Pith. sign in

REVIEW 4 major objections 5 minor 58 references

The Panaceas for Improving Low-Rank Decomposition in Communication-Efficient Federated Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read FedMUD decomposes only the model update — not the full weights — using block-wise Kronecker factors and aggregation-aware products, and converges to a stationary point of the federated objective with a faster bound than FedLMT.

desk verdict Useful paper with three solid techniques; the convergence proof leans on an unverified assumption, so treat the theory as suggestive rather than definitive. read the letter →

arxiv 2505.23176 v2 pith:RCLPCXOQ submitted 2025-05-29 cs.LG cs.DC

classification cs.LGcs.DC
keywords federatedlearningcommunicationefficiencylow-rankdecompositionmodelupdateKroneckerparameteraggregationconvergenceanalysisnon-convexoptimization
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 main bottleneck is communication, and low-rank decomposition is a popular cure: replace a weight matrix by a product of two smaller matrices. The paper's claim is that previous cures aim at the wrong target. Instead of decomposing the model parameters, one should decompose the model update, add block-wise Kronecker structure to raise the rank of the recovered matrix, and aggregate the factors in a way that makes the order of averaging and recovery irrelevant. The paper packages the three choices as FedMUD and proves that it converges to a stationary point of the federated objective under standard assumptions, with a rate bound that degrades as the reset interval $S$ grows and that reduces to FedLMT as the special case $S\ge T$. Empirically, the combined method at $1/32$ communication cost lands within a few points of un-compressed FedAvg on FMNIST, SVHN, CIFAR-10, and CIFAR-100, and beats the low-rank baselines by up to 12 points on CIFAR-100.

What carries the argument

There are three mechanisms, one per issue. MUD freezes the base parameters and trains low-rank factors $U\in\mathbb{R}^{m\times r}$, $V\in\mathbb{R}^{n\times r}$ whose product is the communication content; $V$ starts at zero so the first transmitted update is exactly the local change. BKD tiles the update into $k^2$ blocks and writes each block as $U\otimes V$ with square factors, which gives a rank bound of $\operatorname{rank}(U)\operatorname{rank}(V)$ per block and dynamic compression through $k$. AAD decouples every trainable product $UV^\top$ into $U\tilde V^\top + \tilde U V^\top$ with fixed random tilde factors, so that the server's average of factors recovers the same matrix as the average of recovered matrices. The convergence proof's load-bearing step is Lemma 1: under the assumption that at least one averaged factor per layer has smallest singular value at least $\psi_{uv}$, the gradient norm in factored space dominates $\psi_{uv}^2\|\nabla f_l(\bar w)\|_F^2$, which converts the factored update into progress on the original objective.

What would settle it

Record the smallest singular values of the averaged factors $\bar U_l, \bar V_l$ across rounds for FedMUD on a non-IID dataset such as CIFAR-100 with Dirichlet-0.1 partitions; if any layer ever has both values at machine precision, Assumption 4 is violated and the stated bound no longer applies. A second test: fix identical total parameter budgets for FedMUD with $s=1$ and FedLMT, and compare required rounds to reach a target accuracy; if the smaller reset interval does not converge faster in practice, Remark 3's claimed advantage is contradicted.

Watch

Extended reading notes

Core claim

The central discovery is that the information loss in low-rank federated learning comes from three separate, correctable sources, and the paper corrects all three. Decomposing whole parameters corrupts the static base weights; FedMUD freezes $W$ and learns $U_t(V_t)^\top$ as the update only. Ordinary matrix-product factors cap the recovered rank at $r$; Block-wise Kronecker Decomposition partitions the update into $k^2$ blocks, each represented as a Kronecker product, lifting the rank upper bound to full rank with compression ratio $2k/\sqrt{mn}$. Direct aggregation of factors is biased, because averaging $U_i(V_i)^\top$ over clients differs from $\bar U(\bar V)^\top$ by the cross term $\operatorname{A}(\Delta U)\operatorname{A}(\Delta V)^\top$; Aggregation-Aware Decomposition writes $UV^\top = U\tilde V^\top + \tilde U V^\top$ with fixed $\tilde U,\tilde V$, making the cross term vanish. Theorem 1 bounds the average squared gradient norm by $\frac{2}{\eta c T}(f(\bar w^0)-f(w^*)) + O(\eta^{2-c})\bigl(1+(L-\rho)O(\Gamma_u^4+\Gamma_v^4)\bigr)$, with the $\Gamma$ terms growing in the reset interval $S$; Remark 3 concludes that FedLMT, the $S\ge T$ special case, has a slower rate bound.

Load-bearing premise

At every iteration and every decomposed layer, at least one of the two server-averaged factor matrices must have a smallest singular value bounded away from zero by $\psi_{uv}$, which the paper supports by citing random-matrix theory rather than proving it for aggregated, re-initialized factors. If $\psi_{uv}$ can drift to zero, Lemma 1 and the convergence bound lose their grip.

Editorial extensions

If this is right

  • At compression ratio $1/32$, FedMUD+BKD+AAD reaches 89.0%, 86.6%, 75.9%, and 41.2% accuracy on FMNIST, SVHN, CIFAR-10, and CIFAR-100 (Non-IID-1), narrowing the gap to FedAvg to 1.3, 3.1, 5.1, and 8.6 points respectively.
  • Setting the reset interval $s=1$ is both the theoretical optimum (smallest $\Gamma_u,\Gamma_v$) and the empirical best choice; increasing $s$ moves the method toward FedLMT's slower behavior.
  • AAD is not specific to matrix products or to BKD; the paper states it applies to any product of trainable parameters that has to be aggregated, including Kronecker products, so the technique transfers to other compressed-training schemes.
  • With $\rho=L$ the theorem collapses to the FedAvg rate, so FedMUD adds the low-rank machinery without changing the standard non-convex convergence order $O(1/\sqrt{T})$ (when $\eta=1/\sqrt{T}$ and $c\to 1$).
  • Compared with FedBAT, a quantization method that compresses only the uplink, FedMUD+BKD+AAD still leads by up to 3.8 points on CIFAR-100 Non-IID-1, suggesting the decomposition approach remains competitive with quantization methods.

Reading between the lines

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

  • The paper's 'faster than FedLMT' argument compares upper bounds (Remark 3) rather than tracking all constants, so a reader should not read it as a proof of a smaller rate exponent; a direct test is to plot FedMUD with $s=1$ against FedLMT at matched parameter counts and compare actual rounds to a target accuracy.
  • Assumption 4 is the fragile link: the paper supports it by citing random-matrix theory, but the averaged factors $\bar U, \bar V$ are correlated through training and reinitialization, so the uniform lower bound $\psi_{uv}$ remains an assumption that an empirical trace of smallest singular values could confirm or break.
  • Because BKD improves the rank-per-parameter tradeoff that also limits LoRA-style adapters, the block-Kronecker structure suggests a direct extension to parameter-efficient fine-tuning of large models; the paper does not explore that direction.
  • The reset mechanism in MUD, which periodically merges the low-rank update into the frozen base and reinitializes the factors, acts like a sequence of low-rank corrections; a natural untested prediction is that the number of resets, not the rank alone, controls the model's effective capacity in the low-communication regime.
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 paper proposes three complementary techniques for communication-efficient federated learning with low-rank decomposition: Model Update Decomposition (MUD), Block-wise Kronecker Decomposition (BKD), and Aggregation-Aware Decomposition (AAD). MUD decomposes only the model update instead of the full model parameters, BKD increases the rank upper bound of the recovered matrix at a fixed compression ratio, and AAD removes the aggregation bias that arises when sub-matrices are averaged directly. The paper presents a convergence theorem for FedMUD under smoothness, bounded-gradient, bounded-factor, and a positive-smallest-singular-value assumption, and it claims that FedMUD converges faster than FedLMT. Experiments on FMNIST, SVHN, CIFAR-10, and CIFAR-100 under non-IID partitions show that the full method FedMUD+BKD+AAD improves accuracy over the baselines, with ablations for reset interval, initialization scale, and compression ratio.

Significance. If the theoretical claims were fully established, this would be a useful contribution to communication-efficient FL: MUD avoids introducing decomposition error into the frozen base model, BKD provides a higher rank upper bound than standard low-rank decomposition, and AAD gives an exact, linear aggregation property (Eqs. 6-9) that is clean and easy to verify. The AAD derivation is crisp, the experimental study is reasonably complete, and the code is released. The convergence proof is structurally coherent under the stated assumptions. However, the central speed-up claim over FedLMT currently rests on comparing upper-bound constants rather than on a proven rate separation, and the proof relies on Assumption 4 being satisfied uniformly over training, which is asserted rather than established for the algorithm's actual iterates. The empirical gains are credible, but they do not by themselves settle the theoretical claims.

major comments (4)
  1. [Section 4, Assumption 4 (Eq. (28)) and Lemmas 1, 6] Assumption 4 is load-bearing and is not verified for the actual algorithm. It requires a uniform lower bound psi_uv^2 on delta_min(bar_U_l)^2 + delta_min(bar_V_l)^2 at every iteration and every decomposed layer. The text justifies this by Marchenko-Pastur theory, but that theory provides a high-probability lower bound for a single random matrix with i.i.d. entries, whereas bar_U and bar_V are server averages of data-dependent factor matrices that are correlated through training and periodically reinitialized. At reset times bar_V_l = 0, so the condition reduces to delta_min(U0)^2 >= psi_uv^2 for a single shared random initialization; no deterministic or uniform-over-rounds guarantee is given. Lemma 1 and Lemma 6 use this assumption to convert factored-space gradient norms into original-space gradient norms, so Theorem 1 is conditional on an unproven property. If delta_min is small or zero for some layer or round, the learning-rate restriction eta <= (psi_uv^2/2)^(1/(c-1)) forces eta to be very small or the bound to be vacuous. Please provide a proof for the averaged, correlated factors, weaken the assumption to a verifiable condition, or add empirical monitoring of delta_min across all layers and rounds.
  2. [Appendix D.3, Lemma 4 (Eq. (44)) and Lemma 5 (Eq. (51))] The proof of Lemma 4 uses Gamma_v^2 as if it were a deterministic upper bound on ||V||_F^2, but Lemma 2 only establishes an expectation bound, E||V_{i,l}^t||_F^2 <= Gamma_v^2. In Eq. (44), the step E||A_l||_F^2 <= eta^2 Gamma_v^2 E||sum||^2 is not justified without an almost-sure bound on ||bar_V_l^{t_e}||_F^2 or an independence argument that is not available. The same issue appears in Eq. (51) for Lemma 5, where the bound E||A_l||_F^2 <= eta^2 G^2 Gamma_v^4 is obtained by treating products of expectation-bounded norms as if they were almost-surely bounded. Since Lemma 4 feeds into Lemma 6 and hence into Theorem 1, the convergence proof needs repair, for example by deriving an almost-sure version of Lemma 2 or by adjusting the constants in the stated bounds.
  3. [Section 4, Remark 3 and Theorem 1 (Eq. (10))] The claim that FedMUD converges faster than FedLMT is not established by Theorem 1. The theorem bounds the average squared gradient norm by a constant depending on Gamma_u and Gamma_v, and Remark 2 notes that Gamma_u and Gamma_v grow with S. Remark 3 then concludes that FedLMT, which corresponds to a much larger S, converges slower. This is a comparison of upper-bound constants, not a proof of a rate separation. Two algorithms can have the same O(1/sqrt(T)) rate with different constants, and a tighter upper bound does not imply faster actual convergence. To substantiate the advertised speed-up, the paper should provide a lower-bound construction or formally prove a rate separation, or it should explicitly state the claim as a comparison of theoretical upper bounds rather than as a proven convergence-rate advantage.
  4. [Section 4, Remark 1 and Remark 3] The informal 'c -> 1' limit used to claim an O(1/sqrt(T)) rate is not a rigorous corollary of the theorem as stated. For fixed c in (1,2), the learning rate must satisfy eta <= (psi_uv^2/2)^(1/(c-1)); when psi_uv^2/2 < 1, the right-hand side tends to 0 as c approaches 1, so one cannot simply let c -> 1 after setting eta = 1/sqrt(T). A formal statement should specify how c depends on T or assume psi_uv is bounded away from 0 independently of T. In addition, Remark 3 reduces FedMUD to FedLMT by setting S >= T, W_l = 0, and initializing both U and V randomly, whereas the FedMUD algorithm described in Section 3.1 and Appendix D.1 initializes V to zero; the reduction is therefore from a variant of FedMUD rather than from the default method.
minor comments (5)
  1. [Appendix D.2, Eq. (27)] Assumption 3 appears to contain a typo: the condition ||U_l||_F^2 <= epsilon_v^2 should presumably read ||V_l||_F^2 <= epsilon_v^2, since the second inequality governs the V matrices.
  2. [Appendix A, Table 3] The entry for EF21-P on CIFAR-100 (IID) reads '28.1(±04.)' and should be '28.1(±0.4)'.
  3. [Section 5.3] The sentence beginning 'ADD decouples the multiplication...' uses 'ADD' instead of 'AAD'.
  4. [Appendix D.4, Eq. (71)] The final big-O simplification in Eq. (71) hides terms involving Gamma_u^2 + Gamma_v^2 and Gamma_u^2 Gamma_v^2 inside O(eta^(2-c))(1 + (L-rho)O(Gamma_u^4 + Gamma_v^4)); the authors should clarify how these terms are absorbed into the stated big-O notation.
  5. [Section 5.4 and Appendix D.1] The reset interval s is defined in rounds in the experiments, while the theory uses S iterations between resets; the paper should state the relationship (e.g., S = s * E) explicitly to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the FedMUD convergence proof is a conditional optimization analysis under explicit assumptions, and the FedLMT comparison is a limiting-case reduction rather than a fitted or self-citational loop.

full rationale

The central derivation (Theorem 1) is built from Lemmas 1-6, which are linear-algebra identities and drift bounds under Assumptions 1-4; no parameter is fitted to data and then reported as a prediction. Assumption 4 is a stated regularity condition, not a consequence of Theorem 1; the paper cites Marchenko-Pastur theory via FedLMT (Liu et al., 2024a) for plausibility, but the cited work is external, not authored by the present authors, so this is imported support rather than a self-citation chain. The relation between FedMUD and FedLMT in Remark 3 is a parameter identification (S >= T, W_l = 0, and random U,V reduces FedMUD to FedLMT), not a circular definition. The empirical comparison includes the authors' own FedBAT baseline, but FedBAT plays no role in the convergence proof and is not load-bearing. The main weaknesses are non-circularity concerns: Assumption 4 is not verified for server-averaged, training-correlated factor matrices (especially at reset times when V=0), and the speed-up over FedLMT in Remark 3 compares upper-bound constants rather than realized convergence rates. These are soundness and inference caveats, not reductions of the claimed result to its inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard smoothness and gradient assumptions plus one strong, unproven technical condition (Assumption 4). The method's hyperparameters (reset interval, initialization scale, compression budget) are hand-chosen and affect the reported performance. No new physical entities or forces are introduced.

free parameters (3)
  • Reset interval s = 1 (default), occasionally 2 or 4
    Controls how often low-rank updates are folded into frozen weights. It appears in the convergence bound through Gamma_u and Gamma_v, and the paper notes a slight empirical improvement for s=2 or 4 under heterogeneity.
  • Initialization scale a for sub-matrices = Tuned across {0.01, 0.05, 0.1, 0.5, 1, 5, 10}
    The paper shows accuracy is sensitive to this scale, and BKD requires larger values than standard low-rank decomposition. It is a hand-chosen hyperparameter that materially affects the reported results.
  • Compression ratio / rank r or block count k = 1/32 in main experiments; varying compression tested
    The compression budget is an external design choice that the method depends on. The paper compares methods at a fixed budget to make the comparison fair, but the choice of budget affects all accuracy numbers.
assumptions (5)
  • standard math Assumption 1: Each loss function f_i is differentiable and L_s-smooth.
    Invoked throughout the convergence proof, standard in distributed optimization.
  • standard math Assumption 2: Stochastic gradients are unbiased with bounded variance and norm.
    Standard in FL convergence analysis, used in Lemmas 2-6.
  • domain assumption Assumption 3: Norms of U and V are bounded during initialization and training, with epsilon_u, epsilon_v much smaller than kappa_u, kappa_v.
    The paper argues this holds exactly because periodic reinitialization prevents unbounded growth. It is plausible but depends on the finite local update steps.
  • ad hoc to paper Assumption 4: At least one of the averaged factor matrices U_bar_l and V_bar_l has a positive smallest singular value, uniformly bounded below by psi_uv, at every iteration.
    Load-bearing for Lemma 1 and Lemma 6. It is asserted to be supported by Marchenko-Pastur theory, but is not proven for aggregated, trained factors and may be fragile.
  • domain assumption For BKD, the matrix dimensions are assumed to factor as m = k a^2 and n = k b^2; padding or reshaping is used otherwise.
    The paper handles arbitrary dimensions by reshaping and truncating, which may introduce mild bias that is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Panaceas for Improving Low-Rank Decomposition in Communication-Efficient Federated Learning." pith.science (2026). https://pith.science/paper/RCLPCXOQ

@misc{pith2026250523176,
  author       = {Pith},
  title        = {Pith review of: The Panaceas for Improving Low-Rank Decomposition in Communication-Efficient Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RCLPCXOQ}},
  note         = {Machine review of arXiv:2505.23176}
}
read the original abstract

To improve the training efficiency of federated learning (FL), previous research has employed low-rank decomposition techniques to reduce communication overhead. In this paper, we seek to enhance the performance of these low-rank decomposition methods. Specifically, we focus on three key issues related to decomposition in FL: what to decompose, how to decompose, and how to aggregate. Subsequently, we introduce three novel techniques: Model Update Decomposition (MUD), Block-wise Kronecker Decomposition (BKD), and Aggregation-Aware Decomposition (AAD), each targeting a specific issue. These techniques are complementary and can be applied simultaneously to achieve optimal performance. Additionally, we provide a rigorous theoretical analysis to ensure the convergence of the proposed MUD. Extensive experimental results show that our approach achieves faster convergence and superior accuracy compared to relevant baseline methods. The code is available at https://github.com/Leopold1423/fedmud-icml25.

Figures

Figures reproduced from arXiv: 2505.23176 by the authors.

Figure 1
Figure 1. An illustration of three key issues related to decomposition in FL and their corresponding solutions: (a) What to decompose to minimize information loss? Existing methods decomposes the entire parameters, while MUD decomposes only the model update, effectively reducing information loss. (b) How to decompose to achieve a higher rank? Standard low-rank decomposition achieves rank(W) ≤ min r ≪ min{m, n} with (m+n)r par… view at source ↗
Figure 2
Figure 2. Convergence curves of different methods under the Non-IID-1 data distribution [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Accuracy of FedMUD under different reset intervals [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Accuracy of FedMUD under different initialization values [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Accuracy of FedMUD under different compression ratios. In [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Then, we evaluate the performance of FedMUD under the Non-IID-2 data distribution, considering various reset intervals, initialization values, and compression ratios. The experimental results, presented in [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Ablation on hyperparameters for the proposed methods under Non-IID-2 and IID data distribution [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 52 canonical work pages

  1. [1]

    Federated dynamic sparse training: Computing less, communicating less, yet learning better

    Bibikar, S., Vikalo, H., Wang, Z., and Chen, X. Federated dynamic sparse training: Computing less, communicating less, yet learning better. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2022

  2. [2]

    Large-scale machine learning with stochastic gradient descent

    Bottou, L. Large-scale machine learning with stochastic gradient descent. In International Conference on Computational Statistics (COMPSTAT), pp.\ 177--186. Physica-Verlag, 2010

  3. [3]

    Adaptive federated dropout: Improving communication efficiency and generalization for federated learning

    Bouacida, N., Hou, J., Zang, H., and Liu, X. Adaptive federated dropout: Improving communication efficiency and generalization for federated learning. In 2021 IEEE Conference on Computer Communications Workshops, INFOCOM Workshops , pp.\ 1--6. IEEE , 2021

  4. [4]

    B., and Talwalkar, A

    Caldas, S., Kone c n \'y , J., McMahan, H. B., and Talwalkar, A. Expanding the reach of federated learning by reducing client resource requirements. CoRR, abs/1812.07210, 2018

  5. [5]

    On generalization bounds of a family of recurrent neural networks

    Chen, M., Li, X., and Zhao, T. On generalization bounds of a family of recurrent neural networks. In The 23rd International Conference on Artificial Intelligence and Statistics (AISTATS), 2020

  6. [6]

    U., and Yurtsever, A

    Dadras, A., Stich, S. U., and Yurtsever, A. Personalized federated learning via low-rank matrix factorization. In Advances in Neural Information Processing Systems (NeurIPS) Workshop OPT, 2024

  7. [7]

    Dorfman, R., Vargaftik, S., Ben - Itzhak, Y., and Levy, K. Y. Docofl: Downlink compression for cross-device federated learning. In International Conference on Machine Learning (ICML), 2023

  8. [8]

    S., Kobyzev, I., Nia, V

    Edalati, A., Tahaei, M. S., Kobyzev, I., Nia, V. P., Clark, J. J., and Rezagholizadeh, M. Krona: Parameter efficient tuning with kronecker adapter. In Advances in Neural Information Processing Systems (NeurIPS) Workshop, 2023

Show all 58 references
  1. [9]

    Deep sparse rectifier neural networks

    Glorot, X., Bordes, A., and Bengio, Y. Deep sparse rectifier neural networks. In International Conference on Artificial Intelligence and Statistics (AISTATS), pp.\ 315--323, 2011

  2. [10]

    EF21-P and friends: Improved theoretical communication complexity for distributed optimization with bidirectional compression

    Gruntkowska, K., Tyurin, A., and Richt \' a rik, P. EF21-P and friends: Improved theoretical communication complexity for distributed optimization with bidirectional compression. In International Conference on Machine Learning (ICML), 2023

  3. [11]

    Lora+: Efficient low rank adaptation of large models

    Hayou, S., Ghosh, N., and Yu, B. Lora+: Efficient low rank adaptation of large models. In International Conference on Machine Learning (ICML), 2024

  4. [12]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR , pp.\ 770--778. IEEE Computer Society, 2016

  5. [13]

    H \" o nig, R., Zhao, Y., and Mullins, R. D. Dadaquant: Doubly-adaptive quantization for communication-efficient federated learning. In International Conference on Machine Learning (ICML), 2022

  6. [14]

    J., Shen, Y., Wallis, P., Allen - Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W

    Hu, E. J., Shen, Y., Wallis, P., Allen - Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022

  7. [15]

    Fedpara: Low-rank hadamard product for communication-efficient federated learning

    Hyeon - Woo, N., Ye - Bin, M., and Oh, T. Fedpara: Low-rank hadamard product for communication-efficient federated learning. In International Conference on Learning Representations (ICLR), 2022

  8. [16]

    and Szegedy, C

    Ioffe, S. and Szegedy, C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning (ICML), 2015

  9. [17]

    Sparse random networks for communication-efficient federated learning

    Isik, B., Pase, F., G \" u nd \" u z, D., Weissman, T., and Zorzi, M. Sparse random networks for communication-efficient federated learning. In International Conference on Learning Representations (ICLR), 2023

  10. [18]

    P., Rebjock, Q., Stich, S

    Karimireddy, S. P., Rebjock, Q., Stich, S. U., and Jaggi, M. Error feedback fixes signsgd and other gradient compression schemes. In International Conference on Machine Learning (ICML), 2019

  11. [19]

    B., Yu, F

    Kone c n \'y , J., McMahan, H. B., Yu, F. X., Richt \' a rik, P., Suresh, A. T., and Bacon, D. Federated learning: Strategies for improving communication efficiency. CoRR, 2016

  12. [20]

    and Hinton, G

    Krizhevsky, A. and Hinton, G. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009

  13. [21]

    and Yang, X

    Le, Y. and Yang, X. Tiny imagenet visual recognition challenge, 2015. Stanford CS231n Course Project

  14. [22]

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

    Li, Q., Diao, Y., Chen, Q., and He, B. Federated learning on non-iid data silos: An experimental study. In International Conference on Data Engineering (ICDE), 2022

  15. [23]

    Adaptive low-precision training for embeddings in click-through rate prediction

    Li, S., Guo, H., Hou, L., Zhang, W., Tang, X., Tang, R., Zhang, R., and Li, R. Adaptive low-precision training for embeddings in click-through rate prediction. In Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI) , Washington, DC, USA, 2023

  16. [24]

    Masked random noise for communication-efficient federated learning

    Li, S., Cheng, Y., Wang, H., Tang, X., Xu, S., Luo, W., Li, Y., Liu, D., He, X., and Li, R. Masked random noise for communication-efficient federated learning. In ACM International Conference on Multimedia (ACM MM), 2024 a

  17. [25]

    Embedding compression in recommender systems: A survey

    Li, S., Guo, H., Tang, X., Tang, R., Hou, L., Li, R., and Zhang, R. Embedding compression in recommender systems: A survey. ACM Comput. Surv. , 56 0 (5): 0 130:1--130:21, 2024 b

  18. [26]

    Mixed-precision embeddings for large-scale recommendation models

    Li, S., Hu, Z., Tang, X., Wang, H., Xu, S., Luo, W., Li, Y., He, X., and Li, R. Mixed-precision embeddings for large-scale recommendation models. arXiv preprint arXiv:2409.20305, 2024 c

  19. [27]

    Fedbat: Communication-efficient federated learning via learnable binarization

    Li, S., Xu, W., Wang, H., Tang, X., Qi, Y., Xu, S., Luo, W., Li, Y., He, X., and Li, R. Fedbat: Communication-efficient federated learning via learnable binarization. In International Conference on Machine Learning (ICML), 2024 d

  20. [28]

    Beyond zero initialization: Investigating the impact of non-zero initialization on lora fine-tuning dynamics

    Li, S., Luo, X., Tang, X., Wang, H., Chen, H., Luo, W., Li, Y., He, X., and Li, R. Beyond zero initialization: Investigating the impact of non-zero initialization on lora fine-tuning dynamics. In International Conference on Machine Learning (ICML), 2025 a

  21. [29]

    Bora: Towards more expressive low-rank adaptation with block diversity

    Li, S., Luo, X., Wang, H., Tang, X., Cui, Z., Liu, D., Li, Y., xiuqiang He, and Li, R. Bora: Towards more expressive low-rank adaptation with block diversity. CoRR, 2025 b

  22. [30]

    On the convergence of fedavg on non-iid data

    Li, X., Huang, K., Yang, W., Wang, S., and Zhang, Z. On the convergence of fedavg on non-iid data. In International Conference on Learning Representations (ICLR), 2020

  23. [31]

    Relora: High-rank training through low-rank updates

    Lialin, V., Muckatira, S., Shivagunde, N., and Rumshisky, A. Relora: High-rank training through low-rank updates. In International Conference on Learning Representations (ICLR), 2024

  24. [32]

    Liu, J., Zhou, Y., Wu, D., Hu, M., Guizani, M., and Sheng, Q. Z. Fedlmt: Tackling system heterogeneity of federated learning via low-rank model training with theoretical guarantees. In International Conference on Machine Learning (ICML), 2024 a

  25. [33]

    Mgr: Multi-generator based rationalization

    Liu, W., Wang, H., Wang, J., Li, R., Li, X., Zhang, Y., and Qiu, Y. Mgr: Multi-generator based rationalization. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 12771--12787, 2023 a

  26. [34]

    D-separation for causal self-explanation

    Liu, W., Wang, J., Wang, H., Li, R., Deng, Z., Zhang, Y., and Qiu, Y. D-separation for causal self-explanation. Advances in Neural Information Processing Systems, 36: 0 43620--43633, 2023 b

  27. [35]

    Is the mmi criterion necessary for interpretability? degenerating non-causal features to plain noise for self-rationalization

    Liu, W., Deng, Z., Niu, Z., Wang, J., Wang, H., Zhang, Y., and Li, R. Is the mmi criterion necessary for interpretability? degenerating non-causal features to plain noise for self-rationalization. Advances in Neural Information Processing Systems, 37: 0 117636--117656, 2024 b

  28. [36]

    Enhancing the rationale-input alignment for self-explaining rationalization

    Liu, W., Wang, H., Wang, J., Deng, Z., Zhang, Y., Wang, C., and Li, R. Enhancing the rationale-input alignment for self-explaining rationalization. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pp.\ 2218--2230. IEEE, 2024 c

  29. [37]

    Exploring practical gaps in using cross entropy to implement maximum mutual information criterion for rationalization

    Liu, W., Deng, Z., Niu, Z., Wang, J., Wang, H., and Li, R. Exploring practical gaps in using cross entropy to implement maximum mutual information criterion for rationalization. Transactions of the Association for Computational Linguistics, 13: 0 577--594, 2025 a

  30. [38]

    Breaking free from mmi: A new frontier in rationalization by probing input utilization

    Liu, W., Deng, Z., Niu, Z., Wang, J., Wang, H., Zeng, Z., and Li, R. Breaking free from mmi: A new frontier in rationalization by probing input utilization. In The Thirteenth International Conference on Learning Representations, 2025 b

  31. [39]

    Malik, O. A. and Becker, S. Low-rank tucker decomposition of large tensors using tensorsketch. In Advances in Neural Information Processing Systems (NeurIPS), 2018

  32. [40]

    A survey on lora of large language models

    Mao, Y., Ge, Y., Fan, Y., Xu, W., Mi, Y., Hu, Z., and Gao, Y. A survey on lora of large language models. CoRR, 2024

  33. [41]

    McMahan, B., Moore, E., Ramage, D., Hampson, S., and y Arcas, B. A. Communication-efficient learning of deep networks from decentralized data. In International Conference on Artificial Intelligence and Statistics (AISTATS), 2017

  34. [42]

    Resource-adaptive federated learning with all-in-one neural composition

    Mei, Y., Guo, P., Zhou, M., and Patel, V. Resource-adaptive federated learning with all-in-one neural composition. In Advances in Neural Information Processing Systems (NeurIPS), 2022

  35. [43]

    Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., and Ng, A. Y. Reading digits in natural images with unsupervised feature learning. In Advances in Neural Information Processing Systems (NeurIPS) Workshop, 2011

  36. [44]

    Novikov, A., Izmailov, P., Khrulkov, V., Figurnov, M., and Oseledets, I. V. Tensor train decomposition on tensorflow (T3F) . CoRR, 2018

  37. [45]

    Qiu, X., Fern \' a ndez - Marqu \' e s, J., de Gusmao, P. P. B., Gao, Y., Parcollet, T., and Lane, N. D. Zerofl: Efficient on-device training for federated learning with local sparsity. In International Conference on Learning Representations (ICLR), 2022

  38. [46]

    Fedpaq: A communication-efficient federated learning method with periodic averaging and quantization

    Reisizadeh, A., Mokhtari, A., Hassani, H., Jadbabaie, A., and Pedarsani, R. Fedpaq: A communication-efficient federated learning method with periodic averaging and quantization. In International Conference on Artificial Intelligence and Statistics (AISTATS), 2020

  39. [47]

    N., Kingsbury, B., Sindhwani, V., Arisoy, E., and Ramabhadran, B

    Sainath, T. N., Kingsbury, B., Sindhwani, V., Arisoy, E., and Ramabhadran, B. Low-rank matrix factorization for deep neural network training with high-dimensional output targets. In IEEE International Conference on Acoustics, Speech and Signal Processing, (ICASSP) , 2013

  40. [48]

    Stripelis, D., Gupta, U., Ver Steeg, G., and Ambite, J. L. Federated progressive sparsification (purge-merge-tune)+. In Advances in Neural Information Processing Systems (NeurIPS) Workshop, 2022

  41. [49]

    Improving lora in privacy-preserving federated learning

    Sun, Y., Li, Z., Li, Y., and Ding, B. Improving lora in privacy-preserving federated learning. In International Conference on Learning Representations (ICLR), 2024

  42. [50]

    Doublesqueeze: Parallel stochastic gradient descent with double-pass error-compensated compression

    Tang, H., Yu, C., Lian, X., Zhang, T., and Liu, J. Doublesqueeze: Parallel stochastic gradient descent with double-pass error-compensated compression. In International Conference on Machine Learning (ICML), 2019

  43. [51]

    Decoupling general and personalized knowledge in federated learning via additive and low-rank decomposition

    Wu, X., Liu, X., Niu, J., Wang, H., Tang, S., Zhu, G., and Su, H. Decoupling general and personalized knowledge in federated learning via additive and low-rank decomposition. In ACM International Conference on Multimedia (ACM MM), 2024

  44. [52]

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

    Xiao, H., Rasul, K., and Vollgraf, R. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. CoRR, 2017

  45. [53]

    Communication-efficient federated learning with binary neural networks

    Yang, Y., Zhang, Z., and Yang, Q. Communication-efficient federated learning with binary neural networks. IEEE J. Sel. Areas Commun. , 39 0 (12): 0 3836--3850, 2021

  46. [54]

    Fedhm: Efficient federated learning for heterogeneous models via low-rank factorization

    Yao, D., Pan, W., Wan, Y., Jin, H., and Sun, L. Fedhm: Efficient federated learning for heterogeneous models via low-rank factorization. CoRR, 2021

  47. [55]

    Parallel restarted SGD with faster convergence and less communication: Demystifying why model averaging works for deep learning

    Yu, H., Yang, S., and Zhu, S. Parallel restarted SGD with faster convergence and less communication: Demystifying why model averaging works for deep learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019

  48. [56]

    H., Hoang, T

    Yurochkin, M., Agarwal, M., Ghosh, S., Greenewald, K. H., Hoang, T. N., and Khazaeni, Y. Bayesian nonparametric federated learning of neural networks. In International Conference on Machine Learning (ICML), 2019

  49. [57]

    M., and Gu, Q

    Zou, D., Long, P. M., and Gu, Q. On the global convergence of training deep linear resnets. In International Conference on Learning Representations (ICLR), 2020

  50. [58]

    write newline

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

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.