pith. sign in

arxiv: 2606.11480 · v1 · pith:66Y7YXSPnew · submitted 2026-06-09 · 💻 cs.LG

Accurate and Resource-Efficient Federated Continual Learning

Jebacyril Arockiaraj , Dhruv Parikh , Jayashree Adivarahan , Rajgopal Kannan , Viktor Prasanna This is my paper

Pith reviewed 2026-06-27 13:38 UTC · model grok-4.3

classification 💻 cs.LG
keywords federated continual learningrandom featuresGram matrixtruncated SVDridge regressionpseudo-labelingresource efficiencyanalytic learning
0
0 comments X

The pith

FedRAN replaces gradient updates in federated continual learning with truncated-SVD summaries of Gram matrices to cut communication by up to 121x while raising accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that analytic methods based on random features can outperform iterative optimization in federated continual learning when communication, computation, and labels are scarce. It does so by having clients send only low-rank approximations of their second-order statistics, letting the server merge subspaces across clients and tasks before solving a closed-form classifier. A sympathetic reader would care because this removes the need for repeated local training rounds and full supervision, making distributed learning feasible on bandwidth-limited devices that see data arrive sequentially. The approach also adds prototype-based pseudo-labeling to handle cases where only 20 percent of labels are available.

Core claim

FedRAN transmits a truncated-SVD summary of each client's Gram matrix instead of gradients or full statistics, reducing the dominant communication cost from quadratic to linear in the random feature dimension for fixed rank; the server then applies a two-level QR-SVD merge first across clients and then across tasks before computing a ridge-regression solution in closed form, yielding up to 4.8 percentage points higher average accuracy than the strongest baseline on CIFAR-100, ImageNet-R, and VTAB while using 30.6-121.8 times less per-client communication and running 190.3 times faster on average than gradient baselines.

What carries the argument

The truncated-SVD summary of the Gram matrix, which compresses the second-order random-feature statistic for linear communication cost, together with the server's two-level QR-SVD subspace merge that combines information spatially across clients and temporally across tasks before closed-form ridge regression.

If this is right

  • Average accuracy rises by as much as 4.8 points over the strongest prior FCL method across the tested image datasets.
  • Per-client communication drops by factors between 30.6 and 121.8 compared with optimization-based approaches.
  • Training runs 190.3 times faster on average than gradient-based baselines.
  • When only 20 percent of labels are present, prototype pseudo-labeling adds up to 6.61 accuracy points.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same truncated-summary approach could be tested on non-image modalities such as text or sensor streams to check whether the rank needed for good accuracy scales similarly.
  • Because the server never sees raw gradients or full models, the method may reduce certain privacy leakage vectors that iterative methods expose.
  • If the random-feature dimension must grow with task complexity, the linear communication saving might shrink, suggesting a need to adapt the truncation rank dynamically.

Load-bearing premise

The low-rank SVD truncation of each client's Gram matrix keeps enough information that the server's merged ridge classifier remains competitive with iterative optimization across sequential tasks.

What would settle it

Run FedRAN and a standard optimization-based FCL baseline on a new streaming dataset with the same communication budget; if FedRAN accuracy falls more than 2 points below the baseline while still reporting the claimed communication reduction, the central claim is falsified.

Figures

Figures reproduced from arXiv: 2606.11480 by Dhruv Parikh, Jayashree Adivarahan, Jebacyril Arockiaraj, Rajgopal Kannan, Viktor Prasanna.

Figure 1
Figure 1. Figure 1: Overview of federated continual learning. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Motivation for resource-constrained FCL. Left and middle: sequential non-IID client updates of trainable baselines [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed FedRAN system. Each client extracts frozen pretrained features, optionally assigns pseudo [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Task-wise accuracy comparison across CIFAR-100, ImageNet-R, and VTAB with [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of proxy-label under limited labeled data. FedRAN with proxy labels consistently improves average accuracy at [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of the SVD rank under different random projection dimensions [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation of FedRAN components on CIFAR-100 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
read the original abstract

Federated continual learning (FCL) must learn from distributed task streams under limited resources, such as communication, computation, memory, and label availability. Existing FCL methods often rely on repeated local optimization, replay, and full supervision. Analytic alternatives avoid iterative training and replay, but using high-dimensional random features to improve accuracy requires a second-order feature statistic, the Gram matrix, which has a quadratic communication cost in the random feature size $M$. We propose FedRAN, a resource-aware analytic FCL framework that replaces gradient-based updates with compact random feature statistics. Each client transmits a truncated-SVD summary of its Gram matrix, reducing the dominant second-order upload from quadratic to linear in $M$ for fixed rank. The server performs a two-level QR-SVD subspace merge, spatially across clients and temporally across tasks, and solves a ridge classifier in closed form. FedRAN further supports label scarcity through prototype-based pseudo-labeling. Across CIFAR-100, ImageNet-R, and VTAB datasets, FedRAN improves average accuracy by up to 4.8 percentage points over the strongest baseline, uses 30.6-121.8$\times$ less per-client communication than optimization-based FCL, and is 190.3$\times$ faster on average than gradient-based baselines; with only 20% labels, pseudo-labeling improves average accuracy by up to 6.61 points. These results show that FedRAN enables accurate and resource-efficient FCL under communication, computation, and label constraints. The source code is available at https://github.com/JebacyrilArockiaraj/Fed-RAN-SSL.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes FedRAN, an analytic federated continual learning method that uses random features with truncated-SVD summaries of per-client Gram matrices to reduce communication from quadratic to linear in feature dimension M. The server performs a two-level QR-SVD merge (spatial across clients, temporal across tasks) followed by closed-form ridge regression; prototype-based pseudo-labeling handles label scarcity. On CIFAR-100, ImageNet-R and VTAB it reports up to 4.8 pp higher average accuracy than the strongest baseline, 30.6–121.8× lower per-client communication than optimization-based FCL, and 190.3× faster training than gradient-based methods; with 20 % labels the pseudo-labeling variant gains up to 6.61 pp.

Significance. If the truncated-SVD approximation is shown to be sufficient, the work supplies a genuinely resource-efficient, non-iterative alternative to existing FCL methods, with concrete communication and wall-clock gains that are directly tied to the analytic formulation. The public code release and the explicit handling of label scarcity are additional strengths that would make the result immediately usable by the community.

major comments (2)
  1. [Method description (around the FedRAN algorithm and QR-SVD merge)] The central accuracy claim rests on the assumption that a fixed-rank truncated SVD of each client’s Gram matrix supplies a subspace whose subsequent two-level QR-SVD merge yields a ridge classifier competitive with iterative optimization. No approximation bounds on the discarded singular components are supplied, nor is it shown that the retained rank-r summary remains sufficient when later tasks introduce discriminative directions orthogonal to those retained from earlier tasks (see the skeptic note on the temporal merge).
  2. [Experiments section] The SVD truncation rank is listed as the sole free parameter, yet no ablation or sensitivity analysis on its value is reported; the communication–accuracy trade-off therefore cannot be assessed from the given experiments.
minor comments (2)
  1. [Abstract] The abstract states empirical gains but does not mention the truncation rank or any assumption on the retained subspace; adding a brief qualifier would improve clarity.
  2. [Tables and figures] Table captions and axis labels should explicitly state the random-feature dimension M and the chosen truncation rank so that the reported communication numbers can be reproduced.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on FedRAN. Below we respond point-by-point to the major comments, clarifying the empirical grounding of the truncated-SVD approach and committing to additional experiments where appropriate.

read point-by-point responses

  1. Referee: [Method description (around the FedRAN algorithm and QR-SVD merge)] The central accuracy claim rests on the assumption that a fixed-rank truncated SVD of each client’s Gram matrix supplies a subspace whose subsequent two-level QR-SVD merge yields a ridge classifier competitive with iterative optimization. No approximation bounds on the discarded singular components are supplied, nor is it shown that the retained rank-r summary remains sufficient when later tasks introduce discriminative directions orthogonal to those retained from earlier tasks (see the skeptic note on the temporal merge).

    Authors: We acknowledge the absence of theoretical approximation bounds. FedRAN is motivated by the practical observation that low-rank Gram-matrix summaries retain the dominant directions needed for accurate ridge regression in the random-feature space; this is supported by the consistent gains (up to 4.8 pp) across CIFAR-100, ImageNet-R and VTAB. The two-level QR-SVD merge is explicitly designed to accommodate new discriminative directions: the spatial merge aggregates client subspaces and the temporal merge recomputes an updated orthonormal basis from the concatenated summaries of all tasks seen so far, so directions orthogonal to earlier subspaces are incorporated when they appear in later Gram matrices. While formal bounds would be desirable, the analytic closed-form solution and the reported communication and wall-clock advantages are tied directly to the low-rank representation, and the empirical results across three benchmarks indicate that the retained rank is sufficient for the evaluated continual-learning regimes. revision: no

  2. Referee: [Experiments section] The SVD truncation rank is listed as the sole free parameter, yet no ablation or sensitivity analysis on its value is reported; the communication–accuracy trade-off therefore cannot be assessed from the given experiments.

    Authors: We agree that an explicit sensitivity study on the truncation rank would strengthen the presentation of the communication–accuracy trade-off. In the revised manuscript we will add an ablation that varies the retained rank r on the primary datasets, reporting both average accuracy and per-client communication volume for each setting. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation uses standard linear-algebra primitives

full rationale

The paper constructs FedRAN from random-feature Gram matrices, truncated SVD compression, two-level QR-SVD merging, and closed-form ridge regression. These operations are standard, externally defined linear-algebra steps whose outputs are not redefined in terms of the target accuracy metric. No equation reduces a claimed performance quantity to a fitted parameter by construction, no self-citation supplies a load-bearing uniqueness theorem, and no ansatz is smuggled via prior work. Empirical accuracy, communication, and speed numbers are reported from experiments rather than derived tautologically from the method equations themselves. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Review performed from abstract only; free parameters and axioms cannot be exhaustively enumerated without the methods section.

free parameters (1)
  • SVD truncation rank
    The rank chosen for the truncated SVD summary directly controls the communication-accuracy trade-off and must be selected per experiment.
axioms (1)
  • domain assumption Random Fourier features yield a sufficiently accurate finite-dimensional approximation to the kernel for the downstream ridge classifier.
    Invoked to replace iterative gradient training with closed-form analytic updates.

pith-pipeline@v0.9.1-grok · 5848 in / 1416 out tokens · 19878 ms · 2026-06-27T13:38:53.261759+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

74 extracted references · 5 linked inside Pith

  1. [1]

    Zahir Alsulaimawi. 2026. One-Shot Federated Ridge Regression: Exact Recovery via Sufficient Statistic Aggregation.arXiv preprint arXiv:2601.08216(2026)

    arXiv 2026

  2. [2]

    Gaurav Bagwe, Xiaoyong Yuan, Miao Pan, and Lan Zhang. 2023. Fed-cprompt: Contrastive prompt for rehearsal-free federated continual learning.arXiv preprint arXiv:2307.04869(2023)

    arXiv 2023

  3. [3]

    Yavuz Faruk Bakman, Duygu Nur Yaldiz, Yahya H Ezzeldin, and Salman Aves- timehr. 2023. Federated orthogonal training: Mitigating global catastrophic forgetting in continual federated learning.arXiv preprint arXiv:2309.01289(2023)

    arXiv 2023

  4. [4]

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. 2021. A continual learning survey: Defying forgetting in classification tasks.IEEE transactions on pattern analysis and machine intelligence44, 7 (2021), 3366–3385

    2021

  5. [5]

    Yongheng Deng, Sheng Yue, Tuowei Wang, Guanbo Wang, Ju Ren, and Yaoxue Zhang. 2023. Fedinc: An exemplar-free continual federated learning framework with small labeled data. InProceedings of the 21st ACM Conference on Embedded Networked Sensor Systems. 56–69. Accurate and Resource-Efficient Federated Continual Learning Conference’17, Washington, DC, USA, ,

    2023

  6. [6]

    Jiahua Dong, Hongliu Li, Yang Cong, Gan Sun, Yulun Zhang, and Luc Van Gool

  7. [7]

    No one left behind: Real-world federated class-incremental learning.IEEE Transactions on Pattern Analysis and Machine Intelligence46, 4 (2023), 2054–2070

    2023

  8. [8]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    Pith/arXiv arXiv 2020

  9. [9]

    Armin Eftekhari, Raphael A Hauser, and Andreas Grammenos. 2019. MOSES: A streaming algorithm for linear dimensionality reduction.IEEE transactions on pattern analysis and machine intelligence42, 11 (2019), 2901–2911

    2019

  10. [10]

    Jianqing Fan, Dong Wang, Kaizheng Wang, and Ziwei Zhu. 2019. Distributed estimation of principal eigenspaces.Annals of statistics47, 6 (2019), 3009

    2019

  11. [11]

    Kejia Fan, Jianheng Tang, Zhirui Yang, Feijiang Han, Jiaxu Li, Run He, Yajiang Huang, Anfeng Liu, Houbing Herbert Song, Yunhuai Liu, et al . 2025. Apfl: Analytic personalized federated learning via dual-stream least squares.arXiv preprint arXiv:2508.10732(2025)

    arXiv 2025

  12. [13]

    Eros Fanì, Raffaello Camoriano, Barbara Caputo, and Marco Ciccone. 2024. Ac- celerating heterogeneous federated learning with closed-form classifiers.arXiv preprint arXiv:2406.01116(2024)

    arXiv 2024

  13. [14]

    Mina Ghashami, Edo Liberty, Jeff M Phillips, and David P Woodruff. 2016. Fre- quent directions: Simple and deterministic matrix sketching.SIAM J. Comput.45, 5 (2016), 1762–1792

    2016

  14. [15]

    Masoume Gholizade, Fabrizio Ruffini, Pietro Ducange, and Francesco Marcelloni

  15. [16]

    Federated continual learning: A comprehensive survey on lifelong and privacy-preserving learning over distributed and non-stationary data.Neuro- computing(2026), 133929

    2026

  16. [17]

    Dipam Goswami, Simone Magistri, Kai Wang, Bartłomiej Twardowski, Andrew Bagdanov, and Joost van de Weijer. 2026. Covariances for free: Exploiting mean distributions for training-free federated learning.Advances in Neural Information Processing Systems38 (2026), 65081–65115

    2026

  17. [18]

    Andreas Grammenos, Rodrigo Mendoza Smith, Jon Crowcroft, and Cecilia Mas- colo. 2020. Federated principal component analysis.Advances in neural informa- tion processing systems33 (2020), 6453–6464

    2020

  18. [19]

    Zenghao Guan, Guojun Zhu, Yucan Zhou, Wu Liu, Weiping Wang, Jiebo Luo, and Xiaoyan Gu. 2026. Enhancing Federated Class-Incremental Learning via Spatial- Temporal Statistics Aggregation. InProceedings of the ACM Web Conference 2026. 5356–5367

    2026

  19. [20]

    Haiyang Guo, Fei Zhu, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. 2024. Pilora: Prototype guided incremental lora for federated class-incremental learning. InEuropean Conference on Computer Vision. Springer, 141–159

    2024

  20. [21]

    Nathan Halko, Per-Gunnar Martinsson, and Joel A Tropp. 2011. Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions.SIAM review53, 2 (2011), 217–288

    2011

  21. [22]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition. 770–778

    2016

  22. [23]

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. 2021. The many faces of robustness: A critical analysis of out-of-distribution generalization. InProceedings of the IEEE/CVF international conference on computer vision. 8340– 8349

    2021

  23. [24]

    Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. 2019. Measuring the effects of non-identical data distribution for federated visual classification.arXiv preprint arXiv:1909.06335(2019)

    Pith/arXiv arXiv 2019

  24. [25]

    Mark A Iwen and Benjamin W Ong. 2016. A distributed and incremental SVD algorithm for agglomerative data analysis on large networks.SIAM J. Matrix Anal. Appl.37, 4 (2016), 1699–1718

    2016

  25. [26]

    Wonyong Jeong, Jaehong Yoon, Eunho Yang, and Sung Ju Hwang. 2020. Federated semi-supervised learning with inter-client consistency & disjoint learning.arXiv preprint arXiv:2006.12097(2020)

    arXiv 2020

  26. [27]

    Ellango Jothimurugesan, Kevin Hsieh, Jianyu Wang, Gauri Joshi, and Phillip B Gibbons. 2023. Federated learning under distributed concept drift. InInternational Conference on Artificial Intelligence and Statistics. PMLR, 5834–5853

    2023

  27. [28]

    Peter Kairouz and H Brendan McMahan. 2021. Advances and open problems in federated learning.Foundations and trends in machine learning14, 1-2 (2021), 1–210

    2021

  28. [30]

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks.Proceedings of the national academy of sciences114, 13 (2017), 3521– 3526

    2017

  29. [31]

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

    2009

  30. [32]

    Gwen Legate, Nicolas Bernier, Lucas Page-Caccia, Edouard Oyallon, and Eugene Belilovsky. 2023. Guiding the last layer in federated learning with pre-trained models.Advances in Neural Information Processing Systems36 (2023), 69832– 69848

    2023

  31. [33]

    Haolin Li and Hoda Bidkhori. 2025. FedGTEA: Federated Class-Incremental Learning with Gaussian Task Embedding and Alignment.arXiv preprint arXiv:2510.12927(2025)

    arXiv 2025

  32. [34]

    Yichen Li, Yuying Wang, Jiahua Dong, Haozhao Wang, Yining Qi, Rui Zhang, and Ruixuan Li. 2026. Resource-constrained federated continual learning: What does matter?Advances in Neural Information Processing Systems38 (2026), 79493– 79518

    2026

  33. [35]

    Zhizhong Li and Derek Hoiem. 2017. Learning without forgetting.IEEE transac- tions on pattern analysis and machine intelligence40, 12 (2017), 2935–2947

    2017

  34. [36]

    Mark D McDonnell, Dong Gong, Amin Parvaneh, Ehsan Abbasnejad, and Anton Van den Hengel. 2023. Ranpac: Random projections and pre-trained models for continual learning.Advances in Neural Information Processing Systems36 (2023), 12022–12053

    2023

  35. [38]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-efficient learning of deep networks from decentralized data. InArtificial intelligence and statistics. Pmlr, 1273–1282

    2017

  36. [39]

    Chuiyang Meng, Ming Tang, and Vincent WS Wong. 2026. FLoRG: Federated Fine-tuning with Low-rank Gram Matrices and Procrustes Alignment.arXiv preprint arXiv:2602.17095(2026)

    arXiv 2026

  37. [40]

    Saleh Momeni, Changnan Xiao, and Bing Liu. 2026. Anacp: Toward upper- bound continual learning via analytic contrastive projection.Advances in Neural Information Processing Systems38 (2026), 72923–72944

    2026

  38. [41]

    Byoungjun Park, Pedro Porto Buarque de GusmÃG, o, Dongjin Ji, and Minhoe Kim

  39. [42]

    CATCHFed: Efficient Unlabeled Data Utilization for Semi-Supervised Fed- erated Learning in Limited Labels Environments.arXiv preprint arXiv:2511.11778 (2025)

    arXiv 2025

  40. [43]

    Liangzu Peng, Juan Elenter, Joshua Agterberg, Alejandro Ribeiro, and René Vidal

  41. [44]

    Loranpac: Low-rank random features and pre-trained models for bridging theory and practice in continual learning.arXiv preprint arXiv:2410.00645(2024)

    arXiv 2024

  42. [45]

    Ameya Prabhu, Shiven Sinha, Ponnurangam Kumaraguru, Philip H Torr, Ozan Sener, and Puneet K Dokania. 2024. RanDumb: Random representations outper- form online continually learned representations.Advances in Neural Information Processing Systems37 (2024), 37988–38006

    2024

  43. [46]

    Daiqing Qi, Handong Zhao, and Sheng Li. 2023. Better generative replay for continual federated learning.arXiv preprint arXiv:2302.13001(2023)

    arXiv 2023

  44. [47]

    Yijun Quan, Wentai Wu, and Giovanni Montana. 2026. Exact Federated Contin- ual Unlearning for Ridge Heads on Frozen Foundation Models.arXiv preprint arXiv:2603.12977(2026)

    arXiv 2026

  45. [48]

    Ali Rahimi and Benjamin Recht. 2007. Random features for large-scale kernel machines.Advances in neural information processing systems20 (2007)

    2007

  46. [49]

    Ali Rahimi and Benjamin Recht. 2008. Weighted sums of random kitchen sinks: Replacing minimization with randomization in learning.Advances in neural information processing systems21 (2008)

    2008

  47. [50]

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. 2017. icarl: Incremental classifier and representation learning. InPro- ceedings of the IEEE conference on Computer Vision and Pattern Recognition. 2001– 2010

    2017

  48. [51]

    Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska- Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. 2018. Progress & compress: A scalable framework for continual learning. InInternational conference on machine learning. PMLR, 4528–4537

    2018

  49. [52]

    Maksym Shamrai. 2026. Concatenated Matrix SVD: Compression Bounds, In- cremental Approximation, and Error-Constrained Clustering.arXiv preprint arXiv:2601.11626(2026)

    arXiv 2026

  50. [53]

    James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira

  51. [54]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11909–11919

  52. [55]

    Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. 2022. Fedproto: Federated prototype learning across hetero- geneous clients. InProceedings of the AAAI conference on artificial intelligence, Vol. 36. 8432–8440

    2022

  53. [56]

    Jianheng Tang, Yajiang Huang, Kejia Fan, Feijiang Han, Jiaxu Li, Jinfeng Xu, Run He, Anfeng Liu, Houbing Herbert Song, Huiping Zhuang, et al. 2026. DeepAFL: Deep analytic federated learning.arXiv preprint arXiv:2603.00579(2026). Conference’17, Washington, DC, USA, , Jebacyril Arockiaraj, Dhruv Parikh, Jayashree Adivarahan, Rajgopal Kannan, and Viktor Prasanna

    arXiv 2026

  54. [57]

    Jianheng Tang, Zhirui Yang, Jingchao Wang, Kejia Fan, Jinfeng Xu, Huiping Zhuang, Anfeng Liu, Houbing Herbert Song, Leye Wang, and Yunhuai Liu

  55. [58]

    FedHiP: Heterogeneity-Invariant Personalized Federated Learning Through Closed-Form Solutions.arXiv preprint arXiv:2508.04470(2025)

    arXiv 2025

  56. [59]

    Jianheng Tang, Huiping Zhuang, Jingyu He, Run He, Jingchao Wang, Kejia Fan, Anfeng Liu, Tian Wang, Leye Wang, Zhanxing Zhu, et al. 2025. Afcl: Analytic federated continual learning for spatio-temporal invariance of non-iid data.arXiv preprint arXiv:2505.12245(2025)

    arXiv 2025

  57. [60]

    Fabio Turazza, Marco Picone, and Marco Mamei. 2026. The Gaussian-Head OFL Family: One-Shot Federated Learning from Client Global Statistics.arXiv preprint arXiv:2602.01186(2026)

    Pith/arXiv arXiv 2026

  58. [61]

    Vinita Vasudevan and M Ramakrishna. 2017. A hierarchical singular value decomposition algorithm for low rank matrices.arXiv preprint arXiv:1710.02812 (2017)

    Pith/arXiv arXiv 2017

  59. [62]

    Yeshwanth Venkatesha, Youngeun Kim, Hyoungseob Park, Yuhang Li, and Priyadarshini Panda. 2022. Addressing client drift in federated continual learning with adaptive optimization.A vailable at SSRN 4188586(2022)

    2022

  60. [63]

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. 2022. Dualprompt: Complementary prompting for rehearsal-free continual learning. InEuropean conference on computer vision. Springer, 631–648

    2022

  61. [64]

    Kunlun Xu, Yibo Feng, Jiangmeng Li, Yongsheng Qi, and Jiahuan Zhou. 2026. C2 Prompt: Class-aware Client Knowledge Interaction for Federated Continual Learning.Advances in Neural Information Processing Systems38 (2026), 44109– 44139

    2026

  62. [65]

    Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang

  63. [66]

    InInter- national conference on machine learning

    Federated continual learning with weighted inter-client transfer. InInter- national conference on machine learning. PMLR, 12073–12086

  64. [67]

    Hao Yu, Xin Yang, Le Zhang, Hanlin Gu, Tianrui Li, Lixin Fan, and Qiang Yang

  65. [68]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Handling spatial-temporal data heterogeneity for federated continual learning via tail anchor. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4874–4883

  66. [69]

    Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. 2019. A large-scale study of representation learning with the visual task adaptation benchmark.arXiv preprint arXiv:1910.04867 (2019)

    Pith/arXiv arXiv 2019

  67. [71]

    Jie Zhang, Chen Chen, Weiming Zhuang, and Lingjuan Lyu. 2023. Target: Feder- ated class-continual learning via exemplar-free distillation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 4782–4793

    2023

  68. [72]

    Zhengyi Zhong, Weidong Bao, Ji Wang, Jianguo Chen, Lingjuan Lyu, and Wei Yang Bryan Lim. 2025. Sacfl: Self-adaptive federated continual learning for resource-constrained end devices.IEEE Transactions on Neural Networks and Learning Systems(2025)

    2025

  69. [73]

    Huiping Zhuang, Yizhu Chen, Di Fang, Run He, Kai Tong, Hongxin Wei, Ziqian Zeng, and Cen Chen. 2024. GACL: Exemplar-free generalized analytic continual learning.Advances in neural information processing systems37 (2024), 83024– 83047

    2024

  70. [74]

    Huiping Zhuang, Run He, Kai Tong, Ziqian Zeng, Cen Chen, and Zhiping Lin

  71. [75]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    DS-AL: A dual-stream analytic learning for exemplar-free class-incremental learning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 17237–17244

  72. [76]

    Huiping Zhuang, Yuchen Liu, Run He, Kai Tong, Ziqian Zeng, Cen Chen, Yi Wang, and Lap-Pui Chau. 2024. F-oal: Forward-only online analytic learning with fast training and low memory footprint in class incremental learning.Advances in Neural Information Processing Systems37 (2024), 41517–41538

    2024

  73. [77]

    Huiping Zhuang, Zhenyu Weng, Hongxin Wei, Renchunzi Xie, Kar-Ann Toh, and Zhiping Lin. 2022. ACIL: Analytic class-incremental learning with absolute memorization and privacy protection.Advances in Neural Information Processing Systems35 (2022), 11602–11614

    2022

  74. [78]

    Radim Řehřek. 2011. Subspace tracking for latent semantic analysis. InEuropean Conference on Information Retrieval. Springer, 289–300. Accurate and Resource-Efficient Federated Continual Learning Conference’17, Washington, DC, USA, , A Additional FedRAN Algorithms and Proofs This appendix provides helper algorithms and proofs for the state- ments in Secti...

    2011