Pith. sign in

REVIEW 4 major objections 4 minor 36 references

Federated Low-Rank Tensor Estimation for Multimodal Image Reconstruction

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

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

Pith's one-line read A federated reconstruction method transmits only Tucker components and still beats FIRM and Top-k sparsification on synthetic tomography.

desk verdict A coherent, correctly-mathed extension of federated tensor aggregation to multimodal tomography, but the empirical case needs error bars, a non-oracle curve, and a check on the heterogeneous-rank communication claim before the results can be trusted. read the letter →

arxiv 2502.02761 v1 pith:4LGK27HH submitted 2025-02-04 cs.LG cs.CVcs.DC

Anh Van Nguyen , Diego Klabjan , Minseok Ryu , Kibaek Kim , Zichao Di This is my paper · ORCID
classification cs.LGcs.CVcs.DC
keywords federatedlearningtensordecompositionTuckerimagereconstructiontomographiccommunicationcompressionjointfactorizationrandomizedsketching
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

This paper proposes a way to run federated image reconstruction, where multiple sites collaborate without sharing raw scans, by exchanging small Tucker decomposition components instead of full images. The method, CompJF and its faster variant CompRandJF, combines clients' factor matrices on the server through joint factorization, so the server never has to reconstruct full-size tensors. On synthetic tomographic data, the authors report that these methods achieve higher SSIM and PSNR than the FIRM baseline and surpass Top-k sparsification and CSR encoding in communication efficiency. The approach also supports heterogeneous ranks, letting each client choose its own Tucker rank based on bandwidth or prior knowledge.

What carries the argument

The load-bearing identity is a relation between the concatenated mode-$k$ unfolding of the client tensors and the block matrix formed by each client's factor matrix times its core unfolding: they share the same left singular vectors up to a unitary block diagonal transformation (Lemma II.1 from the joint-factorization literature). This lets the server update the shared factor matrices by doing an SVD on a matrix of size $n \times N \prod_{j\ne k} r_j$, or, with Gaussian sketching, a matrix of size $n \times r_k$, instead of reconstructing and decomposing each full tensor. Lemma III.1 proves that sketching the concatenated unfolding is equivalent to sketching each client's components with independent Gaussian matrices. Proposition 1 supplies rank thresholds for $d=2$ and $d=3$ that guarantee each round communicates fewer numbers than a full tensor, thresholds looser than the prior bound from the literature.

What would settle it

Run the synthetic tomography benchmark with a ground-truth phantom whose Tucker rank exceeds the chosen $r$ (for example, a phantom with many fine details), and check whether CompJF and CompRandJF still beat FIRM in SSIM; if they no longer do, the low-rank joint factorization is the source of the gain rather than a generally valid regularizer. Alternatively, track the distance between the server's factor subspace and the dominant subspace of the clean tensor across rounds: if the distance grows instead of shrinking, the alternating scheme is not converging to the true low-rank structure.

Watch

Extended reading notes

Core claim

The paper claims that a federated Tucker decomposition with a shared factor subspace and client-specific cores, aggregated by joint factorization, can solve multimodal tomographic reconstruction more accurately and with less communication than existing federated reconstruction methods. In CompJF, the server computes new factor matrices as the leading left singular vectors of a block matrix assembled from clients' cores and factor matrices, avoiding any full-tensor reconstruction. CompRandJF replaces that SVD with a randomized sketch and QR decomposition to reduce server cost. The FIRM multimodality constraint, that one client's reconstruction is a weighted sum of the others', is enforced directly on the core tensors after re-projecting them into the updated subspace. On synthetic Shepp-Logan phantoms with speckle noise, the authors report better SSIM and PSNR than FIRM and FullDecomp, and higher Gamma Communication Efficiency than FIRM with CSR or Top-k compression, while still allowing heterogeneous ranks across clients.

Load-bearing premise

The method assumes the federated alternating optimization, where clients project onto their own Tucker subspaces and the server replaces the shared subspace each round, keeps making useful progress even though no convergence analysis is provided and a naive averaging variant collapses during training.

Editorial extensions

If this is right

  • Federated tomography clients can send low-rank components in both upstream and downstream directions, reducing per-round communication from $O(n^d)$ to $O(r^d + dnr)$ while Proposition 1 identifies ranks that guarantee compression for $d=2$ and $d=3$.
  • The method's robustness to noise and undersampling means ill-posed inverse problems in distributed settings, such as collaborative computed tomography, can be regularized without a central full-data reconstruction step.
  • Heterogeneous ranks are supported, so clients with different bandwidths or domain knowledge can pick their own Tucker ranks without breaking the joint-factorization aggregation.
  • The FIRM multimodality constraint transfers cleanly to the core space, so the XRF/XRT linear relationship is enforced while communication still consists of compressed components.

Reading between the lines

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

  • The joint-factorization aggregation scheme is not specific to tomography; it could apply to any federated low-rank tensor estimation problem where the server needs to combine client subspaces, such as distributed recommendation or neuroimaging.
  • CompRandJF samples a fresh Gaussian sketch each round, so a principled choice of sketch size and number of iterations could trade off the observed degradation at small ranks, potentially extending its useful range below the reported $r=40$ threshold.
  • The failure of CompAvg suggests that averaging factor matrices independently destroys alignment; understanding exactly why averaging fails could inform other federated subspace methods that rely on shared latent factors.
  • The experiments assume the coefficient $c_j$ in the multimodality constraint is well characterized; testing sensitivity to misspecified $c_j$ would show how the method behaves when the physical model is only approximately known.
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 / 4 minor

Summary. The paper proposes a federated low-rank tensor estimation framework for multimodal image reconstruction. Clients perform local projected-gradient updates with Tucker decomposition (via ST-HOSVD) and send the resulting core tensors and factor matrices to the server; the server aggregates them through joint factorization, optionally accelerated by randomized sketching, giving the CompJF and CompRandJF variants. The server then enforces the linear multimodality constraint on the cores. The authors provide a compression-ratio bound (Proposition 1), a randomized sketching lemma (Lemma III.1), complexity comparisons (Table I), and synthetic tomographic reconstruction experiments against FIRM, FullDecomp, Top-k sparsification, and CSR encoding, claiming superior reconstruction quality and communication efficiency.

Significance. If validated, the paper contributes a coherent aggregation scheme for federated tensor-based inverse problems that avoids full-size tensor reconstruction on the server, supports heterogeneous Tucker ranks across clients, and comes with concrete communication-complexity savings. The mathematical statements I checked are correct: Lemma III.1 is a valid extension of the Gaussian sketching argument, and Proposition 1 provides genuine improvements over the earlier Dai et al. bound for d=2 and d=3. The algorithms are specified in enough detail to be reimplemented. However, the headline empirical claims currently rest on oracle-selected epochs and single-seed synthetic experiments, so the practical significance is conditional on the evaluation being placed on a sounder footing.

major comments (4)
  1. [Section IV-C, Fig. 3(a) and Fig. 2] The 'best result overall' in Fig. 3(a) and the statement in Fig. 2 that 'the reconstructed images correspond to the highest SSIM values' select the epoch using ground truth. Such oracle selection cannot be implemented when ground truth is unavailable, and it systematically inflates the apparent reconstruction quality of any method that overfits during optimization. The realizable comparison is the early-stopping version in Fig. 3(b); the paper should either make Fig. 3(b) the primary evidence or explicitly justify why oracle epoch selection is an accepted evaluation convention in this setting.
  2. [Section IV-A and Figs. 1, 3, 4] The main reconstruction-quality and communication-efficiency results are reported without error bars or multiple independent runs. This matters because CompRandJF uses Gaussian random sketching, and the initial conditions and noise realizations are not fixed or analyzed. The heterogeneous-rank experiment in Fig. 5 correctly reports mean and one standard deviation over 10 simulations, and the same reporting standard should be applied to Figs. 1, 3, and 4 so the reader can judge whether the observed gaps are robust or seed-specific.
  3. [Algorithm 1 and Section V] There is no convergence or stability analysis for the alternating update in which the server re-estimates the shared factor subspace S_k(t) each round and re-projects each client's core onto that new subspace in line 12. Because the feasible set changes between rounds, a client's local progress can be discarded by the re-projection, so the optimization is not a standard projected-gradient method. Appendix A reinforces this concern by showing that a closely related aggregation variant, CompAvg, fails to train entirely. I am not asking for a full convergence theorem, but the paper should provide at least a simplified descent or stationarity argument, or an empirical stability study over initializations and rank choices, before the 'superior reconstruction quality' claim is accepted.
  4. [Section IV-C.2 and Abstract] The paper's own results show that the proposed methods outperform FIRM only for Tucker ranks r >= 40, and that at r=10 the drop in SSIM 'overshadows the improvement in communication' (Fig. 4a). The abstract and conclusion nevertheless state that the method 'achieves superior reconstruction quality' without this rank qualification. The claims should be qualified to the rank regime in which the experiments actually support them, or additional results should be provided for the low-rank regime.
minor comments (4)
  1. [Proposition 1] The statement 'If n >= 3' is imprecise for d=3: at n=3 the bound n((n-3)/n)^(1/3) is zero, and the proof as written divides by n-3. The proposition should state n > 3, or the d=3 case should be handled separately.
  2. [Algorithm 3, line 2] The phrase 'i'th i.i.d. sample of a Gaussian matrix' is unclear; it should say that for each client i and each mode k, an independent r_k x prod_{j != k} r_j Gaussian matrix is drawn. The text should also clarify whether fresh sketches are drawn every epoch or reused.
  3. [Table I] The complexity expressions for FullDecomp contain a parenthetical grouping that is hard to parse (for example, n^{d-k+2}r^{k-1} appears without a clear multiplication structure). Rewriting the expressions with explicit parentheses and a short derivation would improve verifiability.
  4. [Section IV-B.2] The adaptation of GCE replaces test accuracy with SSIM, but SSIM is not an accuracy measure, and the choice of gamma=0.01 deserves more discussion. In particular, because the denominator contains (1-SSIM)^gamma, the metric can become very sensitive when SSIM is close to 1; the paper should justify that the chosen gamma and SSIM range make the reported gains meaningful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proposed algorithm is evaluated empirically and its theoretical components are self-contained or attributed to non-overlapping prior work.

full rationale

The paper contains no derivation step in which a predicted quantity is equivalent by construction to an input. The core contribution is algorithmic: Algorithm 1 combines projected gradient descent with ST-HOSVD (lines 4-6), server-side joint factorization (lines 9-10), core re-projection (line 12), and the FIRM-style core update (lines 14-16). Each component is either defined in the paper or taken from cited prior work; no component is fitted to the target reconstruction-quality claim. Proposition 1 is a self-contained algebraic bound on the compression ratio phi = n^d/(r^d + dnr); its proof is a direct inequality manipulation and does not assume the conclusion. Lemma III.1 is a short linear-algebra identity following from the mode-k unfolding of a Tucker decomposition and properties of Gaussian matrices; it is proved in the text. The randomized joint factorization (Algorithm 3) is adapted from Rand-Tucker [34], and the joint-factorization lemma (Lemma II.1) is taken from Gao et al. [7], which is not a self-citation. FIRM [2] is a self-citation with overlapping authors (Ryu, Di, Kim), but it is used as a problem formulation, a baseline, and a component update; the paper's central claims are supported by numerical experiments measuring SSIM, PSNR, and GCE, not by an appeal to FIRM's authority. The convergence guarantee of FIRM is not transferred to CompJF/CompRandJF; the paper explicitly defers convergence analysis to future work. Appendix A's CompAvg failure is an acknowledged empirical limitation of a naive aggregation variant, not a circular step. The experimental comparisons, including the top-k/CSR baselines and heterogeneous-rank settings, are external measurements. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' own prior work. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The method rests on standard tensor algebra lemmas, an inherited physical constraint, a low-rank assumption on the data, a randomized sketching accuracy assumption, and an unproved stability assumption for the federated alternating procedure. No new physical entities are introduced.

free parameters (2)
  • Tucker rank r (per mode) = Tested at {10,20,40,60,80,100}; no automatic selection rule
    The method's reconstruction quality and compression ratio both depend strongly on r. Proposition 1 bounds r for positive compression, but the optimal r for quality is chosen by hand per experiment and not fitted to data.
  • Learning rate eta = Not reported in the experimental section
    Local gradient descent step size in Algorithm 1; its value affects convergence and stability but is not specified in the experiments, making it a hidden free parameter.
assumptions (5)
  • standard math The modes of the concatenated tensor have the same left singular space as the joint factorization matrix Y (Lemma II.1), cited from Gao et al. and not re-proved in this paper.
    The server aggregation in Algorithm 2 relies on this lemma to avoid reconstructing full tensors. If the lemma were false, the shared factor computation would be wrong.
  • domain assumption The linear multimodality constraint X_N = sum_{j<N} c_j X_j holds with known coefficients c_j.
    This constraint is inherited from FIRM and is used in the core-space update (lines 14-16). If the physical model is inaccurate, the update enforces a wrong relationship between modalities.
  • domain assumption The ground-truth images are approximately low multilinear rank, so projecting onto a shared rank-(r,...,r) Tucker subspace preserves the signal.
    The method's regularizing effect depends on the Shepp-Logan phantom being low-rank; no rank analysis of the ground truth is provided.
  • ad hoc to paper A single-pass randomized sketch (QR of Y_k Omega, no power iteration or oversampling) preserves enough of the column space for accurate joint factorization at the tested ranks.
    Algorithm 3 uses a random projection without typical randomized-SVD refinements. The paper acknowledges that for small r the approximation degrades (Section III-B), making this a load-bearing design choice.
  • domain assumption The alternating local and global updates converge or at least remain stable across rounds.
    No convergence analysis is provided; the appendix shows that a closely related averaging scheme (CompAvg) fails to train, indicating sensitivity to the aggregation design.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Federated Low-Rank Tensor Estimation for Multimodal Image Reconstruction." pith.science (2026). https://pith.science/paper/4LGK27HH

@misc{pith2026250202761,
  author       = {Pith},
  title        = {Pith review of: Federated Low-Rank Tensor Estimation for Multimodal Image Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4LGK27HH}},
  note         = {Machine review of arXiv:2502.02761}
}
read the original abstract

Low-rank tensor estimation offers a powerful approach to addressing high-dimensional data challenges and can substantially improve solutions to ill-posed inverse problems, such as image reconstruction under noisy or undersampled conditions. Meanwhile, tensor decomposition has gained prominence in federated learning (FL) due to its effectiveness in exploiting latent space structure and its capacity to enhance communication efficiency. In this paper, we present a federated image reconstruction method that applies Tucker decomposition, incorporating joint factorization and randomized sketching to manage large-scale, multimodal data. Our approach avoids reconstructing full-size tensors and supports heterogeneous ranks, allowing clients to select personalized decomposition ranks based on prior knowledge or communication capacity. Numerical results demonstrate that our method achieves superior reconstruction quality and communication compression compared to existing approaches, thereby highlighting its potential for multimodal inverse problems in the FL setting.

Figures

Figures reproduced from arXiv: 2502.02761 by the authors.

Figure 1
Figure 1. Evaluating the performance of various methods given the strongest noise level (σ = 0.1). For CompJF, CompRandJF and FullDecomp, we present the results for Tucker rank r = 100, the largest value that guarantee communication compression. The results for other ranks r ≥ 40 is similar. reconstruction starts or b) resampled every epoch prior to upstream communication. B. Performance metrics 1) Reconstruction quality: We … view at source ↗
Figure 2
Figure 2. Ground-truth and reconstructed images from various methods for client N with XRT imaging modality. The images are reconstructed from noisy data (σ = 0.1). For FullDecomp, CompJF, and CompRandJF, we use Tucker rank r = 100. The reconstructed images correspond to the highest SSIM values. 2) Stable performance across Tucker ranks: In addition to better reconstruction, CompJF and CompRandJF are able to maintain good per… view at source ↗
Figure 3
Figure 3. Performance of CompJF and CompRandJF for varying Tucker decomposition ranks and varying level of noise. (a) Includes the highest SSIM value, averaged across clients, obtained by the approaches. (b) Includes the average SSIM obtained when the discrepancy principle is satisfied. With early stopping condition, all three approaches terminate at the same epoch. 3) Communication compression: We next illustrate the superio… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance of our methods, CompJF and CompRandJF, Top-k and CSR encoding in balancing reconstruction quality with communication compression. a) GCE is computed using the SSIM values when the early stopping is satisfied and is averaged across clients. b) For CompJF and…
Figure 6
Figure 6. Figure 6: Evaluating the performance of various methods given noisy data (σ = 0.1). For CompJF, CompRandJF and FullDecomp, Tucker rank is r = 100. training compared to other methods. This is likely because the factor matrices are obtained via tensor decomposition after the full-…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages

  1. [1]

    Tensor regression using low-rank and sparse tucker decompositions

    Talal Ahmed, Haroon Raja, and Waheed U Bajwa. Tensor regression using low-rank and sparse tucker decompositions. SIAM Journal on Mathematics of Data Science , 2(4):944–966, 2020

    work page 2020

  2. [2]

    FIRM: Federated Image Reconstruction using Multimodal Tomographic Data

    Geunyeong Byeon, Minseok Ryu, Zichao Wendy Di, and Kibaek Kim. Firm: Federated image reconstruction using multimodal tomographic data. arXiv preprint arXiv:2501.05642 , 2025

    work page Pith review arXiv 2025

  3. [3]

    Federated condition generalization on low-dose ct reconstruction via cross-domain learning

    Shixuan Chen, Boxuan Cao, Yinda Du, Yaoduo Zhang, Ji He, Zhaoying Bian, Dong Zeng, and Jianhua Ma. Federated condition generalization on low-dose ct reconstruction via cross-domain learning. In Interna- tional Conference on Medical Image Computing and Computer-Assisted Intervention, pages 47–56. Springer, 2023

    work page 2023

  4. [4]

    Deep learning model compression with rank reduction in tensor decomposition

    Wei Dai, Jicong Fan, Yiming Miao, and Kai Hwang. Deep learning model compression with rank reduction in tensor decomposition. IEEE Transactions on Neural Networks and Learning Systems , 2023

    work page 2023

  5. [5]

    Optimization-based approach for joint x-ray fluorescence and transmission tomographic inversion

    Zichao Di, Sven Leyffer, and Stefan M Wild. Optimization-based approach for joint x-ray fluorescence and transmission tomographic inversion. SIAM Journal on Imaging Sciences , 9(1):1–23, 2016

    work page 2016

  6. [6]

    2D & 3D Shepp- Logan phantom standards for MRI

    H Michael Gach, Costin Tanase, and Fernando Boada. 2D & 3D Shepp- Logan phantom standards for MRI. pages 521–526, 2008

    work page 2008

  7. [7]

    Federated tensor decomposition-based feature extraction approach for industrial iot

    Yuan Gao, Guangming Zhang, Chunchun Zhang, Jinke Wang, Laurence T Yang, and Yaliang Zhao. Federated tensor decomposition-based feature extraction approach for industrial iot. IEEE Transactions on Industrial Informatics, 17(12):8541–8549, 2021

    work page 2021

  8. [8]

    Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions

    Nathan Halko, Per-Gunnar Martinsson, and Joel A Tropp. Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions. SIAM Review, 53(2):217–288, 2011

    work page 2011

Show all 36 references
  1. [9]

    An optimal statistical and computational framework for generalized tensor estimation

    Rungang Han, Rebecca Willett, and Anru R Zhang. An optimal statistical and computational framework for generalized tensor estimation. The Annals of Statistics , 50(1):1–29, 2022

    2022

  2. [10]

    Attribute-enhanced face recognition with neural tensor fusion networks

    Guosheng Hu, Yang Hua, Yang Yuan, Zhihong Zhang, Zheng Lu, Sankha S Mukherjee, Timothy M Hospedales, Neil M Robertson, and Yongxin Yang. Attribute-enhanced face recognition with neural tensor fusion networks. In Proceedings of the IEEE International Conference on Computer Visi...

    2017

  3. [11]

    Low-rank tensor regression for x-ray tomography

    Sanket R Jantre and Zichao Wendy Di. Low-rank tensor regression for x-ray tomography. In 2021 IEEE International Conference on Image Processing (ICIP), pages 2833–2837. IEEE, 2021

    2021

  4. [12]

    Tensor decompositions and applications

    Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM Review, 51(3):455–500, 2009

    2009

  5. [13]

    Communication-efficient federated learning for resource-constrained edge devices

    Guangchen Lan, Xiao-Yang Liu, Yijing Zhang, and Xiaodong Wang. Communication-efficient federated learning for resource-constrained edge devices. IEEE Transactions on Machine Learning in Communications and Networking, 2023

    2023

  6. [14]

    Semi-centralized federated learning network for low-dose CT imaging

    Danyang Li, Hao Wang, Jingyi Liao, Mingqiang Meng, Cuidie Zeng, Yuting Wang, Dong Zeng, and Jianhua Ma. Semi-centralized federated learning network for low-dose CT imaging. In Medical Imaging 2023: Physics of Medical Imaging , volume 12463, pages 1024–1028. SPIE, 2023

    2023

  7. [15]

    Tucker tensor regression and neuroimaging analysis

    Xiaoshan Li, Da Xu, Hua Zhou, and Lexin Li. Tucker tensor regression and neuroimaging analysis. Statistics in Biosciences , 10:520–545, 2018

    2018

  8. [16]

    Marvel: Towards efficient federated learning on IoT devices

    Libin Liu and Xiuting Xu. Marvel: Towards efficient federated learning on IoT devices. Computer Networks, 245:110375, 2024

    2024

  9. [17]

    Fedzip: A compression framework for communication-efficient federated learning

    Amirhossein Malekijoo, Mohammad Javad Fadaeieslam, Hanieh Maleki- jou, Morteza Homayounfar, Farshid Alizadeh-Shabdiz, and Reza Rawas- sizadeh. Fedzip: A compression framework for communication-efficient federated learning. arXiv preprint arXiv:2102.01593 , 2021

    2021 arXiv

  10. [18]

    Communication-efficient learning of deep networks from decentralized data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics , pages 1273–1282. PMLR, 2017

    2017

  11. [19]

    Interpretable multimodal tucker fusion model with information filtering for multimodal sentiment analysis

    Xin Nie, Laurence T Yang, Zhe Li, Xianjun Deng, Fulan Fan, and Zecan Yang. Interpretable multimodal tucker fusion model with information filtering for multimodal sentiment analysis. IEEE Transactions on Computational Social Systems , 2024

    2024

  12. [20]

    Communication-efficient federated low-rank update algorithm and its connection to implicit regularization

    Haemin Park and Diego Klabjan. Communication-efficient federated low-rank update algorithm and its connection to implicit regularization. arXiv preprint arXiv:2409.12371 , 2024

    2024

  13. [21]

    FDLoRA: Personalized federated learning of large language model via dual LoRA tuning

    Jiaxing Qi, Zhongzhi Luan, Shaohan Huang, Carol Fung, Hailong Yang, and Depei Qian. FDLoRA: Personalized federated learning of large language model via dual LoRA tuning. arXiv preprint arXiv:2406.07925, 2024

    2024 arXiv

  14. [22]

    Federated learning of tensor generalized linear models with low separation rank

    Jose Hoyos Sanchez, Batoul Taki, Waheed U Bajwa, and Anand D Sarwate. Federated learning of tensor generalized linear models with low separation rank. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 2136–2140. IEEE, 2024

    2024

  15. [23]

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

    Felix Sattler, Simon Wiedemann, Klaus-Robert Müller, and Wojciech Samek. Robust and communication-efficient federated learning from non-iid data. IEEE Transactions on Neural Networks and Learning Systems, 31(9):3400–3413, 2019

    2019

  16. [24]

    Tensor- based formulation and nuclear norm regularization for multienergy computed tomography

    Oguz Semerci, Ning Hao, Misha E Kilmer, and Eric L Miller. Tensor- based formulation and nuclear norm regularization for multienergy computed tomography. IEEE Transactions on Image Processing , 23(4):1678–1693, 2014

    2014

  17. [25]

    Federated learning via decentralized dataset distillation in resource-constrained edge environments

    Rui Song, Dai Liu, Dave Zhenyu Chen, Andreas Festag, Carsten Trinitis, Martin Schulz, and Alois Knoll. Federated learning via decentralized dataset distillation in resource-constrained edge environments. In 2023 International Joint Conference on Neural Networks (IJCNN) , pages...

    2023

  18. [26]

    Structured data fusion

    Laurent Sorber, Marc Van Barel, and Lieven De Lathauwer. Structured data fusion. IEEE Journal of Selected Topics in Signal Processing , 9(4):586–600, 2015

    2015

  19. [27]

    A new truncation strategy for the higher-order singular value decomposition

    Nick Vannieuwenhoven, Raf Vandebril, and Karl Meerbergen. A new truncation strategy for the higher-order singular value decomposition. SIAM Journal on Scientific Computing , 34(2):A1027–A1052, 2012

    2012

  20. [28]

    Tensor-based recurrent neural network and multi-modal prediction with its applications in traffic network management

    Qing Wu, Zhe Jiang, Kewei Hong, Huazhong Liu, Laurence T Yang, and Jihong Ding. Tensor-based recurrent neural network and multi-modal prediction with its applications in traffic network management. IEEE Transactions on Network and Service Management , 18(1):780–792, 2021

    2021

  21. [29]

    Tenalign: Joint tensor alignment and coupled factorization

    Yunshu Wu, Uday Singh Saini, Jia Chen, and Evangelos E Papalexakis. Tenalign: Joint tensor alignment and coupled factorization. In 2022 IEEE International Conference on Data Mining (ICDM) , pages 568–577. IEEE, 2022

    2022

  22. [30]

    Hypernetwork-based physics-driven personalized federated learning for CT imaging

    Ziyuan Yang, Wenjun Xia, Zexin Lu, Yingyu Chen, Xiaoxiao Li, and Yi Zhang. Hypernetwork-based physics-driven personalized federated learning for CT imaging. IEEE Transactions on Neural Networks and Learning Systems, 2023

    2023

  23. [31]

    FedLoRA: Model- heterogeneous personalized federated learning with LoRA tuning

    Liping Yi, Han Yu, Gang Wang, and Xiaoguang Liu. FedLoRA: Model- heterogeneous personalized federated learning with LoRA tuning. arXiv preprint arXiv:2310.13283, 2023

    2023 arXiv

  24. [32]

    Federated learning using coupled tensor train decomposition

    Xiangtao Zhang, Eleftherios Kofidis, Ce Zhu, Le Zhang, and Yipeng Liu. Federated learning using coupled tensor train decomposition. arXiv preprint arXiv:2403.02898, 2024

    2024 arXiv

  25. [33]

    A distributed hierarchical deep computation model for federated learning in edge computing

    Haifeng Zheng, Min Gao, Zhizhang Chen, and Xinxin Feng. A distributed hierarchical deep computation model for federated learning in edge computing. IEEE Transactions on Industrial Informatics , 17(12):7946– 7956, 2021

    2021

  26. [34]

    Decomposition of big tensors with low multilinear rank

    Guoxu Zhou, Andrzej Cichocki, and Shengli Xie. Decomposition of big tensors with low multilinear rank. arXiv preprint arXiv:1412.1885 , 2014

    2014 arXiv

  27. [35]

    Tensor regression with applications in neuroimaging data analysis

    Hua Zhou, Lexin Li, and Hongtu Zhu. Tensor regression with applications in neuroimaging data analysis. Journal of the American Statistical Association, 108(502):540–552, 2013

    2013

  28. [36]

    Multi-source heterogeneous data fusion algorithm based on federated learning

    Jincheng Zhou and Yang Lei. Multi-source heterogeneous data fusion algorithm based on federated learning. In International Conference on Soft Computing in Data Science , pages 46–60. Springer, 2023. APPENDIX A RESULTS FOR COMPAVG We present the results of CompA VG, the naive e...

    2023

Pith tools

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