point of the paper: Assumption 1 directly supplies the mechanism by which FR compensates compression error, and Appendix B uses it exactly to bound E||u-v||^2. However, this is an unverified hypothesis and a conditional-theorem structure, not circularity: no parameter is fitted to data and renamed a prediction, no quantity is defined in terms of the result it is used to prove, and the paper explicitly frames the rate advantage as conditional on Assumption 1 being well satisfied. The prior self-citations ([15], [3]) describe earlier frameworks and the adopted VQ loss, but none is load-bearing for the convergence or DP conclusions, and the experimental comparisons to FedAvg, FedDMA, FedLol, SKBS, and FedSFD are external benchmarks. The concern that epsilon may not be small under the paper's learning-rate schedules is a correctness/validation risk, not a circularity, and under the hard rules it does not raise the circularity score.

Assumptions & free parameters 4 free parameters · 9 assumptions · 0 invented entities

The framework rests on standard FL assumptions (smoothness, unbiased gradients, bounded variance) plus several paper-specific premises. The most consequential is Assumption 1, which assumes the server FR update is nearly aligned with the global compression error; the convergence improvement in Theorem 1 follows directly from this assumption. The surrogate-loss proofs add a Lipschitz-continuity premise with an uncontrolled constant and an i.i.d. Gaussian Jacobian premise for the linear version. The DP advantage requires a symmetric encoder-decoder architecture to halve the sensitivity. No new physical entities are introduced.

free parameters (4)
  • Server learning rate η_s^(0) = 0.0001 (CIFAR-10), 0.001 (CelebA)
    Hand-tuned; Fig. 2(b) and 3(b) show the FR benefit strongly depends on it; performance degrades for 0.1 and 0.01, and the convergence condition requires η_s < η_c.
  • Sparsification ratios Sm/D and So/D = 0.2 and 0.1 (CIFAR-10); 0.78/0.39 used for the SKBS baseline
    Hand-chosen budget parameters that define the compression level, directly affecting the error memory m_k and the convergence bound via (K-Km)^2/Km^2; baselines are re-tuned to different ratios, complicating iso-budget comparison.
  • Number of feature vectors per client per round = 50 (CIFAR-10), 25 (CelebA)
    Chosen to satisfy the budget Bo < Bm; it sets the size of the server-side FR dataset D_s and hence the quality of the reconstruction step.
  • FR loss coefficients α_c=α_s=1, β_c=0.25α_c, β_s=0.25α_s = 1 and 0.25
    Hand-fixed, inherited from [3,16]; these constants define the objective being minimized and are not subject to sensitivity analysis in this paper.
assumptions (9)
  • ad hoc to paper Assumption 1: ||a-b||^2/(||a||^2+||b||^2) ≤ ε for global compression error a and server FR update b
    Postulates the FR benefit; used to prove Theorem 1 (Section IV.A).
  • standard math F is β-smooth and lower-bounded
    Assumption 2, standard in FL convergence proofs.
  • standard math Stochastic gradients are unbiased
    Assumption 3.
  • standard math Bounded gradient norms: E||∇F_k||^2 ≤ G_k^2, E||∇F_s||^2 ≤ G_s^2
    Assumption 4.
  • domain assumption Compression operator satisfies E||x-Compress(x)||^2 ≤ (1-ν)||x||^2
    Eq. (24), standard error-feedback assumption [12]; used for Lemma 3.
  • domain assumption Decoder f_ϕ^{-1} is L-Lipschitz
    Section III.C.2; needed for the surrogacy bound, no Lipschitz constant estimated.
  • ad hoc to paper Jacobian elements of encoder are i.i.d. standard Gaussian
    Section III.C.1; used only to get the linear surrogate relationship E[FR] ≈ Nd E[image error].
  • ad hoc to paper Symmetric encoder-decoder architecture implies ||g_θ||_1 ≤ Q/2
    Theorem 2; needed for the factor-2 DP advantage; no evidence that the 0.51M-parameter model is symmetric in practice.
  • domain assumption Clients have access to a shared public dataset P_k ⊂ D_k
    Section III.A; required for feature-vector clients to produce Y_k without leaking private data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Federated Learning Enhanced by Feature Reconstruction for Semantic Communication Module Updates of Agents." pith.science (2026). https://pith.science/paper/OI4YBQWQ

@misc{pith2026250803248,
  author       = {Pith},
  title        = {Pith review of: Federated Learning Enhanced by Feature Reconstruction for Semantic Communication Module Updates of Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OI4YBQWQ}},
  note         = {Machine review of arXiv:2508.03248}
}
read the original abstract

Recent advancements in semantic communication have primarily focused on image transmission, where neural network-based joint source-channel coding modules play a central role. However, such systems often experience semantic communication errors due to mismatched knowledge bases between agents and performance degradation from outdated models, necessitating regular model updates. To address these challenges in vector quantization (VQ)-based image semantic communication systems, we propose FedSFR, a novel federated learning framework that incorporates semantic feature reconstruction (FR). FedSFR introduces an FR step at the parameter server and allows a subset of clients to transmit compact feature vectors in lieu of sending full local model updates, thereby improving training stability and communication efficiency. To enable effective FR learning, we design a loss function tailored for VQ-based image semantic communication and demonstrate its validity as a surrogate for image reconstruction error. We further establish a rigorous convergence analysis of FedSFR. Experimental results on two benchmark datasets validate the superiority of FedSFR over existing baselines, especially in capacity-constrained settings, confirming both its effectiveness and robustness.

Figures

Figures reproduced from arXiv: 2508.03248 by the authors.

Figure 1
Figure 1. Overall procedure of FedSFR with the numbered algorithmic steps. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. PSNR of (a) the proposed scheme and the baselines, (b) the proposed scheme with varying learning rates, and (c) the proposed scheme and the [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. PSNR of (a) the proposed scheme and the baselines, (b) the proposed scheme with varying learning rates, and (c) the proposed scheme and the [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

  1. [1]

    Task- oriented communications for 6G: Vision, principles, and technologies,

    Y . Shi, Y . Zhou, D. Wen, Y . Wu, C. Jiang, and K. B. Letaief, “Task- oriented communications for 6G: Vision, principles, and technologies,” IEEE Wireless Communications, vol. 30, no. 3, pp. 78–85, 2023

  2. [2]

    Semantic communications: Overview, open issues, and future research directions,

    X. Luo, H.-H. Chen, and Q. Guo, “Semantic communications: Overview, open issues, and future research directions,” IEEE Wireless Communi- cations, vol. 29, no. 1, pp. 210–219, 2022. 5Most compression methods satisfy this assumption [12]

  3. [3]

    Universal joint source-channel coding for modulation-agnostic semantic communication,

    Y . Huh, H. Seo, and W. Choi, “Universal joint source-channel coding for modulation-agnostic semantic communication,” IEEE Journal on Selected Areas in Communications, vol. 43, no. 7, pp. 2560–2574, 2025

  4. [4]

    A multi-user deep semantic communication system based on federated learning with dynamic model aggregation,

    H. Xing, H. Zhang, X. Wang, L. Xu, Z. Xiao, B. Zhao, S. Luo, L. Feng, and Y . Dai, “A multi-user deep semantic communication system based on federated learning with dynamic model aggregation,” in 2023 IEEE International Conference on Communications Workshops (ICC WKSHPS). IEEE, 2023, pp. 1612–1616

  5. [5]

    An efficient federated learning framework for training semantic communication systems,

    L. X. Nguyen, H. Q. Le, Y . L. Tun, P. S. Aung, Y . K. Tun, Z. Han, and C. S. Hong, “An efficient federated learning framework for training semantic communication systems,” IEEE Transactions on Vehicular Technology, vol. 73, no. 10, pp. 15 872–15 877, 2024

  6. [6]

    Federated learning-based co- operative model training for task-oriented semantic communication,

    H. Sun, H. Tian, W. Ni, and J. Zheng, “Federated learning-based co- operative model training for task-oriented semantic communication,” in IEEE INFOCOM 2024-IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS). IEEE, 2024, pp. 1–6

  7. [7]

    Efficient knowledge base synchro- nization in semantic communication network: A federated distillation approach,

    X. Lu, K. Zhu, J. Li, and Y . Zhang, “Efficient knowledge base synchro- nization in semantic communication network: A federated distillation approach,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC). IEEE, 2024, pp. 1–6

  8. [8]

    Federated KD-assisted image semantic communication in IoT edge learning,

    X. Xu, Y . Xu, H. Dou, M. Chen, and L. Wang, “Federated KD-assisted image semantic communication in IoT edge learning,” IEEE Internet of Things Journal, vol. 11, no. 21, pp. 34 215–34 228, 2024

Show all 25 references
  1. [9]

    Advances and open problems in federated learning,

    P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummings et al. , “Advances and open problems in federated learning,” Foundations and Trends® in Machine Learning , vol. 14, no. 1–2, pp. 1–210, 2021

  2. [10]

    Communication-efficient wireless federated fine- tuning for large-scale AI models,

    B. Kim and W. Choi, “Communication-efficient wireless federated fine- tuning for large-scale AI models,” arXiv preprint arXiv:2505.00333 , 2025

  3. [11]

    Base station dataset-assisted broad- band over-the-air aggregation for communication-efficient federated learning,

    J.-P. Hong, S. Park, and W. Choi, “Base station dataset-assisted broad- band over-the-air aggregation for communication-efficient federated learning,” IEEE Transactions on Wireless Communications , vol. 22, no. 11, pp. 7259–7272, 2023

  4. [12]

    Sparsified SGD with mem- ory,

    S. U. Stich, J.-B. Cordonnier, and M. Jaggi, “Sparsified SGD with mem- ory,” Advances in Neural Information Processing Systems (NeurIPS) , vol. 31, 2018

  5. [13]

    Error feedback fixes signsgd and other gradient compression schemes,

    S. P. Karimireddy, Q. Rebjock, S. Stich, and M. Jaggi, “Error feedback fixes signsgd and other gradient compression schemes,” in International Conference on Machine Learning (ICML) . PMLR, 2019, pp. 3252– 3261

  6. [14]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial Intelligence and Statistics Conference (AISTATS) . PMLR, 2017, pp. 1273–1282

  7. [15]

    Feature reconstruction aided fed- erated learning for image semantic communication,

    Y . Huh, B. Kim, and W. Choi, “Feature reconstruction aided fed- erated learning for image semantic communication,” arXiv preprint arXiv:2508.02048, 2025

  8. [16]

    Neural discrete representa- tion learning,

    A. Van Den Oord, O. Vinyals et al. , “Neural discrete representa- tion learning,” Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

  9. [17]

    On the convergence of FedAvg on non-IID data,

    X. Li, K. Huang, W. Yang, S. Wang, and Z. Zhang, “On the convergence of FedAvg on non-IID data,” arXiv preprint arXiv:1907.02189 , 2019

  10. [18]

    16 federated knowl- edge distillation,

    H. Seo, J. Park, S. Oh, M. Bennis, and S.-L. Kim, “16 federated knowl- edge distillation,” Machine Learning and Wireless Communications, vol. 457, 2022

  11. [19]

    Federated learning with non-IID data,

    Y . Zhao, M. Li, L. Lai, N. Suda, D. Civin, and V . Chandra, “Federated learning with non-IID data,” arXiv preprint arXiv:1806.00582 , 2018

  12. [20]

    Privacy-enhanced over-the-air fed- erated learning via client-driven power balancing,

    B. Kim, H. Seo, and W. Choi, “Privacy-enhanced over-the-air fed- erated learning via client-driven power balancing,” arXiv preprint arXiv:2410.05907, 2024

  13. [21]

    The algorithmic foundations of differential privacy,

    C. Dwork, A. Roth et al. , “The algorithmic foundations of differential privacy,” Foundations and Trends® in Theoretical Computer Science , vol. 9, no. 3–4, pp. 211–407, 2014

  14. [22]

    Oneshot differentially private top-k selection,

    G. Qiao, W. Su, and L. Zhang, “Oneshot differentially private top-k selection,” in International Conference on Machine Learning (ICML) . PMLR, 2021, pp. 8672–8681

  15. [23]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky and G. Hinton, “Learning multiple layers of features from tiny images,” 2009, [Online]. Available: https://www.cs.toronto.edu/∼kriz/learning-features-2009-TR.pdf

  16. [24]

    QSGD: Communication-efficient SGD via gradient quantization and encoding,

    D. Alistarh, D. Grubic, J. Li, R. Tomioka, and M. V ojnovic, “QSGD: Communication-efficient SGD via gradient quantization and encoding,” Advances in Neural Information Processing Systems (NeurIPS) , vol. 30, 2017

  17. [25]

    Deep learning face attributes in the wild,

    Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” in Proceedings of International Conference on Computer Vision (ICCV), December 2015

Pith tools

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