pith. machine review for the scientific record. sign in

arxiv: 2604.08617 · v1 · submitted 2026-04-09 · 💻 cs.LG · cs.AI· cs.CV

Recognition: unknown

From Selection to Scheduling: Federated Geometry-Aware Correction Makes Exemplar Replay Work Better under Continual Dynamic Heterogeneity

Zhuang Qi , Ying-Peng Tang , Lei Meng , Guoqing Chao , Lei Wu , Han Yu , Xiangxu Meng
Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:02 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CV
keywords federated continual learningexemplar replaygeometric structure alignmentenergy-based geometric correctionrepresentation collapseclass imbalanceequiangular tight frame
0
0 comments X

The pith

Federated geometry-aware correction prevents rare-class features from collapsing toward frequent classes in continual learning.

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

This paper introduces FEAT to make exemplar replay more effective in federated continual learning by countering the tendency of rare-class features to drift toward frequent classes due to client imbalances. It does so through two modules that enforce a stable geometric reference using fixed prototypes and strip away biased directional components in features. A sympathetic reader would care because real federated systems face shifting, non-uniform data across clients and tasks, where standard replay methods amplify forgetting and bias against minorities. The approach seeks to preserve class separability without exchanging raw samples, supporting longer-term model stability in dynamic environments.

Core claim

FEAT alleviates imbalance-induced representation collapse via the Geometric Structure Alignment module, which aligns pairwise angular similarities between feature representations and fixed shared Equiangular Tight Frame prototypes to promote geometric consistency across tasks and clients, and the Energy-based Geometric Correction module, which removes task-irrelevant directional components from embeddings to reduce majority-class prediction bias and improve minority-class sensitivity under class-imbalanced federated continual learning.

What carries the argument

The Geometric Structure Alignment module, which performs structural knowledge distillation by matching pairwise angular similarities of features to fixed shared Equiangular Tight Frame prototypes serving as a class-discriminative reference structure.

If this is right

  • Enhances sensitivity to minority classes under imbalanced client distributions in continual settings.
  • Reduces overall prediction bias toward majority classes during replay-based training.
  • Mitigates representation drift and catastrophic forgetting across dynamic heterogeneous clients and tasks.
  • Allows exemplar replay to maintain performance without additional data sharing.

Where Pith is reading between the lines

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

  • The fixed-prototype strategy might extend to non-federated continual learning with severe imbalance by providing a stable geometric anchor.
  • Combining this with adaptive prototype updates could handle even faster distribution shifts.
  • It points to geometry as a lightweight alternative to complex importance sampling for replay selection in distributed systems.

Load-bearing premise

Aligning features to fixed shared prototypes will produce geometric consistency that stops client-specific imbalances from dragging rare-class representations toward frequent ones.

What would settle it

A federated simulation with controlled task shifts and increasing client imbalance where removing the alignment module causes measurable increase in angular collapse of minority features and drop in their accuracy, while the full method maintains separation.

Figures

Figures reproduced from arXiv: 2604.08617 by Guoqing Chao, Han Yu, Lei Meng, Lei Wu, Xiangxu Meng, Ying-Peng Tang, Zhuang Qi.

Figure 1
Figure 1. Figure 1: (a) Continual dynamic heterogeneity: clients see dif [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the proposed FEAT framework. It contains two main modules: 1) the Geometry Structure Alignment module aligns local representations with global ETF prototypes to alleviate inter-client heterogeneity; and 2) the Energy-based Geometric Correction module removes biased components from the feature space during inference. lenges such as client heterogeneity and task-level imbalance remain [29], w… view at source ↗
Figure 3
Figure 3. Figure 3: Cross-client feature alignment in an incremental 5-task [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The GSA module reduces the tendency of tail classes to [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evaluation on CIFAR-10 (5 tasks), CIFAR-100 (10 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Sensitivity analysis of hyperparameters λ, ρ, and τ , adjusted over the ranges {0.05, 0.1, 0.5}, {0.5, 0.7, 0.9}, and {0.07, 0.5}, respectively. For each parameter, the default value is fixed while the other two are varied. It achieves robust perfor￾mance under a wide range of these parameters [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Impact of per-task memory capacity M on model per￾formance across different datasets, where FEATF shows the best overall performance. This section evaluates FEATF and FEATR against baselines under different replay budgets on three datasets (5-task split), with replay sizes set to {100, 150, 300} (CIFAR￾10), {125, 250, 500} (CIFAR-100), and {250, 500} (TinyImageNet-Subset). As shown in [PITH_FULL_IMAGE:fig… view at source ↗
Figure 10
Figure 10. Figure 10: Geometric correction effect of the EGC module on [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of visual attention. The GSA module cor [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
read the original abstract

Exemplar replay has become an effective strategy for mitigating catastrophic forgetting in federated continual learning (FCL) by retaining representative samples from past tasks. Existing studies focus on designing sample-importance estimation mechanisms to identify information-rich samples. However, they typically overlook strategies for effectively utilizing the selected exemplars, which limits their performance under continual dynamic heterogeneity across clients and tasks. To address this issue, this paper proposes a Federated gEometry-Aware correcTion method, termed FEAT, which alleviates imbalance-induced representation collapse that drags rare-class features toward frequent classes across clients. Specifically, it consists of two key modules: 1) the Geometric Structure Alignment module performs structural knowledge distillation by aligning the pairwise angular similarities between feature representations and their corresponding Equiangular Tight Frame prototypes, which are fixed and shared across clients to serve as a class-discriminative reference structure. This encourages geometric consistency across tasks and helps mitigate representation drift; 2) the Energy-based Geometric Correction module removes task-irrelevant directional components from feature embeddings, which reduces prediction bias toward majority classes. This improves sensitivity to minority classes and enhances the model's robustness under class-imbalanced distributions.

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 / 1 minor

Summary. The manuscript proposes FEAT, a method to improve exemplar replay in federated continual learning under continual dynamic heterogeneity. It introduces two modules: (1) Geometric Structure Alignment, which performs structural distillation by aligning pairwise angular similarities of learned features to fixed, shared Equiangular Tight Frame (ETF) prototypes across clients and tasks to encourage geometric consistency and reduce representation drift; (2) Energy-based Geometric Correction, which removes task-irrelevant directional components from embeddings to reduce prediction bias toward majority classes and improve sensitivity to minority classes.

Significance. If the empirical results and ablations confirm the claims, the work could meaningfully advance federated continual learning by addressing how selected exemplars are utilized rather than merely selected. Enforcing a shared geometric reference via ETF prototypes offers a concrete mechanism for mitigating client-specific and task-induced representation collapse, which is a recognized challenge in imbalanced, non-stationary federated settings. The approach is novel in its geometry-aware focus and could inspire further work on prototype-based regularization in distributed continual learning.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (method): The central claim that aligning pairwise angular similarities to fixed shared ETF prototypes mitigates imbalance-induced collapse and drift rests on the unproven assumption that the ETF geometry remains a valid class-discriminative reference after new classes arrive in later tasks and under client-specific drifts. No derivation, stability analysis, or counterexample test is provided showing that this alignment separates rare-class features from frequent-class directions rather than merely imposing a global regularizer; the skeptic concern therefore lands as a load-bearing correctness risk.
  2. [§3.2] §3.2 (Energy-based Geometric Correction): The module is described as removing task-irrelevant directional components to reduce majority-class bias, yet the manuscript supplies no explicit formulation, energy function definition, or proof that the correction selectively preserves minority-class directions without introducing new collapse modes under dynamic heterogeneity.
minor comments (1)
  1. [Abstract] The abstract would be improved by briefly stating the key quantitative gains (e.g., accuracy or forgetting metrics) and the number of clients/tasks in the primary experiments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful feedback. We appreciate the positive evaluation of the novelty and potential impact of FEAT for addressing representation issues in federated continual learning. We address the major comments point by point below, acknowledging areas where the current manuscript lacks sufficient rigor, and commit to revisions that strengthen the theoretical and formal aspects without misrepresenting the work.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (method): The central claim that aligning pairwise angular similarities to fixed shared ETF prototypes mitigates imbalance-induced collapse and drift rests on the unproven assumption that the ETF geometry remains a valid class-discriminative reference after new classes arrive in later tasks and under client-specific drifts. No derivation, stability analysis, or counterexample test is provided showing that this alignment separates rare-class features from frequent-class directions rather than merely imposing a global regularizer; the skeptic concern therefore lands as a load-bearing correctness risk.

    Authors: We thank the referee for identifying this critical assumption. The ETF prototypes are designed as fixed, shared equiangular references to enforce consistent geometric separation across clients and tasks, with new classes assigned dedicated prototype vectors upon arrival while preserving the overall tight-frame structure. This is intended to anchor rare-class features to their specific directions rather than allowing drift to majority-class vectors. However, the current manuscript does not include a formal derivation or stability analysis under continual class arrival and client drifts. In the revision, we will add a dedicated subsection deriving the alignment's effect on feature separation (showing that pairwise angular matching to ETF pulls embeddings toward orthogonal prototype directions), along with a brief stability argument and new ablation experiments testing prototype validity on counterexample sequences with extreme imbalance and drift. revision: yes

  2. Referee: [§3.2] §3.2 (Energy-based Geometric Correction): The module is described as removing task-irrelevant directional components to reduce majority-class bias, yet the manuscript supplies no explicit formulation, energy function definition, or proof that the correction selectively preserves minority-class directions without introducing new collapse modes under dynamic heterogeneity.

    Authors: We acknowledge that the description of the Energy-based Geometric Correction in §3.2 is high-level and lacks an explicit energy function or formal proof. The module aims to subtract directional components orthogonal to the aligned ETF prototypes to reduce bias. In the revised manuscript, we will provide the full formulation: the energy function E(f) = ||f - proj_P(f)||^2 where P denotes the subspace spanned by the class prototypes, with the correction applied as f' = f - α * (f - proj_P(f)) for a scaling factor α. We will include a proof sketch demonstrating selectivity for minority directions under the ETF alignment (leveraging equiangular margins to ensure minority prototypes retain influence) and additional experiments confirming no new collapse modes across dynamic heterogeneity settings. revision: yes

Circularity Check

0 steps flagged

No circularity: modules introduced as additive mechanisms without reducing claims to self-defined inputs or fits

full rationale

The paper proposes FEAT with two explicit modules—Geometric Structure Alignment (aligning features to fixed shared ETF prototypes) and Energy-based Geometric Correction (removing directional components)—to address representation collapse under federated continual learning. These are presented as independent design choices that encourage geometric consistency and reduce bias, with no equations, derivations, or self-citations shown that make the claimed mitigation equivalent to the inputs by construction. The ETF reference is adopted as an external class-discriminative structure rather than fitted or redefined within the method, and no 'prediction' reduces to a parameter estimated from the target data. The derivation chain remains self-contained as a proposed algorithmic addition, consistent with the absence of load-bearing self-referential steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim depends on the assumption that Equiangular Tight Frame prototypes provide a useful fixed reference structure for alignment. No free parameters or additional invented entities are explicitly quantified in the abstract.

axioms (1)
  • domain assumption Equiangular Tight Frame prototypes serve as a fixed, shared, class-discriminative reference structure for aligning pairwise angular similarities across clients and tasks.
    Directly invoked in the Geometric Structure Alignment module description.
invented entities (1)
  • Equiangular Tight Frame prototypes no independent evidence
    purpose: Provide a fixed geometric reference to enforce consistency and reduce representation drift.
    Used as the target for structural knowledge distillation; no independent falsifiable evidence supplied in abstract.

pith-pipeline@v0.9.0 · 5529 in / 1344 out tokens · 60042 ms · 2026-05-10T17:02:04.423833+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

80 extracted references · 11 canonical work pages · 2 internal anchors

  1. [1]

    Don’t memorize; mimic the past: Federated class incremental learning without episodic memory.arXiv preprint arXiv:2307.00497, 2023

    Sara Babakniya, Zalan Fabian, Chaoyang He, Mahdi Soltanolkotabi, and Salman Avestimehr. Don’t memorize; mimic the past: Federated class incremental learning without episodic memory.arXiv preprint arXiv:2307.00497, 2023. 2

    work page arXiv 2023

  2. [2]

    Ten years of generative adversarial nets (gans): a sur- vey of the state-of-the-art.Machine Learning: Science and Technology, 5(1):011001, 2024

    Tanujit Chakraborty, Ujjwal Reddy KS, Shraddha M Naik, et al. Ten years of generative adversarial nets (gans): a sur- vey of the state-of-the-art.Machine Learning: Science and Technology, 5(1):011001, 2024. 2

    2024

  3. [3]

    General federated class-incremental learning with lightweight generative re- play.IEEE Internet of Things Journal, 2024

    Yuanlu Chen, Alysa Ziying Tan, Siwei Feng, Han Yu, Tao Deng, Libang Zhao, and Feng Wu. General federated class-incremental learning with lightweight generative re- play.IEEE Internet of Things Journal, 2024. 2

    2024

  4. [4]

    Class-level structural relation modeling and smoothing for visual representation learning

    Zitan Chen et al. Class-level structural relation modeling and smoothing for visual representation learning. InProceedings of the 31st ACM International Conference on Multimedia, pages 2964–2972, 2023. 5

    2023

  5. [5]

    Federated class-incremental learning

    Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, et al. Federated class-incremental learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10164–10173, 2022. 2

    2022

  6. [6]

    Federated incremental semantic segmentation

    Jiahua Dong, Duzhen Zhang, Yang Cong, Wei Cong, Henghui Ding, and Dengxin Dai. Federated incremental semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3934–3943, 2023. 3

    2023

  7. [7]

    Learning from each other: Generalized federated incremen- tal semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    Jiahua Dong, Wenqi Liang, Yang Cong, Gan Sun, Lixu Wang, Henghui Ding, Yulun Zhang, and Luc Van Gool. Learning from each other: Generalized federated incremen- tal semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026. 3

    2026

  8. [8]

    Federated learning with bilateral curation for partially class-disjoint data.Advances in Neural Information Processing Systems, 36:32006–32019, 2023

    Ziqing Fan, Jiangchao Yao, Bo Han, Ya Zhang, Yanfeng Wang, et al. Federated learning with bilateral curation for partially class-disjoint data.Advances in Neural Information Processing Systems, 36:32006–32019, 2023. 3

    2023

  9. [9]

    Prism: Progressive robust learning for open-world continual category discovery

    Wei Feng, Sijin Zhou, Yiwen Jiang, and Zongyuan Ge. Prism: Progressive robust learning for open-world continual category discovery. InThe Fourteenth International Confer- ence on Learning Representations. 3

  10. [10]

    Neighbor-guided unbiased framework for generalized category discovery in medical image classification.IEEE Journal of Biomedical and Health Informatics, 2025

    Wei Feng, Sijin Zhou, Yiwen Jiang, et al. Neighbor-guided unbiased framework for generalized category discovery in medical image classification.IEEE Journal of Biomedical and Health Informatics, 2025. 3

    2025

  11. [11]

    Be- yond federated prototype learning: Learnable semantic an- chors with hyperspherical contrast for domain-skewed data

    Lele Fu, Sheng Huang, Yanyi Lai, Tianchi Liao, et al. Be- yond federated prototype learning: Learnable semantic an- chors with hyperspherical contrast for domain-skewed data. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 16648–16656, 2025. 1

    2025

  12. [12]

    Zijian Gao, Kele Xu, et al. Rethinking obscured sub- optimality in analytic learning for exemplar-free class- incremental learning.IEEE Transactions on Circuits and Systems for Video Technology, 36(10):1123–1136, 2025. 2

    2025

  13. [13]

    Dynamical variational autoencoders: A comprehensive review.arXiv preprint arXiv:2008.12595, 2020

    Laurent Girin, Simon Leglaive, Xiaoyu Bie, Julien Diard, Thomas Hueber, and Xavier Alameda-Pineda. Dynamical variational autoencoders: A comprehensive review.arXiv preprint arXiv:2008.12595, 2020. 2

    work page arXiv 2008

  14. [14]

    Generative adversarial nets.Advances in neural information processing systems, 27, 2014

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Advances in neural information processing systems, 27, 2014. 2

    2014

  15. [15]

    Fedmut: Generalized federated learning via stochastic mutation

    Ming Hu, Yue Cao, Anran Li, Zhiming Li, et al. Fedmut: Generalized federated learning via stochastic mutation. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 12528–12537, 2024. 1

    2024

  16. [16]

    Is aggregation the only choice? federated learning via layer- wise model recombination

    Ming Hu, Zhihao Yue, Xiaofei Xie, Cheng Chen, et al. Is aggregation the only choice? federated learning via layer- wise model recombination. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1096–1107, 2024. 1

    2024

  17. [17]

    Fedcross: Towards accurate feder- ated learning via multi-model cross-aggregation

    Ming Hu, Peiheng Zhou, Zhihao Yue, Zhiwei Ling, Yihao Huang, Anran Li, et al. Fedcross: Towards accurate feder- ated learning via multi-model cross-aggregation. InIEEE In- ternational Conference on Data Engineering (ICDE), pages 2137–2150. IEEE, 2024. 1

    2024

  18. [18]

    Soft- consensual federated learning for data heterogeneity via mul- tiple paths

    Sheng Huang, Lele Fu, Fanghua Ye, Tianchi Liao, et al. Soft- consensual federated learning for data heterogeneity via mul- tiple paths. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 5

    2025

  19. [19]

    Overcoming catastrophic forgetting in neu- ral networks.Proceedings of the national academy of sci- ences, 114(13):3521–3526, 2017

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska- Barwinska, et al. Overcoming catastrophic forgetting in neu- ral networks.Proceedings of the national academy of sci- ences, 114(13):3521–3526, 2017. 5

    2017

  20. [20]

    Exploiting multi-label correlation in label distribution learning

    Zhiqiang Kou, Jing Wang, Jiawei Tang, et al. Exploiting multi-label correlation in label distribution learning. InPro- ceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 4326–4334, 2024. 3

    2024

  21. [21]

    Label distribution learning with biased anno- tations assisted by multi-label learning

    Zhiqiang Kou, Si Qin, Hailin Wang, Jing Wang, Mingkun Xie, et al. Label distribution learning with biased anno- tations assisted by multi-label learning. InProceedings of the Thirty-Fourth International Joint Conference on Artifi- cial Intelligence, 2025

    2025

  22. [22]

    Rankmatch: A novel approach to semi-supervised label distribution learn- ing leveraging rank correlation between labels

    Zhiqiang Kou, Yucheng Xie, Hailin Wang, et al. Rankmatch: A novel approach to semi-supervised label distribution learn- ing leveraging rank correlation between labels. InProceed- ings of the 39th Conference on Neural Information Process- ing Systems (NeurIPS 2025), 2025. 3

    2025

  23. [23]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5

    2009

  24. [24]

    Tiny imagenet visual recognition challenge.CS 231N, 7(7):3, 2015

    Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge.CS 231N, 7(7):3, 2015. 5

    2015

  25. [25]

    Model- contrastive federated learning

    Qinbin Li, Bingsheng He, and Dawn Song. Model- contrastive federated learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 7

    2021

  26. [26]

    Vt- fsl: Bridging vision and text with llms for few-shot learning

    Wenhao Li, Qiangchang Wang, Xianjing Meng, et al. Vt- fsl: Bridging vision and text with llms for few-shot learning. arXiv preprint arXiv:2509.25033, 2025. 8

    work page arXiv 2025

  27. [27]

    Dvla-rl: Dual-level vision-language alignment with reinforcement learning gating for few-shot learning

    Wenhao Li, Xianjing Meng, Qiangchang Wang, Zhongyi Han, et al. Dvla-rl: Dual-level vision-language alignment with reinforcement learning gating for few-shot learning. arXiv preprint arXiv:2602.00795, 2026. 3

    work page arXiv 2026

  28. [28]

    Cross-modal learning using privileged information for long-tailed image classification

    Xiangxian Li, Yuze Zheng, et al. Cross-modal learning using privileged information for long-tailed image classification. Computational Visual Media, 10(5):981–992, 2024. 3

    2024

  29. [29]

    Towards efficient re- play in federated incremental learning

    Yichen Li, Qunwei Li, Haozhao Wang, Ruixuan Li, Wen- liang Zhong, and Guannan Zhang. Towards efficient re- play in federated incremental learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12820–12829, 2024. 2, 3, 5

    2024

  30. [30]

    Re-fed+: A better replay strategy for federated incre- mental learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Yichen Li, Haozhao Wang, Yining Qi, Wei Liu, and Ruixuan Li. Re-fed+: A better replay strategy for federated incre- mental learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2, 5

    2025

  31. [31]

    Unleashing the power of continual learning on non-centralized devices: A survey.IEEE Communica- tions Surveys & Tutorials, 2025

    Yichen Li, Haozhao Wang, Wenchao Xu, Tianzhe Xiao, Hong Liu, et al. Unleashing the power of continual learning on non-centralized devices: A survey.IEEE Communica- tions Surveys & Tutorials, 2025. 1

    2025

  32. [32]

    Rehearsal-free continual fed- erated learning with synergistic synaptic intelligence.Inter- national Conference on Machine Learning, 2025

    Yichen Li, Yuying Wang, Haozhao Wang, Yining Qi, Tianzhe Xiao, and Ruixuan Li. Rehearsal-free continual fed- erated learning with synergistic synaptic intelligence.Inter- national Conference on Machine Learning, 2025. 5

    2025

  33. [33]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelli- gence, 40(12):2935–2947, 2017. 5

    2017

  34. [34]

    No fear of classifier biases: Neural collapse inspired federated learning with synthetic and fixed classifier

    Zexi Li, Xinyi Shang, Rui He, Tao Lin, and Chao Wu. No fear of classifier biases: Neural collapse inspired federated learning with synthetic and fixed classifier. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 5319–5329, 2023. 3

    2023

  35. [35]

    Diffusion-driven data replay: A novel approach to combat forgetting in fed- erated class continual learning

    Jinglin Liang, Jin Zhong, Hanlin Gu, et al. Diffusion-driven data replay: A novel approach to combat forgetting in fed- erated class continual learning. InEuropean Conference on Computer Vision, pages 303–319. Springer, 2024. 2

    2024

  36. [36]

    Federated domain generalization with decision insight matrix

    Tianchi Liao, Binghui Xie, Lele Fu, Sheng Huang, Bowen Deng, Chuan Chen, and Zibin Zheng. Federated domain generalization with decision insight matrix. InProceedings of the Thirty-Fourth International Joint Conference on Arti- ficial Intelligence, pages 5689–5697, 2025. 5

    2025

  37. [37]

    Fedbcgd: Communication-efficient accelerated block coordinate gradient descent for federated learning

    Junkang Liu, Fanhua Shang, Yuanyuan Liu, Hongying Liu, et al. Fedbcgd: Communication-efficient accelerated block coordinate gradient descent for federated learning. InPro- ceedings of the 32nd ACM International Conference on Mul- timedia, pages 2955–2963, 2024. 1

    2024

  38. [38]

    Im- proving generalization in federated learning with highly het- erogeneous data via momentum-based stochastic controlled weight averaging

    Junkang Liu, Yuanyuan Liu, Fanhua Shang, et al. Im- proving generalization in federated learning with highly het- erogeneous data via momentum-based stochastic controlled weight averaging. InForty-second International Conference on Machine Learning, 2025. 1

    2025

  39. [39]

    FedAdamW: A Communication-Efficient Optimizer with Convergence and Generalization Guarantees for Federated Large Models

    Junkang Liu, Fanhua Shang, Kewen Zhu, et al. Fedadamw: A communication-efficient optimizer with convergence and generalization guarantees for federated large models.arXiv preprint arXiv:2510.27486, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  40. [40]

    Cross-training with prototypical distilla- tion for improving the generalization of federated learning

    Tianhan Liu et al. Cross-training with prototypical distilla- tion for improving the generalization of federated learning. In2023 IEEE International Conference on Multimedia and Expo (ICME), pages 648–653. IEEE, 2023. 5

    2023

  41. [41]

    Neural collapse under cross-entropy loss.Applied and Computational Harmonic Analysis, 59:224–241, 2022

    Jianfeng Lu and Stefan Steinerberger. Neural collapse under cross-entropy loss.Applied and Computational Harmonic Analysis, 59:224–241, 2022. 3

    2022

  42. [42]

    Geo- metric knowledge-guided localized global distribution align- ment for federated learning

    Yanbiao Ma, Wei Dai, Wenke Huang, and Jiayi Chen. Geo- metric knowledge-guided localized global distribution align- ment for federated learning. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 20958– 20968, 2025. 1

    2025

  43. [43]

    Improving global generalization and local personalization for federated learning.IEEE Transactions on Neural Networks and Learn- ing Systems, 36(1):76–87, 2024

    Lei Meng, Zhuang Qi, Lei Wu, Xiaoyu Du, et al. Improving global generalization and local personalization for federated learning.IEEE Transactions on Neural Networks and Learn- ing Systems, 36(1):76–87, 2024. 1

    2024

  44. [44]

    Causal inference over visual-semantic-aligned graph for im- age classification

    Lei Meng, Xiangxian Li, Xiaoshuo Yan, Haokai Ma, et al. Causal inference over visual-semantic-aligned graph for im- age classification. InProceedings of the AAAI Conference on Artificial Intelligence, pages 19449–19457, 2025. 8

    2025

  45. [45]

    Generative adversarial networks (gans) in networking: A comprehensive survey & evaluation.Computer Networks, 194:108149, 2021

    Hojjat Navidan, Parisa Fard Moshiri, Mohammad Nabati, Reza Shahbazian, Seyed Ali Ghorashi, Vahid Shah- Mansouri, and David Windridge. Generative adversarial networks (gans) in networking: A comprehensive survey & evaluation.Computer Networks, 194:108149, 2021. 2

    2021

  46. [46]

    Overcoming catas- trophic forgetting in federated class-incremental learn- ing via federated global twin generator.arXiv preprint arXiv:2407.11078, 2024

    Thinh Nguyen, Khoa D Doan, et al. Overcoming catas- trophic forgetting in federated class-incremental learn- ing via federated global twin generator.arXiv preprint arXiv:2407.11078, 2024. 2

    work page arXiv 2024

  47. [47]

    Milad Khademi Nori et al. Federated class-incremental learning: A hybrid approach using latent exemplars and data- free techniques to address local and global forgetting.arXiv preprint arXiv:2501.15356, 2025. 2

    work page arXiv 2025

  48. [48]

    Prevalence of neural collapse during the terminal phase of deep learning training.Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020

    Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training.Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020. 3

    2020

  49. [49]

    Vari- ational autoencoder

    Lucas Pinheiro Cinelli, Matheus Ara ´ujo Marins, Ed- uardo Ant´unio Barros da Silva, and S´ergio Lima Netto. Vari- ational autoencoder. InVariational methods for machine learning with applications to deep networks, pages 111–149. Springer, 2021. 2

    2021

  50. [50]

    Better genera- tive replay for continual federated learning.arXiv preprint arXiv:2302.13001, 2023

    Daiqing Qi, Handong Zhao, and Sheng Li. Better genera- tive replay for continual federated learning.arXiv preprint arXiv:2302.13001, 2023. 2

    work page arXiv 2023

  51. [51]

    Federated learning for science: A survey on the path to a trustworthy collaboration ecosystem.Authorea Preprints, 2025

    Xin Qi, Meixuan Li, Sijin Zhou, et al. Federated learning for science: A survey on the path to a trustworthy collaboration ecosystem.Authorea Preprints, 2025. 2

    2025

  52. [52]

    Federated learning in oncology: bridg- ing artificial intelligence innovation and privacy protection

    Xin Qi, Tao Xu, et al. Federated learning in oncology: bridg- ing artificial intelligence innovation and privacy protection. Information Fusion, page 104154, 2026. 1

    2026

  53. [53]

    Cross-silo prototypical calibra- tion for federated learning with non-iid data

    Zhuang Qi, Lei Meng, et al. Cross-silo prototypical calibra- tion for federated learning with non-iid data. InProceedings of the 31st ACM International Conference on Multimedia, pages 3099–3107, 2023. 1

    2023

  54. [54]

    Cross-silo feature space align- ment for federated learning on clients with imbalanced data

    Zhuang Qi, Lei Meng, et al. Cross-silo feature space align- ment for federated learning on clients with imbalanced data. InThe 39th Annual AAAI Conference on Artificial Intelli- gence (AAAI-25), pages 19986–19994, 2025. 7

    2025

  55. [55]

    Class-wise balancing data replay for federated class-incremental learning

    Zhuang Qi, Ying-Peng Tang, et al. Class-wise balancing data replay for federated class-incremental learning. InThe Thirty-Ninth Annual Conference on Neural Information Pro- cessing Systems, 2025. 2, 3, 5

    2025

  56. [56]

    Global prompt re- finement with non-interfering attention masking for one-shot federated learning

    Zhuang Qi, Pan Yu, Lei Meng, et al. Global prompt re- finement with non-interfering attention masking for one-shot federated learning. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 5

    2025

  57. [57]

    Can: Leveraging clients as navigators for generative replay in fed- erated continual learning

    Xuankun Rong, Jianshu Zhang, Kun He, and Mang Ye. Can: Leveraging clients as navigators for generative replay in fed- erated continual learning. ICML, 2025. 2

    2025

  58. [58]

    Learning equi-angular representations for online continual learning

    Minhyuk Seo, Hyunseo Koh, Wonje Jeung, et al. Learning equi-angular representations for online continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 23933–23942, 2024. 3

    2024

  59. [59]

    Relaxed contrastive learning for federated learning

    Seonguk Seo, Jinkyu Kim, Geeho Kim, and Bohyung Han. Relaxed contrastive learning for federated learning. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12279–12288, 2024. 7

    2024

  60. [60]

    Asynchronous federated continual learning

    Donald Shenaj, Marco Toldo, Alberto Rigon, and Pietro Zanuttigh. Asynchronous federated continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 5055–5063, 2023. 1

    2023

  61. [61]

    Clip-guided federated learning on heterogeneity and long-tailed data

    Jiangming Shi, Shanshan Zheng, et al. Clip-guided federated learning on heterogeneity and long-tailed data. InProceed- ings of the AAAI Conference on Artificial Intelligence, pages 14955–14963, 2024. 7

    2024

  62. [62]

    Protoconnet: Prototypical aug- mentation and alignment for open-set few-shot image classi- fication.Displays, page 103364, 2026

    Kexuan Shi, Zhuang Qi, et al. Protoconnet: Prototypical aug- mentation and alignment for open-set few-shot image classi- fication.Displays, page 103364, 2026. 8

    2026

  63. [63]

    Exemplar- condensed federated class-incremental learning.arXiv preprint arXiv:2412.18926, 2024

    Rui Sun, Yumin Zhang, Varun Ojha, Tejal Shah, Hao- ran Duan, Bo Wei, and Rajiv Ranjan. Exemplar- condensed federated class-incremental learning.arXiv preprint arXiv:2412.18926, 2024. 2

    work page arXiv 2024

  64. [64]

    Text-enhanced data-free approach for federated class-incremental learning

    Minh-Tuan Tran, Trung Le, Xuan-May Le, Mehrtash Ha- randi, and Dinh Phung. Text-enhanced data-free approach for federated class-incremental learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23870–23880, 2024. 2, 5

    2024

  65. [65]

    Fedcda: Feder- ated learning with cross-rounds divergence-aware aggrega- tion

    Haozhao Wang, Haoran Xu, Yichen Li, et al. Fedcda: Feder- ated learning with cross-rounds divergence-aware aggrega- tion. InThe Twelfth International Conference on Learning Representations, 2023. 1

    2023

  66. [66]

    Data-free federated class incremental learn- ing with diffusion-based generative memory.arXiv preprint arXiv:2405.17457, 2024

    Naibo Wang, Yuchen Deng, Wenjie Feng, Jianwei Yin, and See-Kiong Ng. Data-free federated class incremental learn- ing with diffusion-based generative memory.arXiv preprint arXiv:2405.17457, 2024. 2

    work page arXiv 2024

  67. [67]

    Enhancing Visual Representation with Textual Semantics: Textual Semantics-Powered Prototypes for Heterogeneous Federated Learning

    Xinghao Wu, Jianwei Niu, Xuefeng Liu, Guogang Zhu, Ji- ayuan Zhang, and Shaojie Tang. Enhancing visual repre- sentation with textual semantics: Textual semantics-powered prototypes for heterogeneous federated learning.arXiv preprint arXiv:2503.13543, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  68. [68]

    Empow- ering vision transformers with multi-scale causal interven- tion for long-tailed image classification

    Xiaoshuo Yan, Zhaochuan Li, Lei Meng, et al. Empow- ering vision transformers with multi-scale causal interven- tion for long-tailed image classification. InProceedings of the Thirty-Fourth International Joint Conference on Artifi- cial Intelligence, pages 6785–6793, 2025. 8

    2025

  69. [69]

    Federated continual learning via knowledge fu- sion: A survey.IEEE Transactions on Knowledge and Data Engineering, 36(8):3832–3850, 2024

    Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. Federated continual learning via knowledge fu- sion: A survey.IEEE Transactions on Knowledge and Data Engineering, 36(8):3832–3850, 2024. 2

    2024

  70. [70]

    Fedgh: Hetero- geneous federated learning with generalized global header

    Liping Yi, Gang Wang, Xiaoguang Liu, et al. Fedgh: Hetero- geneous federated learning with generalized global header. InProceedings of the 31st ACM international conference on multimedia, pages 8686–8696, 2023. 1

    2023

  71. [71]

    Federated model heterogeneous matryoshka represen- tation learning.Advances in Neural Information Processing Systems, 37:66431–66454, 2024

    Liping Yi, Han Yu, Chao Ren, Gang Wang, Xiaoxiao Li, et al. Federated model heterogeneous matryoshka represen- tation learning.Advances in Neural Information Processing Systems, 37:66431–66454, 2024. 1

    2024

  72. [72]

    Personalized federated continual learning via multi-granularity prompt

    Hao Yu, Xin Yang, Xin Gao, Yan Kang, Hao Wang, Junbo Zhang, and Tianrui Li. Personalized federated continual learning via multi-granularity prompt. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4023–4034, 2024. 1

    2024

  73. [73]

    Overcoming spatial-temporal catas- trophic forgetting for federated class-incremental learning

    Hao Yu, Xin Yang, et al. Overcoming spatial-temporal catas- trophic forgetting for federated class-incremental learning. InProceedings of the 32nd ACM International Conference on Multimedia, pages 5280–5288, 2024. 1, 2

    2024

  74. [74]

    Handling spatial-temporal data het- erogeneity for federated continual learning via tail anchor

    Hao Yu, Xin Yang, Le Zhang, Hanlin Gu, Tianrui Li, Lixin Fan, and Qiang Yang. Handling spatial-temporal data het- erogeneity for federated continual learning via tail anchor. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4874–4883, 2025. 1

    2025

  75. [75]

    Fedagc: Federated continual learn- ing with asymmetric gradient correction

    Chengchao Zhang, Fanhua Shang, Hongying Liu, Liang Wan, and Wei Feng. Fedagc: Federated continual learn- ing with asymmetric gradient correction. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 3841–3850, 2025. 1

    2025

  76. [76]

    Target: Federated class- continual learning via exemplar-free distillation

    Jie Zhang, Chen Chen, et al. Target: Federated class- continual learning via exemplar-free distillation. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 4782–4793, 2023. 2, 5

    2023

  77. [77]

    Few-shot class-incremental learning for classifi- cation and object detection: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Jinghua Zhang, Li Liu, Olli Silv ´en, Matti Pietik ¨ainen, and Dewen Hu. Few-shot class-incremental learning for classifi- cation and object detection: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2

    2025

  78. [78]

    Causality inspired feder- ated learning for ood generalization

    Jiayuan Zhang, Xuefeng Liu, Jianwei Niu, Shaojie Tang, Haotian Yang, and Xinghao Wu. Causality inspired feder- ated learning for ood generalization. InForty-second Inter- national Conference on Machine Learning, 2025. 8

    2025

  79. [79]

    Decou- pled spatio-temporal consistency learning for self-supervised tracking

    Yaozong Zheng, Bineng Zhong, Qihua Liang, et al. Decou- pled spatio-temporal consistency learning for self-supervised tracking. InProceedings of the AAAI Conference on Artifi- cial Intelligence, pages 10635–10643, 2025. 3

    2025

  80. [80]

    Towards universal modal tracking with online dense temporal token learning

    Yaozong Zheng, Bineng Zhong, et al. Towards universal modal tracking with online dense temporal token learning. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2025. 8

    2025