Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

E-3SFC: Communication-Efficient Federated Learning with Double-way Features Synthesizing

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Compressing gradients into tiny synthetic features — regenerated by the model's own backward pass — lets federated learning cut communication by up to 111.6× while matching or beating uncompressed accuracy.

desk verdict The empirical package is broad and the extension is real, but the convergence theorems lean on an unproved bound (Assumption V.5) that is the only bridge between the compressor and the rates. read the letter →

arxiv 2502.03092 v1 pith:GQQAHG47 submitted 2025-02-05 cs.LG cs.AIcs.DC

classification cs.LGcs.AIcs.DC
keywords federatedlearninggradientcompressionsyntheticfeatureserrorfeedbackdouble-waybudgetschedulerconvergenceanalysisnon-IIDdata
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

Federated learning's bottleneck is the cost of shipping model updates between clients and server. This paper claims a way to ship almost nothing: instead of sending gradients, each client sends a handful of synthetic input-like features that, when passed through the frozen model, regenerate the gradients on the receiving side. The compressor, 3SFC (Single-Step Synthetic Features Compressor), picks these features by a one-step optimization that minimizes the angle between regenerated and true gradients, and error feedback carries the leftover compression error into the next round. On top of this, E-3SFC also compresses the server-to-client download and schedules a fixed total communication budget across rounds. If the claims hold, federated networks can cut traffic by 111.6× in the paper's experiments while matching or exceeding the accuracy of uncompressed training.

What carries the argument

The load-bearing object is the 3SFC compressor and its compression–decompression symmetry. Compression means solving Eq. 9: a budget-constrained search over trainable synthetic features $D_{\mathrm{syn}}$ (plus a scalar $s$) to maximize the cosine similarity between the gradient the frozen model produces at those features and the target gradient; decompression is a single backward pass through the same frozen model. Error feedback (Eq. 6), inherited from sparsified SGD with memory, feeds the residual $\epsilon_t^i$ into the next compression target so that compression error does not accumulate unchecked. The E-3SFC extensions compress the server-to-client download through the same mechanism (Eq. 11, with clients reconstructing the global model via Eq. 12) and add a budget scheduler that solves Eq. 14 offline, shifting more of a fixed total budget to early rounds where compression efficiency is highest. The convergence analysis treats the compressor as a noisy oracle and expresses the error terms through the variance bound $\kappa_S^2 + \sigma_F^2$.

What would settle it

Measure the squared compression error with error feedback at every communication round on the paper's own MNIST+MLP setup and check whether it stays bounded by $\eta^2\kappa_S^2$ for a constant $\kappa_S$ that does not grow with the round number — Figure 8 reports only cosine efficiency, which is not the quantity the theorems bound. On the double-way scheme, also track the reconstruction residual of Eq. 12 over 200 rounds; if it drifts or amplifies, the approximate decompression is unstable.

Watch

Extended reading notes

Core claim

The central claim is that gradient compression in federated learning can be inverted: the model itself becomes the decompressor, and the compressed payload is a set of synthetic features. With the global model $w_t$ and the loss $F$ frozen as training priors, the 3SFC compressor solves the objective of Eq. 9 — find budget-constrained synthetic features $D_{\mathrm{syn}}$ whose gradient $\nabla_w F(D_{\mathrm{syn}}, w_t)$ points as nearly as possible in the direction of the true accumulated gradient $g_t^i + \epsilon_t^i$ (including the error-feedback residual) — and scales the result by $s_t^i$ via Eq. 8. The server aggregates the reconstructed gradients $s_t^i \nabla_w F(D_{\mathrm{syn}}^t, w_t)$, and clients or the server regenerate updates in one backward pass. The paper argues that this representation carries lower compression error per transmitted byte than sparsification or quantization, that the single simulation step avoids the multi-step optimization collapse of earlier synthetic-data compressors, and that under Assumptions V.1–V.5 the scheme converges linearly with aggregation noise in the strongly convex case and sub-linearly in the non-convex case, with rates $O(1/R)$/$O(1/R^2)$ and $O(1/R^{1/2})$/$O(1/R^{2/3})$ respectively, controlled by a compression-variance constant $\kappa_S^2$.

Load-bearing premise

The convergence theorems assume (Assumption V.5) that the compression error after $S$ inner optimization steps is bounded by a fixed constant $\kappa_S$ — $\mathbb{E}\|\eta\nabla F(w_t^i) - g_t^i\|^2 \le \eta^2\kappa_S^2$ — but this bound is never derived from the compressor's own optimization (Eq. 9) or from $S$, so if the real compressor does not satisfy it, the proved rates do not apply.

Editorial extensions

If this is right

  • If the claimed rates hold, 3SFC stays within the convergence class of sparsified gradient methods — linear under strong convexity with aggregation noise, sub-linear otherwise — while transmitting orders of magnitude fewer bytes.
  • Communication budget and local computation become interchangeable: increasing the inner optimization steps $S$ shrinks the variance term $\kappa_S^2$ that sets the convergence radius, so a fixed bandwidth can be traded for accuracy by spending more compute.
  • The scheduler is solved before training begins, so the relaxed cumulative budget (Eq. 5) costs no extra online signalling and assigns more budget to early rounds, where compression efficiency is highest.
  • At equal compression ratio, the paper reports 3SFC consistently above top-k sparsification with error feedback (up to 13.4% higher accuracy), which implies that each transmitted synthetic bit carries more usable gradient information than a sparsified gradient bit.
  • Because downloads are compressed as well, the 111.6× traffic reduction claimed for E-3SFC comes from both FL communication phases, not just the upload.

Reading between the lines

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

  • Beyond the paper: if the missing derivation of $\kappa_S$ from $S$ were supplied, the theory would produce a concrete design rule — how many inner optimization steps are needed to push compression noise below stochastic-gradient noise — which the paper currently states only qualitatively.
  • Beyond the paper: the compressor reuses the machinery of gradient-inversion attacks (the paper notes the resemblance and argues safety because clients invert their own gradients); a natural next test is whether membership or attribute inference from the transmitted synthetic features is harder than from compressed gradients.
  • Beyond the paper: the compression-efficiency metric (cosine between decompressed and raw gradients, Eq. 20) falls over training rounds even with error feedback; the budget scheduler is designed to counteract that, and a measurable prediction is that late-round per-byte efficiency under E-3SFC stays above that of constant-budget 3SFC.
  • Beyond the paper: the error-feedback residual is never transmitted and Eq. 12 reconstruction is approximate, so an open benchmark is whether the reconstruction drift compounds over very long training horizons; tracking that residual would settle the stability of double-way compression.
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

3 major / 6 minor

Summary. The paper proposes E-3SFC, a communication-efficient federated learning method whose core compressor, 3SFC, represents client gradients by synthetic features optimized through a short back-propagation loop. The claimed contributions are threefold: a single-step synthetic-features compressor with error feedback, a double-way compression scheme that also compresses model downloads, and a communication-budget scheduler. The paper provides convergence analyses for 3SFC under strongly convex and non-convex assumptions, claiming linear and sub-linear rates with aggregation noise, and reports experiments over six datasets and six models showing large communication reductions at comparable or better accuracy than several baselines. The main theoretical claim, however, is that convergence rates follow from an assumed bound on the compression error rather than from a property proved for the proposed compressor.

Significance. If the theoretical claims were established, this would be a valuable contribution: a gradient compression method with reported communication reductions of two to three orders of magnitude and broad empirical validation. The empirical work is extensive, covering six datasets, six models, 10-60 clients, ablations, budget schedulers, and compression-efficiency visualizations, and the paper points to a public code repository. The novel perspective of using the model itself as a decompressor is interesting and the experiments suggest practical promise. However, the central theoretical result is not established as stated: the convergence rates are conditional on an assumption about the compressor's error that is never derived, and the proof of the key lemma applies that assumption inconsistently. The paper's significance therefore rests mostly on its empirical findings until the theory is repaired or appropriately qualified.

major comments (3)
  1. [Section V, Assumption V.5 and Eq. (9)] The convergence theorems are conditional on an assumed bound, not on a property derived for 3SFC. Assumption V.5 asserts E||η∇F(w_t^i) − g_t^i||^2 ≤ η^2κ_S^2, but κ_S is never expressed in terms of the inner-loop objective Eq. (9), the optimization horizon S, the budget B, or the model's loss landscape. Theorems V.3 and V.4 then contain κ_S only through this assumption, so the argument proves that any compressor satisfying the bound converges at those rates; it does not establish that the optimization in Eq. (9) produces such a bound. Remark V.3.1 and the Insight paragraph state that S can be chosen to keep κ_S at the level of σ_F^2, but this is an instruction rather than a consequence of Eq. (9). This is load-bearing for the paper's central theoretical claim and needs either a derivation of κ_S from Eq. (9) or a substantially weakened claim.
  2. [Section IV-B, Eq. (12), and Section V] The formal analysis does not cover the double-way compression used in E-3SFC, which is the paper's titular algorithm. The server compresses the updated global model via Eq. (11), and clients reconstruct it via Eq. (12): w_t = w_{t−1} − s_t∇_{w_{t−1}}F(D_t_syn, w_{t−1}). This reconstruction is exact only if the compressed residual perfectly represents w_t − w_{t−1} + ϵ_t, which is not guaranteed and whose error is never bounded. Assumption V.5 and Lemma V.2 concern only the upload compressor applied to η∇F(w_t^i) + ϵ; they do not bound the download-compression error. Consequently, Theorems V.3 and V.4 do not provide convergence guarantees for Algorithm 1 as presented. Please either extend the analysis to the download compressor or state explicitly that the theoretical results apply to 3SFC without double-way compression.
  3. [Appendix, Lemma V.2 proof, Eqs. (30)-(32)] There is a mismatch between Assumption V.5 and its use in the proof of Lemma V.2. Assumption V.5 bounds E||η∇F(w_t^i) − g_t^i||^2 with the full local gradient ∇F, but Eq. (30) applies it to the stochastic gradient ∇F̃(w^{t+k}) in the first term, yielding E||g − η∇F̃||^2 ≤ η^2κ_S^2. The assumption as stated does not control this quantity; a bound would require an additional relationship between ∇F and ∇F̃ beyond Assumption V.2, which only controls E||∇F̃ − ∇F||^2. The same display also has missing norm symbols on the second and third terms. Since Lemma V.2 is the only route from the compressor to Theorems V.3 and V.4, this gap is load-bearing for the claimed convergence rates.
minor comments (6)
  1. [Eq. (9)] The denominator in the cosine term is printed as ||∇w_t + ϵ_t^i||; it should almost certainly be ||g_t^i + ϵ_t^i|| or ||∇_{w_t}F(D_syn,i, w_t)||. Please correct the typo, as it obscures the stated equivalence with Eq. (7).
  2. [Eq. (14)] The regularization term τ|H−B|/(H−B) is undefined when H=B, which is the constant-budget case used as a baseline in Table VII. Please specify a convention for this case or replace the term with a well-defined penalty.
  3. [Eq. (19) and Tables III-IV] The compression terminology is inconsistent. Eq. (19) defines Comp. Rate = Comp. Size / Uncomp. Size = 1 / Comp. Ratio, but the text and tables call a value such as 250× a 'compression ratio' and use 'higher compression ratio' to mean more compression. Please unify the definitions to avoid confusion.
  4. [Corollary V.3.1] The phrase 'quadratic convergence with total local epochs R' appears to mean a quadratic speed-up O(1/R^2); as written it conflicts with the linear convergence rate stated in Theorem V.3. Please rephrase to avoid a literal reading that contradicts the theorem.
  5. [Appendix, proof of Lemma V.1, Eq. (22)] The displayed expansion of term A contains a typo: the second inner product is written as −2⟨w_{t_a} − w∗, η̃⟩ and appears to be missing the gradient factor, so the step from Eq. (21) to Eq. (23) is difficult to follow.
  6. [Tables III and VI] Several entries appear to be mistyped, for example the 60-client E-3SFC entry for RegNet on Cifar10 reads '42.47 57.6×' rather than the '757.6×' that would match the corresponding column. Please proofread all table cells.

Circularity Check

1 steps flagged · score 5.0 of 10

Convergence claims are conditional on Assumption V.5: the rate bounds are expressed through κ_S, the same compression-variance constant assumed rather than derived from 3SFC's S-step optimization.

  1. self definitional [Section V, Assumption V.5; Lemma V.2; Theorems V.3 and V.4; Remark V.3.1; Insight paragraph]
    "Assumption V.5 (Unbiased Compression Optimization Variance Bound). The expectation of the difference between local gradients and compressed local gradients is upper-bounded, i.e., E||η∇F (wt i) − gt i||2 ≤ η2κ2 S ... Theorem V.3 ... O(F ( ¯w) − F (w∗)) = O(μF e−˜ημF R 2 ∆0) + O( κ2 S + σ2 F μF ) + ..."

    The convergence bounds in Theorems V.3 and V.4 are written directly in terms of κ_S², the very quantity introduced in Assumption V.5 as an upper bound on 3SFC's compression residual. κ_S is never derived from the cosine-based objective in Eq. 9, from the number of inner-loop steps S, or from the communication budget B. Instead, Remark V.3.1 and the Insight paragraph assert that 'S is chosen to maintain the same level variance of compression' and that κ_S 'can be effectively reduced by increasing S,' without proof. Thus the claimed linear and sub-linear convergence of 3SFC is not a derived consequence of the compressor; it is an assumed bound.

full rationale

The paper's central theoretical claims about 3SFC's convergence rates reduce to Assumption V.5. The proof chain is: Lemma V.1 bounds local shift in terms of the compression residual ||g−η∇F||², Lemma V.2 bounds that residual using κ_S² and σ_F², and Theorems V.3 and V.4 then present rates containing κ_S²+σ_F² as an additive term. Critically, κ_S² is never bounded in terms of the algorithm's actual optimization dynamics (Eq. 9's cosine objective, the number of inner steps S, or the budget B). The remark after Corollary V.3.1 simply instructs that S be chosen to keep κ_S² at the same level as σ_F², and the 'Insight' asserts without proof that increasing S reduces κ_S. Consequently, the theorem is conditional on an unverified property of the compressor; it would hold for any compressor satisfying that property, including ones that never run Eq. 9. This is a partial circularity: the prediction of linear/sub-linear convergence is not established for the specific 3SFC algorithm. The empirical comparisons against DGC, STC, etc. are independent and not circular. Because the core theoretical result rests on an assumed bound that is not shown to be satisfied, the score is 5 for a moderately circular, assumption-driven analysis.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The central theoretical claim rests on Assumption V.5, a bound on compression error that is assumed rather than derived. The algorithm also depends on exact reconstruction of the global model through compressed features, which is not proven. No new physical entities are introduced.

free parameters (3)
  • kappa_S
    Introduced in Assumption V.5 as the bound on compressed-gradient error variance. The convergence theorems' constants depend on it, but it is not derived from the 3SFC inner loop or from S, so the stated rates are in terms of an uncharacterized quantity.
  • tau = 3.0
    Hyperparameter in the budget-scheduler objective (Eq. 14) that balances transferred information against deviation from the constant budget; set by hand and not analyzed across values.
  • lambda = 0
    Regularization weight in Eq. 7/9. The paper claims lambda regularizes synthetic features but sets lambda=0 in all experiments, so the objective reduces to cosine-similarity maximization without sparsity control.
assumptions (7)
  • standard math Assumption V.1: bounded local variance
    Standard FL data-heterogeneity bound.
  • standard math Assumption V.2: bounded approximated gradient variance
    Standard bound for minibatch gradient approximation.
  • standard math Assumption V.3: bounded global optimal variance
    Standard assumption that local gradients at the optimum are bounded.
  • standard math Assumption V.4: expected smoothness
    Standard smoothness assumption used in FL convergence proofs.
  • ad hoc to paper Assumption V.5: bounded compression optimization variance
    This is the critical premise: it assumes 3SFC's compressed gradients are within eta^2 * kappa_S^2 of true gradients, without proving that optimizing Eq. 9 for S steps achieves this. The convergence theorems inherit this unproved bound.
  • domain assumption Exact global-model reconstruction in double-way compression
    Algorithm 1 has clients reconstruct w_t from compressed features (Eq. 12), but the reconstruction is approximate and the server and client may disagree on the training prior w_t; the paper does not analyze this error.
  • ad hoc to paper Inner-loop optimization (Eq. 9) converges in S steps
    The compressor is assumed to find synthetic features with small cosine error within S=10 iterations; no convergence guarantee for this non-convex inner problem is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of E-3SFC: Communication-Efficient Federated Learning with Double-way Features Synthesizing." pith.science (2026). https://pith.science/paper/GQQAHG47

@misc{pith2026250203092,
  author       = {Pith},
  title        = {Pith review of: E-3SFC: Communication-Efficient Federated Learning with Double-way Features Synthesizing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GQQAHG47}},
  note         = {Machine review of arXiv:2502.03092}
}
read the original abstract

The exponential growth in model sizes has significantly increased the communication burden in Federated Learning (FL). Existing methods to alleviate this burden by transmitting compressed gradients often face high compression errors, which slow down the model's convergence. To simultaneously achieve high compression effectiveness and lower compression errors, we study the gradient compression problem from a novel perspective. Specifically, we propose a systematical algorithm termed Extended Single-Step Synthetic Features Compressing (E-3SFC), which consists of three sub-components, i.e., the Single-Step Synthetic Features Compressor (3SFC), a double-way compression algorithm, and a communication budget scheduler. First, we regard the process of gradient computation of a model as decompressing gradients from corresponding inputs, while the inverse process is considered as compressing the gradients. Based on this, we introduce a novel gradient compression method termed 3SFC, which utilizes the model itself as a decompressor, leveraging training priors such as model weights and objective functions. 3SFC compresses raw gradients into tiny synthetic features in a single-step simulation, incorporating error feedback to minimize overall compression errors. To further reduce communication overhead, 3SFC is extended to E-3SFC, allowing double-way compression and dynamic communication budget scheduling. Our theoretical analysis under both strongly convex and non-convex conditions demonstrates that 3SFC achieves linear and sub-linear convergence rates with aggregation noise. Extensive experiments across six datasets and six models reveal that 3SFC outperforms state-of-the-art methods by up to 13.4% while reducing communication costs by 111.6 times. These findings suggest that 3SFC can significantly enhance communication efficiency in FL without compromising model performance.

Figures

Figures reproduced from arXiv: 2502.03092 by the authors.

Figure 1
Figure 1. Information Compression Rate v.s. Model Conver￾gence: The rate of convergence reduces as the compression rate decreases. The evaluated MLP model is trained on non￾iid MNIST with 20 FL clients. the server and clients are crucial for FL [7], [8]. Nevertheless, the limited network bandwidth [9] and explosive growth of model size [10]–[13] not only decrease the training efficiency of FL but also hinder the deployment an… view at source ↗
Figure 3
Figure 3. Relationship between E-3SFC and 3SFC itself as a gradient decompressor. 2) Theoretical convergence analyses are provided, showing that the proposed compressor 3SFC has a O( 1 R )/O( 1 R2 ) speedup with/without aggregation noise under the strongly convex case, and a O( 1 R1/2 )/O( 1 R2/3 ) conver￾gence rate with/without aggregation noise under the non￾convex case. 3) Extensive experiments are conducted on six dataset… view at source ↗
Figure 4
Figure 4. Above: When fitting gradients obtained by 128 steps of SGD using 128 steps of simulation [22], it collapsed. Meanwhile, E-3SFC requires only one step of simulation, occupying less computation and storage, but achieves signifi￾cantly better results. Below: Before the collapse, the gradients of its trainable parameters exhibit a phenomenon similar to the gradient explosion, where the magnitude of gradients increases a… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: The general workflow of 3SFC. When compressing in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Illustration of our manual dataset partitions for 20 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Test accuracy and training loss comparisons after 200 [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Compression efficiency comparisons. 3SFC owns sig [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Discerning and quantifying high frequency activities in EEG under normal and epileptic conditions

    nlin.CD 2025-08 unverdicted novelty 5.0 of 10

    The authors claim that gamma-band EEG activity at roughly 50 Hz and 76 Hz shows distinct bi-stable dynamics in epilepsy, enabling 94-95% accuracy in separating normal from ictal states.

Reference graph

Works this paper leans on

63 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [33]

    Communication- efficient federated learning with single-step synthetic features compres- sor for faster convergence,

    Y . Zhou, M. Shi, Y . Li, Y . Sun, Q. Ye, and J. Lv, “Communication- efficient federated learning with single-step synthetic features compres- sor for faster convergence,” in Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , 2023, pp. 5031–5040

  2. [1]

    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 . PMLR, 2017, pp. 1273– 1282

  3. [2]

    A review of applications in federated learning,

    L. Li, Y . Fan, M. Tse, and K.-Y . Lin, “A review of applications in federated learning,” Computers & Industrial Engineering , vol. 149, p. 106854, 2020

  4. [3]

    Large scale distributed deep networks,

    J. Dean, G. Corrado, R. Monga, K. Chen, M. Devin, M. Mao, M. Ran- zato, A. Senior, P. Tucker, K. Yang et al., “Large scale distributed deep networks,” Advances in neural information processing systems , vol. 25, 2012

  5. [4]

    Demystifying parallel and distributed deep learning: An in-depth concurrency analysis,

    T. Ben-Nun and T. Hoefler, “Demystifying parallel and distributed deep learning: An in-depth concurrency analysis,” ACM Computing Surveys (CSUR), vol. 52, no. 4, pp. 1–43, 2019

  6. [5]

    Prior: Personalized prior for reactivating the information overlooked in federated learning,

    M. Shi, Y . Zhou, K. Wang, H. Zhang, S. Huang, Q. Ye, and J. Lv, “Prior: Personalized prior for reactivating the information overlooked in federated learning,” Advances in Neural Information Processing System , 2023

  7. [6]

    Defta: A plug-and-play peer-to-peer decentralized federated learning framework,

    Y . Zhou, M. Shi, Y . Tian, Q. Ye, and J. Lv, “Defta: A plug-and-play peer-to-peer decentralized federated learning framework,” Information Sciences, vol. 670, p. 120582, 2024

  8. [7]

    Robust and communication-efficient federated learning from non-iid data,

    F. Sattler, S. Wiedemann, K.-R. M ¨uller, and W. Samek, “Robust and communication-efficient federated learning from non-iid data,” IEEE transactions on neural networks and learning systems , vol. 31, no. 9, pp. 3400–3413, 2019

Show all 63 references
  1. [8]

    Communication-efficient federated learning with compensated overlap-fedavg,

    Y . Zhou, Q. Ye, and J. Lv, “Communication-efficient federated learning with compensated overlap-fedavg,” IEEE Transactions on Parallel and Distributed Systems, vol. 33, no. 1, pp. 192–205, 2021

  2. [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

  3. [10]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

  4. [11]

    An empirical study of parameter efficient fine-tuning on vision-language pre-train model,

    Y . Tian, M. Yang, Y . Li, D. Liu, X. Ren, X. Peng, and J. Lv, “An empirical study of parameter efficient fine-tuning on vision-language pre-train model,” in IEEE International Conference on Multimedia and Expo, ICME 2024 , pp. 1–6

  5. [12]

    UNITE: multitask learning with sufficient feature for dense prediction,

    Y . Tian, Y . Lin, Q. Ye, J. Wang, X. Peng, and J. Lv, “UNITE: multitask learning with sufficient feature for dense prediction,” IEEE Trans. Syst. Man Cybern. Syst. , vol. 54, no. 8, pp. 5012–5024, 2024

  6. [13]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,

    W. Fedus, B. Zoph, and N. Shazeer, “Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,” 2021

  7. [14]

    Towards federated learning at scale: System design,

    K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman, V . Ivanov, C. Kiddon, J. Kone ˇcn`y, S. Mazzocchi, B. McMahan et al. , “Towards federated learning at scale: System design,” Proceedings of Machine Learning and Systems , vol. 1, pp. 374–388, 2019

  8. [15]

    Scalable distributed dnn training using commodity gpu cloud computing,

    N. Strom, “Scalable distributed dnn training using commodity gpu cloud computing,” in Sixteenth annual conference of the international speech communication association, 2015

  9. [16]

    Deep gradient compression: Reducing the communication bandwidth for distributed training,

    Y . Lin, S. Han, H. Mao, Y . Wang, and W. J. Dally, “Deep gradient compression: Reducing the communication bandwidth for distributed training,” arXiv preprint arXiv:1712.01887 , 2017

  10. [17]

    Gradient sparsification for communication-efficient distributed optimization,

    J. Wangni, J. Wang, J. Liu, and T. Zhang, “Gradient sparsification for communication-efficient distributed optimization,” Advances in Neural Information Processing Systems , vol. 31, 2018

  11. [18]

    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 , vol. 30, 2017

  12. [19]

    signsgd: Compressed optimisation for non-convex problems,

    J. Bernstein, Y .-X. Wang, K. Azizzadenesheli, and A. Anandkumar, “signsgd: Compressed optimisation for non-convex problems,” in Inter- national Conference on Machine Learning. PMLR, 2018, pp. 560–569

  13. [20]

    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 . PMLR, 2019, pp. 3252–3261

  14. [21]

    Knowledge distillation: A survey,

    J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,” International Journal of Computer Vision , vol. 129, pp. 1789– 1819, 2021

  15. [22]

    Federated learning via synthetic data,

    J. Goetz and A. Tewari, “Federated learning via synthetic data,” arXiv preprint arXiv:2008.04489, 2020

  16. [23]

    Communication-efficient federated learning based on compressed sensing,

    C. Li, G. Li, and P. K. Varshney, “Communication-efficient federated learning based on compressed sensing,”IEEE Internet of Things Journal, vol. 8, no. 20, pp. 15 531–15 541, 2021

  17. [24]

    Fedsynth: Gradient compression via synthetic data in federated learning,

    S. Hu, J. Goetz, K. Malik, H. Zhan, Z. Liu, and Y . Liu, “Fedsynth: Gradient compression via synthetic data in federated learning,” arXiv preprint arXiv:2204.01273, 2022

  18. [25]

    Sparse communication for distributed gradient descent,

    A. F. Aji and K. Heafield, “Sparse communication for distributed gradient descent,” arXiv preprint arXiv:1704.05021 , 2017

  19. [26]

    z-signfedavg: A unified stochastic sign-based compression for federated learning,

    Z. Tang, Y . Wang, and T.-H. Chang, “z-signfedavg: A unified stochastic sign-based compression for federated learning,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 14, 2024, pp. 15 301–15 309

  20. [27]

    Optimization methods for large- scale machine learning,

    L. Bottou, F. E. Curtis, and J. Nocedal, “Optimization methods for large- scale machine learning,” SIAM review, vol. 60, no. 2, pp. 223–311, 2018

  21. [28]

    Deep leakage from gradients,

    L. Zhu, Z. Liu, and S. Han, “Deep leakage from gradients,” Advances in neural information processing systems , vol. 32, 2019

  22. [29]

    Sparsified sgd with memory,

    S. U. Stich, J.-B. Cordonnier, and M. Jaggi, “Sparsified sgd with memory,” Advances in Neural Information Processing Systems , vol. 31, 2018

  23. [30]

    A theoretical study of dataset distillation,

    Z. Izzo and J. Zou, “A theoretical study of dataset distillation,” in NeurIPS 2023 Workshop on Mathematics of Modern Machine Learning , 2023

  24. [31]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

  25. [32]

    The convergence of sparsified gradient methods,

    D. Alistarh, T. Hoefler, M. Johansson, N. Konstantinov, S. Khirirat, and C. Renggli, “The convergence of sparsified gradient methods,” Advances in Neural Information Processing Systems , vol. 31, 2018

  26. [34]

    Communication-efficient federated learning,

    M. Chen, N. Shlezinger, H. V . Poor, Y . C. Eldar, and S. Cui, “Communication-efficient federated learning,” Proceedings of the Na- tional Academy of Sciences , vol. 118, no. 17, p. e2024789118, 2021

  27. [35]

    Rethinking gradient sparsification as total error minimiza- tion,

    A. Sahu, A. Dutta, A. M Abdelmoniem, T. Banerjee, M. Canini, and P. Kalnis, “Rethinking gradient sparsification as total error minimiza- tion,” Advances in Neural Information Processing Systems , vol. 34, pp. 8133–8146, 2021

  28. [36]

    Pre- dicting parameters in deep learning,

    M. Denil, B. Shakibi, L. Dinh, M. Ranzato, and N. De Freitas, “Pre- dicting parameters in deep learning,” Advances in neural information processing systems, vol. 26, 2013

  29. [37]

    1-bit stochastic gradient descent and its application to data-parallel distributed training of speech dnns,

    F. Seide, H. Fu, J. Droppo, G. Li, and D. Yu, “1-bit stochastic gradient descent and its application to data-parallel distributed training of speech dnns,” in Fifteenth annual conference of the international speech communication association , 2014. JOURNAL OF LATEX CLASS FILES...

  30. [38]

    Communication-efficient federated learning via knowledge distillation,

    C. Wu, F. Wu, L. Lyu, Y . Huang, and X. Xie, “Communication-efficient federated learning via knowledge distillation,” Nature communications, vol. 13, no. 1, pp. 1–8, 2022

  31. [39]

    Communication-efficient federated deep learning with layerwise asynchronous model update and temporally weighted aggregation,

    Y . Chen, X. Sun, and Y . Jin, “Communication-efficient federated deep learning with layerwise asynchronous model update and temporally weighted aggregation,” IEEE transactions on neural networks and learning systems, vol. 31, no. 10, pp. 4229–4238, 2019

  32. [40]

    A convergence theory for deep learn- ing via over-parameterization,

    Z. Allen-Zhu, Y . Li, and Z. Song, “A convergence theory for deep learn- ing via over-parameterization,” in International conference on machine learning. PMLR, 2019, pp. 242–252

  33. [41]

    A systematic review on overfitting control in shallow and deep neural networks,

    M. M. Bejani and M. Ghatee, “A systematic review on overfitting control in shallow and deep neural networks,” Artificial Intelligence Review , vol. 54, no. 8, pp. 6391–6438, 2021

  34. [42]

    idlg: Improved deep leakage from gradients,

    B. Zhao, K. R. Mopuri, and H. Bilen, “idlg: Improved deep leakage from gradients,” arXiv preprint arXiv:2001.02610 , 2020

  35. [43]

    Gradient leakage attacks in federated learning: Research frontiers, taxonomy and future directions,

    H. Yang, M. Ge, D. Xue, K. Xiang, H. Li, and R. Lu, “Gradient leakage attacks in federated learning: Research frontiers, taxonomy and future directions,” IEEE Network, 2023

  36. [44]

    The mnist database of handwritten digit images for machine learning research,

    L. Deng, “The mnist database of handwritten digit images for machine learning research,” IEEE Signal Processing Magazine , vol. 29, no. 6, pp. 141–142, 2012

  37. [45]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,” arXiv preprint arXiv:1708.07747, 2017

  38. [46]

    Emnist: Extending mnist to handwritten letters,

    G. Cohen, S. Afshar, J. Tapson, and A. Van Schaik, “Emnist: Extending mnist to handwritten letters,” in 2017 international joint conference on neural networks (IJCNN) . IEEE, 2017, pp. 2921–2926

  39. [47]

    Learning multiple layers of features from tiny images,

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

  40. [48]

    Character-level convolutional net- works for text classification,

    X. Zhang, J. Zhao, and Y . LeCun, “Character-level convolutional net- works for text classification,” Advances in neural information processing systems, vol. 28, 2015

  41. [49]

    Maximum likelihood estimation of dirichlet distribution parameters,

    J. Huang, “Maximum likelihood estimation of dirichlet distribution parameters,” CMU Technique report, vol. 76, 2005

  42. [50]

    Tackling the objective inconsistency problem in heterogeneous federated optimiza- tion,

    J. Wang, Q. Liu, H. Liang, G. Joshi, and H. V . Poor, “Tackling the objective inconsistency problem in heterogeneous federated optimiza- tion,” Advances in neural information processing systems , vol. 33, pp. 7611–7623, 2020

  43. [51]

    Federated learning on non-iid data silos: An experimental study,

    Q. Li, Y . Diao, Q. Chen, and B. He, “Federated learning on non-iid data silos: An experimental study,” in 2022 IEEE 38th International Conference on Data Engineering (ICDE) . IEEE, 2022, pp. 965–978

  44. [52]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

  45. [53]

    Designing network design spaces,

    I. Radosavovic, R. P. Kosaraju, R. Girshick, K. He, and P. Doll ´ar, “Designing network design spaces,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 10 428–10 436

  46. [54]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in International conference on machine learning . PMLR, 2015, pp. 448–456

  47. [55]

    Dropout: a simple way to prevent neural networks from over- fitting,

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut- dinov, “Dropout: a simple way to prevent neural networks from over- fitting,” The journal of machine learning research , vol. 15, no. 1, pp. 1929–1958, 2014

  48. [56]

    Why batch normalization damage federated learning on non-iid data?

    Y . Wang, Q. Shi, and T.-H. Chang, “Why batch normalization damage federated learning on non-iid data?” arXiv preprint arXiv:2301.02982 , 2023

  49. [57]

    Personalized federated learning with moreau envelopes,

    C. T Dinh, N. Tran, and J. Nguyen, “Personalized federated learning with moreau envelopes,” Advances in neural information processing systems, vol. 33, pp. 21 394–21 405, 2020

  50. [58]

    Federated cinn clustering for accurate clustered federated learning,

    Y . Zhou, M. Shi, Y . Tian, Y . Li, Q. Ye, and J. Lv, “Federated cinn clustering for accurate clustered federated learning,” in ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 5590–5594

  51. [59]

    Step-ahead error feedback for distributed training with compressed gradient,

    A. Xu, Z. Huo, and H. Huang, “Step-ahead error feedback for distributed training with compressed gradient,” in Proceedings of the AAAI Confer- ence on Artificial Intelligence , vol. 35, no. 12, 2021, pp. 10 478–10 486

  52. [60]

    Detached error feedback for distributed sgd with random sparsification,

    A. Xu and H. Huang, “Detached error feedback for distributed sgd with random sparsification,” in International Conference on Machine Learning. PMLR, 2022, pp. 24 550–24 575

  53. [61]

    Federated reinforcement learning: Techniques, applications, and open challenges,

    J. Qi, Q. Zhou, L. Lei, and K. Zheng, “Federated reinforcement learning: Techniques, applications, and open challenges,” arXiv preprint arXiv:2108.11887, 2021. IX. B IOGRAPHY SECTION Yuhao Zhou is currently working toward a Ph.D. degree in computer science with the College of ...

  54. [62]

    When δ∗ RC1 ≤ ˆη2 n or δ∗ C2 ≤ ˆη3 n, by choosing a proper ˜η = min{ δ∗ 1 2 C 1 2 1 R 1 2 , ∆∗ 1 3 C 1 3 2 R 1 3 }, the following bound can be derived: O(E||∇F (wt∗ )||2) = K(C0 + ∆∗ 1 2 C 1 2 1 R 1 2 + ∆∗ 2 3 C 1 3 2 R 2 3 ). (66)

  55. [63]

    When ∆∗ RC1 ≥ ˆη2 n or ∆∗ C2 ≥ ˆη3 n, by choosing a proper ˜η = ˆηn, the following bound can be derived: O(E||∇F (wt∗ )||2) = K( 1 ˜ηR ∆∗ + C0 + ∆∗ 1 2 C 1 2 1 R 1 2 + ∆∗ 2 3 C 1 3 2 R 2 3 ). (67) Finally, a general bound can be derived as follows, and the Theorem V .4 is ther...

Pith tools

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