Pith. sign in

REVIEW 5 major objections 4 minor 38 references

Knowledge-Aware Evolution for Task-Free Streaming Federated Continual Learning with Arbitrary Class Overlap

T0 review · 5 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read FedKACE claims that streaming federated continual learning can work without task identifiers and with overlapping classes, by combining adaptive inference switching, gradient-balanced replay, and kernel-spectral buffer selection, and that t

desk verdict A plausible empirical recipe for a genuinely new streaming-FCL setting, wrapped in a regret analysis that does not hold up as written. read the letter →

arxiv 2601.19788 v2 pith:AFMLC63O submitted 2026-01-27 cs.LG cs.DC

Sixing Tan , Xianmin Liu This is my paper · ORCID
classification cs.LGcs.DC MSC 68T05
keywords federatedcontinuallearningstreamingdatatask-freeclassoverlapreplaybuffergradient-balancedregretanalysismodelswitching
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

FedKACE is proposed to solve streaming federated continual learning, where clients see each data chunk only once, no task identifier is provided, and the same classes can reappear in later chunks. The paper's central claim is that three mechanisms—automatically deciding when to switch from a client's personal model to the aggregated global model for inference, weighting replay samples by the ratio of squared gradient norms, and maintaining a buffer with informative and decision-boundary-relevant samples—together prevent old knowledge from being confused with new knowledge. The paper reports that FedKACE achieves the highest average accuracy and lowest average regret across eight streaming scenarios on CIFAR-100 and ImageNet-100 compared with seven baselines. It also proves regret bounds showing that the global model strictly outperforms purely local training when all clients eventually see all classes. If correct, this gives a way to deploy continual learning in federated systems without task labels or task boundaries.

What carries the argument

The argument turns on three mechanisms plus a virtual-buffer identification. (1) Adaptive inference switching: clients monitor the change in the gap between the global model's average accuracy and its average predicted probability for ground-truth labels on the client's local buffer; two consecutive rounds of a decreasing gap triggers a permanent switch from local to global inference. (2) Adaptive gradient-balanced replay: each epoch sets the replay loss weight to the squared-L2-norm ratio of output-layer gradients for buffer samples versus new task samples, so the trade-off between retaining old knowledge and acquiring new knowledge is self-adjusting. (3) Kernel spectral boundary buffer mai

What would settle it

Train two models on identical streams: one is FedKACE's federated average, the other is trained centrally on the union of all clients' buffers at each round. If the centrally pooled model consistently outperforms the federated average by more than the claimed regret margin, the virtual-buffer equivalence underpinning Theorem 3 is contradicted. A smaller-scale version with a convex quadratic objective would also reveal whether parameter averaging reproduces pooled-data training.

Watch

Extended reading notes

Core claim

The discovery is that task-free streaming federated continual learning with arbitrary class overlap can be handled by a three-component system. First, each client starts with its local model for inference and switches permanently to the global model after two consecutive rounds in which the global model's accuracy-minus-confidence gap on the client's buffer shrinks, signaling that aggregation has absorbed the client's local distribution. Second, the replay loss weight is set adaptively each epoch to the ratio of squared L2 norms of output-layer gradients for buffered old samples versus new data, so the model tightens old-knowledge constraints when buffer gradients dominate and relaxes them w

Load-bearing premise

The main load-bearing premise is that averaging the clients' trained models behaves as if all clients' saved samples had been gathered into one pool and used for training; averaging and pooling are not the same thing, and the regret guarantee depends on that identification.

Editorial extensions

If this is right

  • If FedKACE is correct, federated continual learning systems can operate without task identifiers and with recurring classes, a condition that breaks existing FCL methods that rely on task boundaries and disjoint category sets.
  • Theorem 3 implies that global aggregation strictly outperforms purely local training after all categories have been seen, with the advantage growing as the number of clients increases.
  • Theorem 2's regret bound implies that kernel-spectral boundary buffer maintenance yields a strictly better constant factor than random sample selection for the same buffer capacity.
  • The experiments show that increasing buffer capacity improves accuracy more than increasing the number of clients, suggesting that storage investment may matter more than communication scale in these settings.
  • The ablation studies indicate that all three components—adaptive switching, adaptive replay weighting, and boundary-aware buffer maintenance—are necessary; removing any one degrades performance.

Reading between the lines

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

  • A natural testable extension the paper does not run: use the adaptive switching rule in a purely online continual learning setting without federated aggregation, since the accuracy-versus-confidence gap is a generic signal for when a model has adapted to a local distribution.
  • If the virtual-global-buffer equivalence holds, then communication frequency could be reduced while enlarging effective buffer size; the paper does not test sparse aggregation schedules, but the bound suggests a trade-off worth measuring.
  • The output-layer-only gradient ratio is a cheap plasticity-stability control that could be applied to any replay-based continual learner, not only federated systems, though its equivalence to full-gradient balancing is only argued through convergence approximations.
  • The regret bound contains O(t^{1−α}) with α > 0.5, which implies that if the stream's distribution shift decays slowly, the advantage of global aggregation may shrink over time; this is my reading of the bound, not a claim the paper makes.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper introduces a streaming federated continual learning (FCL) setting in which clients receive disjoint data batches across FL rounds, with possible category overlap and no task identifiers. It proposes FedKACE, consisting of three components: an adaptive mechanism for switching client inference from the local to the global model, a gradient-norm-ratio replay weighting scheme, and a kernel spectral boundary buffer maintenance strategy. The authors report experiments on CIFAR-100 and ImageNet-100 under four overlap levels, claiming the highest average accuracy and lowest average regret among baselines, and they provide three theorems: local saddle-point convergence of the replay weighting scheme (Theorem 1), a regret upper bound for the buffer maintenance (Theorem 2), and a regret upper bound and strict global-vs-local advantage for FedKACE (Theorem 3).

Significance. The proposed streaming FCL setting is well motivated and responds to a genuine gap: most FCL methods assume task identifiers or non-overlapping categories, while real streaming data often violate both. The three mechanisms are intuitively reasonable, and the experimental setup covers two datasets, four overlap configurations, ablations, and buffer/client-size variations. However, the central theoretical claims are not sound as written. Theorem 1's proof reverses the direction of Assumption 1; Theorem 2 essentially restates Assumption 9; and Theorem 3's strict advantage bound is contradicted by the proof's own additive O(T^{1-α}) term when 0.5<α<1. The virtual-global-buffer construction also equates parameter averaging with data pooling without justification. The experimental evidence is reported as single-run point estimates with no error bars or seeds, which is insufficient to support the claimed superiority, especially for comparisons within fractions of a percent. If the theoretical and statistical issues were repaired, the contribution could be significant, but as the manuscript stands, the regret analysis does not validate the method.

major comments (5)
  1. [Appendix C.2, Eq. (39)-(41)] The proof of Theorem 1 uses Assumption 1's upper bounds as lower bounds. Assumption 1 states ||∇L_task||_2 ≤ L1 and ||∇L_rep||_2 ≤ L2. The proof then claims 'under Assumption 1, ||∇L_task||^2 ≥ L1 and ... ≥ L2' and defines B_θ = L1 + ... + (λ)^2 L2 > 0. This is the reverse of the assumption; norms bounded above can be arbitrarily small, including zero. Consequently, the descent inequality (41) is not established, and the claimed saddle-point convergence, which Theorem 3 explicitly relies on, does not follow.
  2. [Appendix C.3.3-C.3.4, Eq. (86) and (90)] Theorem 2's regret bound is essentially Assumption 9 restated. Assumption 9 already postulates that the buffer is representative with error O(1/√M)+O(1/t^α). The covering-number and condition-number arguments establish improved sample spacing, but the step where the representativeness constant is reduced from C_random=1 to C_κ<1 is asserted rather than derived: the text says 'From Assumption 9 ... kernel spectrum boundary buffer maintenance reduces this constant factor.' Since the efficiency constant C_κ is load-bearing for the claimed superiority over random replay, Theorem 2 is circular with respect to its key assumption.
  3. [Appendix C.5.3, Eq. (137)] The proof derives E[Regret_global − Regret_local] ≤ −C_κ O(√(Cmax/M)(1−1/√K)) + O(T^{1−α}) and asserts that for sufficiently large T this is strictly negative because O(T^{1−α}) grows slower than the constant term. But Assumption 8 only requires α>0.5. For any 0.5<α<1, T^{1−α} → ∞, so the positive additive term is unbounded and cannot be dominated by a fixed constant. The upper bound diverges to +∞ and does not imply a strict negative difference. Thus Eq. (25), the paper's central theoretical claim that the global model strictly outperforms purely local training, is not proven.
  4. [Appendix C.5.2, Eq. (24)] The global regret bound treats the averaged model θ_g as if it were trained on a pooled virtual global buffer M^t_global, with per-category sample count |K^t_cmin|·M/|C^≤t_global|. Parameter averaging is not data pooling, and no lemma establishes that the global model's empirical risk on the union is controlled by that quantity. The construction assumes the benefit of aggregation it is meant to prove. This equivalence is load-bearing for Eq. (24) and for the strict C_κ improvement in Theorem 3, and it is not justified anywhere in the proof.
  5. [Section 5.1, Tables 1-4] All accuracy and regret results are reported as single-run point estimates. No seeds, standard errors, confidence intervals, or repeated trials are given. Several decisive comparisons are within fractions of a percent — e.g., CIFAR100 O=5: AS5 26.54 vs FedKACE 26.59, AS6 26.47 vs FedKACE 26.59. Without error bars, the claims of 'highest average accuracy' and the necessity of individual components are not statistically supported. This is a major weakness for the empirical half of the paper's central claim.
minor comments (4)
  1. [Section 4 heading] The section heading reads 'Methonology' instead of 'Methodology'.
  2. [Section 4.3, paragraph 1] 'capacity M is is nearly equally distributed' contains a duplicated 'is'.
  3. [Theorem 2 statement] The notation f_{M_t^k} is ambiguous: the model is said to be trained on M^{t-1}_k and D^t_k 'to construct M^t_k', but the regret is defined against the buffer M^t_k. Please clarify whether the bound is for the model before or after the buffer update, and align the notation in Eq. (21).
  4. [Figures 2-4] The accuracy-trend figures in Appendix B are not visibly rendered in the submitted text. If these are images, ensure they are embedded correctly; if they are textual placeholders, the figures are missing.

Circularity Check

2 steps flagged · score 6.0 of 10

Theorem 2's regret rate is largely Assumption 9 re-imported; Theorem 3 imports the cooperation advantage by treating parameter averaging as a pooled virtual global buffer.

  1. self definitional [Appendix C.1 (Assumption 9), Appendix C.3.3, Theorem 2 (Section 4.3)]
    "Assumption 9 (Local Buffer Representativeness): ... M_t^k is representative, such that for any model f, M_t^k yields the risk bound |R̂_{M_t^k}(f)−R_{p≤t_k}(f)| ≤ O(1/√M)+O(1/t^α) ... From Assumption 9, the representative error bound is O(1/√M)+O(1/t^α). Under identical assumptions, kernel spectral boundary buffer maintenance reduces this constant factor from C_random=1 for random sampling to C_κ<1."

    The regret bound Theorem 2 claims to prove is, in its leading M-rate and time-decay, already posited by Assumption 9. The proof does not derive the rate; it explicitly imports it ('From Assumption 9, the representative error bound is O(1/√M)+O(1/t^α)') and only changes the constant to C_κ. Thus the buffer-maintenance 'prediction' is the input assumption up to the constant factor and the |C| factor supplied by Assumption 6, making the claimed kernel-buffer improvement partially circular.

  2. other [Appendix C.5.2, Theorem 3 proof, Eq. (24)/(134) and Eq. (25)/(137)]
    "global model aggregation enables each client to indirectly benefit from buffer data of other clients. We therefore construct a virtual global buffer M^t_global = ∪_{k'} M^t_{k'} ... Accordingly, by treating the virtual global buffer as the buffer for the global model and invoking Theorem 2 together with Lemma 1, the regret upper bound for the global model satisfies ..."

    The strict cooperation advantage in Eq. (25) is imported by definitional fiat: the averaged global model is treated as if it were trained on the union of all clients' buffers, and the per-category count |K^t_cmin|·M/|C^t_global| is inserted into the local-buffer Theorem 2. Parameter averaging is never shown to equal data pooling, so the K-dependence of the asserted advantage is an assumption built into the virtual-buffer construction rather than a consequence of the FedKACE aggregation step.

full rationale

The empirical evaluation (Tables 1-4) is self-contained and compares FedKACE against external baselines, so the accuracy claim has independent support and by itself would score low. The theoretical regret analysis, however, is partially circular. Theorem 2's leading rate is not derived: Assumption 9 already postulates buffer risk error O(1/√M)+O(1/t^α), and the proof explicitly re-imports this as 'From Assumption 9...' before attaching the constant C_κ; the kernel buffer maintenance 'prediction' is therefore the assumption up to a constant factor. Theorem 3 then treats the federated average as a pooled virtual global buffer, inserting per-category coverage |K_cmin| into Theorem 2 without showing that parameter averaging equals data pooling, so the K-dependent cooperation advantage is imported by construction. Additionally, even granting the virtual-buffer step, Eq. (137) contains a positive +O(T^{1−α}) term that grows without bound for any 0.5<α<1, so the claimed strict negativity is vacuous; this is an internal inconsistency rather than a circularity, but it reinforces that the central theoretical comparison is not established. No load-bearing self-citation is present: [34]-[37] are external works. Score 6 reflects partial circularity of the central theoretical claim alongside independent empirical content.

Assumptions & free parameters 5 free parameters · 8 assumptions · 0 invented entities

The method adds hand-designed weighting and switching rules rather than fitted constants; the heavy theoretical lifting is done by Assumptions 1-13, of which Assumption 9 (buffer representativeness) already contains the regret rate the paper claims to prove. No new physical or conceptual entities are introduced.

free parameters (5)
  • Initial replay weight λ^{t,1}_k = 1
    Chosen by hand in Section 4.2 to treat new and old tasks as equally important; no justification from data or theory.
  • Switch criterion: two consecutive rounds with Δgap<0 = 2 consecutive rounds
    Arbitrary threshold in Eq. (6); no analysis of why two rounds is optimal.
  • Kernel bandwidth β = |M^{t-1}_k|^2 / d
    Hand-designed adaptive formula in Eq. (13), used for all kernel computations; not derived from data or a target condition.
  • IDV/CDV trade-off factors λ1, λ2 = log|M|/√|M|, 1/√|M|
    Chosen in Section 4.3 to prioritize diversity for small buffers; arbitrary functional forms, no sensitivity analysis.
  • Local training hyperparameters = J=20 epochs, batch size 32, AdamW lr=0.01 cosine, weight decay 0.001
    Manual choices in Appendix A.3 that affect all results; no ablation on these choices.
assumptions (8)
  • domain assumption L_total is L-smooth and μ-strongly convex (Assumptions 10-11)
    Used to invoke convergence results from [37] and to define the saddle-point problem; unrealistic for deep ResNet training.
  • ad hoc to paper Local buffer M_t^k is representative: |Rhat_M - R_p| ≤ O(1/√M)+O(1/t^α) (Assumption 9)
    This is essentially the regret statement Theorem 2 claims to prove; assuming it makes the theorem circular.
  • domain assumption Inter-round distribution shift decays as D_TV(p^t,p^{t-1}) ≤ Cmax/t^α with α>0.5 (Assumption 8)
    Controls every O(1/t^α) term in Theorems 2-3; never measured on the CIFAR100/ImageNet100 streams.
  • domain assumption Output-layer gradient norms proxy for full-model gradient norms (Eq. 11, citing [23])
    The replay-weight formula uses only output-layer gradients; no validation is given for this replay-weighting setting.
  • domain assumption Theorems from [34] apply to the proposed two-player game with β_t=1
    Theorem 1 proof imports equilibrium and stability lemmas from [34] directly; preconditions are not checked.
  • domain assumption Robbins-Monro learning-rate conditions (Assumption 3)
    The proof requires divergent infinite sums of learning rates, while experiments use a fixed J=20 cosine-decay schedule.
  • domain assumption Uniform per-category buffer allocation (Assumption 6)
    Used in Theorems 2 and 3 to get per-category sample count Θ(M/|C|); the algorithm only approximates this with quotas and remainders.
  • domain assumption Compactness, Lipschitz continuity, bounded gradient variance/norm (Assumptions 2,4,5,7,12,13)
    Standard but unverified technical conditions for ResNet-18 on image data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knowledge-Aware Evolution for Task-Free Streaming Federated Continual Learning with Arbitrary Class Overlap." pith.science (2026). https://pith.science/paper/AFMLC63O

@misc{pith2026260119788,
  author       = {Pith},
  title        = {Pith review of: Knowledge-Aware Evolution for Task-Free Streaming Federated Continual Learning with Arbitrary Class Overlap},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFMLC63O}},
  note         = {Machine review of arXiv:2601.19788}
}
read the original abstract

Federated Continual Learning (FCL) leverages inter-client collaboration to better balance new knowledge acquisition and old knowledge retention on non-stationary data. However, existing FCL methods struggle to adapt to streaming scenarios where sequential and ephemerally accessible data chunks lack task identifiers and exhibit arbitrary class overlap, leading to confusion between old and new knowledge and an inability to sustain local inference on all encountered classes. To address this, we propose FedKACE with three components: 1) an adaptive mechanism that determines when to switch the inference model from the local to the global one to improve client-side inference performance; 2) a responsive gradient-balanced replay scheme that utilizes the ratio of the squared L2 gradient norms to balance client-specific knowledge between new acquisition and old retention; 3) a holistic buffer maintenance strategy that preserves highly informative and boundary-significant samples to enhance knowledge retention under class overlap.Experiments across multiple scenarios and theoretical analysis demonstrate the effectiveness of FedKACE.

Figures

Figures reproduced from arXiv: 2601.19788 by the authors.

Figure 1
Figure 1. Illustration of the FedKACE framework. In FL round [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Trends in Method Accuracy and FedKACE Buffer Metrics Across FL Rounds(Tasks) on Cifar100. [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Trends in Method Accuracy and FedKACE Buffer Metrics Across FL Rounds(Tasks) on ImageNet100. [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Trends in Method Accuracy Across FL Rounds(Tasks). [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 1 linked inside Pith

  1. [1]

    Communication- efficient learning of deep networks from decentralized data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication- efficient learning of deep networks from decentralized data. In Aarti Singh and Xiaojin (Jerry) Zhu, editors, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20- 22 April 2017, Fort Lauderdale, FL, ...

    2017

  2. [2]

    Federated learning: Challenges, methods, and future directions.IEEE Signal Process

    Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions.IEEE Signal Process. Mag., 37(3):50–60, 2020

    2020

  3. [3]

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

    2024

  4. [4]

    A comprehensive survey of continual learning: Theory, method and application.IEEE Trans

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application.IEEE Trans. Pattern Anal. Mach. Intell., 46(8):5362–5383, 2024

    2024

  5. [5]

    Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista A

    Peter Kairouz, H. Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista A. Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, Rafael G. L. D’Oliveira, Hubert Eichner, Salim El Rouayheb, David Evans, Josh Gardner, Zachary Garrett, Adrià Gascón, Badih Ghazi, Phillip B. Gibbons, Marco Gruteser, Zaïd Harchaoui, ...

    2021

  6. [6]

    Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A

    James Kirkpatrick, Razvan Pascanu, Neil C. Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks.CoRR, 2016

    2016

  7. [7]

    Online continual learning in image classification: An empirical survey.Neurocomputing, 469:28–51, 2022

    Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo Kim, and Scott Sanner. Online continual learning in image classification: An empirical survey.Neurocomputing, 469:28–51, 2022

    2022

  8. [8]

    Towards efficient replay in federated incremental learning

    Yichen Li, Qunwei Li, Haozhao Wang, Ruixuan Li, Wenliang Zhong, and Guannan Zhang. Towards efficient replay in federated incremental learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 12820–12829. IEEE, 2024

    2024

Show all 38 references
  1. [9]

    Federated continual learning goes online: Uncertainty-aware memory management for vision tasks and beyond

    Giuseppe Serra and Florian Buettner. Federated continual learning goes online: Uncertainty-aware memory management for vision tasks and beyond. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025, 2025

    2025

  2. [10]

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

    Zhuang Qi, Ying-Peng Tang, Lei Meng, Han Yu, Xiaoxiao Li, and Xiangxu Meng. Class-wise balancing data replay for federated class-incremental learning. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 10

    2025

  3. [11]

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

    Minh-Tuan Tran, Trung Le, Xuan-May Le, Mehrtash Harandi, and Dinh Phung. Text-enhanced data-free approach for federated class-incremental learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 23870–2388...

    2024

  4. [12]

    CAN: leveraging clients as navigators for generative replay in federated continual learning

    Xuankun Rong, Jianshu Zhang, Kun He, and Mang Ye. CAN: leveraging clients as navigators for generative replay in federated continual learning. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025. OpenReview.net, 2025

    2025

  5. [13]

    Traceable federated continual learning

    Qiang Wang, Bingyan Liu, and Yawen Li. Traceable federated continual learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 12872–12881. IEEE, 2024

    2024

  6. [14]

    Closed- form merging of parameter-efficient modules for federated continual learning

    Riccardo Salami, Pietro Buzzega, Matteo Mosconi, Jacopo Bonato, Luigi Sabetta, and Simone Calderara. Closed- form merging of parameter-efficient modules for federated continual learning. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapor...

    2025

  7. [15]

    A distillation-based approach integrating continual learning and federated learning for pervasive services.CoRR, abs/2109.04197, 2021

    Anastasiia Usmanova, François Portet, Philippe Lalanda, and Germán Vega. A distillation-based approach integrating continual learning and federated learning for pervasive services.CoRR, abs/2109.04197, 2021

    2021 arXiv

  8. [16]

    Fedssi: Rehearsal-free continual federated learning with synergistic synaptic intelligence

    Yichen Li, Yuying Wang, Haozhao Wang, Yining Qi, Tianzhe Xiao, and Ruixuan Li. Fedssi: Rehearsal-free continual federated learning with synergistic synaptic intelligence. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19,...

    2025

  9. [17]

    Personalized federated class-incremental learning through critical parameter transfer

    Feng Wu, Siwei Feng, Yuanlu Chen, and Libang Zhao. Personalized federated class-incremental learning through critical parameter transfer. In2025 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2025, Hyderabad, India, April 6-11, 2025, pages 1–5...

    2025

  10. [18]

    Decentralized dynamic cooperation of personalized models for federated continual learning

    Danni Yang, Zhikang Chen, Sen Cui, Mengyue Yang, Ding Li, Abudukelimu Wuerkaixi, Haoxuan Li, Jinke Ren, and Mingming Gong. Decentralized dynamic cooperation of personalized models for federated continual learning. InThe Thirty-ninth Annual Conference on Neural Information Proc...

    2025

  11. [19]

    Coreset selection via reducible loss in continual learning

    Ruilin Tong, Yuhang Liu, Javen Qinfeng Shi, and Dong Gong. Coreset selection via reducible loss in continual learning. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025

    2025

  12. [20]

    Fei Ye and Adrian G. Bors. Online task-free continual learning via dynamic expansionable memory distribution. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025, pages 20512–20522. Computer Vision Foundation / IEEE, 2025

    2025

  13. [21]

    Rupam Mahmood

    Mohamed Elsayed and A. Rupam Mahmood. Addressing loss of plasticity and catastrophic forgetting in continual learning. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024

    2024

  14. [22]

    Mitigating catastrophic forgetting in online continual learning by modeling previous task interrelations via pareto optimization

    Yichen Wu, Hong Wang, Peilin Zhao, Yefeng Zheng, Ying Wei, and Long-Kai Huang. Mitigating catastrophic forgetting in online continual learning by modeling previous task interrelations via pareto optimization. In Forty-first International Conference on Machine Learning, ICML 20...

    2024

  15. [23]

    Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal

    Jordan T. Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. In8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020, 2020

    2020

  16. [24]

    Ferret: An efficient online continual learning framework under varying memory constraints

    Yuhao Zhou, Yuxin Tian, Jindi Lv, Mingjia Shi, Yuanxi Li, Qing Ye, Shuhao Zhang, and Jiancheng Lv. Ferret: An efficient online continual learning framework under varying memory constraints. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville,...

    2025

  17. [25]

    Clifton, and Bernard Ghanem

    Sihao Liu, Yibo Yang, Xiaojie Li, David A. Clifton, and Bernard Ghanem. Enhancing online continual learning with plug-and-play state space model and class-conditional mixture of discretization. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashvi...

    2025

  18. [26]

    Rethinking momentum knowledge distillation in online continual learning

    Nicolas Michel, Maorong Wang, Ling Xiao, and Toshihiko Yamasaki. Rethinking momentum knowledge distillation in online continual learning. InForty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024

    2024

  19. [27]

    Forgetting, ignorance or myopia: Revisiting key challenges in online continual learning

    Xinrui Wang, Chuanxing Geng, Wenhai Wan, Shao-Yuan Li, and Songcan Chen. Forgetting, ignorance or myopia: Revisiting key challenges in online continual learning. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang,...

    2024

  20. [28]

    Information-theoretic generalization analysis for expected calibration error

    Futoshi Futami and Masahiro Fujisawa. Information-theoretic generalization analysis for expected calibration error. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang, editors,Advances in Neural Information Proces...

    2024

  21. [29]

    Convergence of adam under relaxed assumptions

    Haochuan Li, Alexander Rakhlin, and Ali Jadbabaie. Convergence of adam under relaxed assumptions. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors,Advances in Neural Information Processing Systems 36: Annual Conference on Neur...

    2023

  22. [30]

    Krizhevsky and G

    A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images.Handbook of Systemic Autoimmune Diseases, 1(4), 2009

    2009

  23. [31]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Peter L. Bartlett, Fernando C. N. Pereira, Christopher J. C. Burges, Léon Bottou, and Kilian Q. Weinberger, editors,Advances in Neural Information Proce...

    2012

  24. [32]

    Resource- constrained federated continual learning: What does matter? InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    Yichen Li, Yuying Wang, Jiahua Dong, Haozhao Wang, Yining Qi, Rui Zhang, and Ruixuan Li. Resource- constrained federated continual learning: What does matter? InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  25. [33]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV , USA, June 27-30, 2016, pages 770–778. IEEE Computer Society, 2016

    2016

  26. [34]

    Formalizing the generalization-forgetting trade-off in continual learning

    Krishnan Raghavan and Prasanna Balaprakash. Formalizing the generalization-forgetting trade-off in continual learning. In Marc’Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, editors,Advances in Neural Information Processing Syst...

    2021

  27. [35]

    Generalization bounds for non-stationary mixing processes.Mach

    Vitaly Kuznetsov and Mehryar Mohri. Generalization bounds for non-stationary mixing processes.Mach. Learn., 106(1):93–117, 2017

    2017

  28. [36]

    Bartlett and Shahar Mendelson

    Peter L. Bartlett and Shahar Mendelson. Rademacher and gaussian complexities: Risk bounds and structural results.J. Mach. Learn. Res., 3:463–482, 2002

    2002

  29. [37]

    Bach and Eric Moulines

    Francis R. Bach and Eric Moulines. Non-asymptotic analysis of stochastic approximation algorithms for machine learning. In John Shawe-Taylor, Richard S. Zemel, Peter L. Bartlett, Fernando C. N. Pereira, and Kilian Q. Weinberger, editors,Advances in Neural Information Processin...

    2011

  30. [38]

    O 1 J +C κ ·O Cmax τ α p M·min(τ, C max) ! +O(1) # =O(1)(123) Forτ > t/2(recent influence), since Qt i=τ+1 γi ≤1and 1 τ α ≤ 2α tα , it follows that: tX τ=t/2+1

    The local training of client k in FL round t converges to client-specific optimalλ t,∗ k andθ t,∗ k satisfying the saddle point condition: Lt total(θt,∗ k |λt,j k )≤L t total(θt,∗ k |λt,∗ k )≤L t total(θt,j k |λt,∗ k )(32) for all θt,j k ∈Ω θ and λt,j k ≥0 . This saddle point ...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.