REVIEW 3 major objections 4 minor 57 references
Federated Split Learning with Improved Communication and Storage Efficiency
T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read CSE-FSL cuts upstream traffic by a factor $h$ and stores a single server-side model, while keeping the $O(1/\sqrt{T})$ convergence rate of standard federated learning.
desk verdict Useful communication/storage-efficient FSL variant with an overclaimed server-side convergence guarantee; the drift-decay assumption is missing from the abstract. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Three components carry the design. First, the auxiliary network: a small task-head appended to the client-side model whose local loss replaces the server's backpropagated gradient, so clients can take $h$ local steps, or run asynchronously, without a round trip. Second, the single shared server-side model, updated serially whenever a client's smashed data arrives (an event-triggered update), which eliminates the $n$ server-side replicas that prior FSL methods require. Third, the batch frequency $h$: clients upload smashed data only when $m \bmod h = 0$, cutting the upstream payload per epoch from $n q |D|$ to $n q |D| / h$. The convergence argument rests on the usual $L$-smoothness and bounded-gradient assumptions and, for the server-side bound, on Assumption 3, which requires the cumulative $\ell^1$ distance $d^t_{c,i}$ between the current client-side output distribution and the optimal one to stay finite; the drift term in Proposition 2 is exactly the part of the bound that Assumption 3 controls. The proofs are standard descent-style expansions of the $L$-smooth inequality around the aggregated client update and the serial server update, with the drift term absorbed through Assumption 3.
What would settle it
Train CSE-FSL on a non-convex task and compute the empirical $\ell^1$ distance $d^t_{c,i}$ between the distribution of smashed data produced by the current client-side model and that produced by the fully converged client-side model, summed over clients and rounds. If this sum grows without bound while the client-side gradient bound in Proposition 1 still holds, the server-side rate in Proposition 2 is false and the average server-side gradient norm will be observed to plateau above zero; if the sum decays, the drift term vanishes and the stated rate should be observed.
Extended reading notes
Core claim
The central claim is that a federated split learning system can keep the convergence rate of standard federated learning while making server storage independent of the number of clients and cutting upstream traffic by a factor $h$. In CSE-FSL, each client owns only the bottom layers of the model plus a small auxiliary network; it trains locally for $h$ mini-batches, computing its local loss at the auxiliary head, and only then uploads the smashed data, the intermediate activations at the cut layer, together with its updated client-side parameters. The server keeps exactly one upper model, updating it sequentially as smashed data arrives, asynchronously and in arbitrary client order. Under the standard $L$-smoothness and bounded-gradient assumptions, Proposition 1 bounds the average squared client-side gradient norm by $O(1/\sqrt{T})$; Proposition 2 gives the same order for the server-side model, with an additional term proportional to the accumulated $\ell^1$ distance between the current and optimal smashed-data distributions, which the authors show vanishes when that drift decays over rounds. On CIFAR-10 and F-EMNIST, CSE-FSL reaches higher top-1 accuracy per byte of communication than the existing FSL baselines, with the largest gains in data-heavy settings where the $h$-periodic uploads matter most.
Load-bearing premise
The load-bearing premise is that the intermediate-layer outputs of the clients' trained models drift toward the outputs of the optimal client-side model fast enough that the total drift over all clients and rounds stays finite, and the paper assumes this rather than deriving it from the client-side convergence result.
Editorial extensions
If this is right
- Server storage becomes $|w|+|a|$ regardless of the number of clients, so a server can serve very large client populations without proportional memory growth.
- Upstream communication per global epoch falls from $n q |D|$ to $n q |D| / h$, and on CIFAR-10 the $h=5$ setting reaches 76.5% top-1 accuracy on 18.1 GB of traffic versus 77.8% at 86.8 GB for the auxiliary-network baseline.
- Clients no longer wait for server gradients, so client and server operate at their own pace; the paper's ordered-versus-random server update experiments show nearly identical accuracy, supporting the asynchronous event-triggered design.
- The client-side model converges at the same $O(1/\sqrt{T})$ order as standard federated averaging, so replacing the server's gradient with a local auxiliary-network loss does not asymptotically slow learning.
- In data-heavy regimes with large local datasets or a large split layer relative to the client model, larger $h$ yields both faster convergence per communication round and lower total traffic, the regime where the experiments show the clearest gains.
Reading between the lines
- The client-side bound's $h$-dependence saturates: its coefficient $2h/(2h-1)$ tends to 1, so the theory predicts that increasing $h$ does not asymptotically worsen the per-round gradient guarantee even as communication falls by $h$; the paper does not state this saturation explicitly.
- The proof treats client arrivals in a fixed order although the protocol is asynchronous; bounding the bias introduced by stale smashed data (up to $h$ batches old) under shuffled arrivals would put the asynchronous claim on the same footing as the convergence analysis, and the randomization experiments suggest this bias is small.
- Assumption 3 is measurable in practice: an operator could track the $\ell^1$ distance between successive smashed-data batches during training, and if the cumulative drift fails to decay, reduce $h$ or down-weight server-side updates to keep the server model in its convergent regime.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes CSE-FSL, a federated split learning method that adds an auxiliary network at the client to enable local client-side updates, keeps a single server-side model instead of one per client, and uploads smashed data only every h local batches. It claims an O(1/sqrt(T)) convergence rate under non-convex losses for both the client-side model (Proposition 1) and the server-side model (Proposition 2), and reports communication and storage gains in Table II. Experiments on CIFAR-10 and F-EMNIST compare CSE-FSL with SplitFed variants and auxiliary-network baselines in IID/non-IID and full/partial participation settings, including accuracy, communication load, and storage comparisons.
Significance. The algorithmic design is well motivated and the experimental study is substantial: it covers multiple datasets, IID and non-IID settings, communication-load comparisons, and auxiliary-network architecture choices. If the convergence claims held as stated, the paper would make a useful contribution to communication-efficient federated split learning. However, the theoretical centerpiece is not established as stated. Proposition 2's bound contains a drift term involving d^t_{c,i}, and Assumption 3 as written does not force that term to vanish; the paper itself acknowledges the residual gap only after the theorem. The client-side proof also appears to ignore the simultaneous update of the auxiliary network. These issues matter because the abstract and conclusions advertise a rigorous convergence guarantee, which is currently stronger than what the proofs support.
major comments (3)
- [Section V-A, Eq. (17), Assumption 3] Assumption 3 as stated (sum_i d^t_{c,i} < infinity) is automatically satisfied for finite n because d^t_{c,i} is an L1 distance between two probability distributions and is therefore bounded by 2; it does not imply that the Cesaro mean (1/T) sum_{t=1}^T sum_{i=1}^n d^t_{c,i} tends to zero. Consequently the residual term 4 G_2^2 / ((2n-1)T) sum_{t=1}^T sum_{i=1}^n d^t_{c,i} in Eq. (17) is not shown to vanish, and the claimed O(1/sqrt(T)) server-side convergence rate does not follow from Proposition 2. The abstract's statement that CSE-FSL is 'rigorously guaranteeing its convergence' is therefore unsupported; the qualification in Section V-A that d^t_{c,i} must decay should be elevated to the abstract and introduction.
- [Section V-A, discussion after Proposition 2] The paper acknowledges that one needs 'another assumption that d^t_{c,i} decays with t', but this assumption is neither stated in the theorem nor derived from Proposition 1. Proposition 1 only bounds the averaged squared gradient norm of the aggregate client loss F_c; this at best implies convergence to a stationary point of F_c, not convergence of x^t_c to the global minimizer x*_c used to define P*_{c,i}, and it provides no information about the sequence of smashed-data output distributions P^t_{c,i}. Thus there is no argument linking the client-side convergence result to the server-side drift term, and the residual gap in Eq. (17) is load-bearing for the advertised convergence guarantee.
- [Appendix A, Eqs. (18)-(31)] Proposition 1 is stated for the client-side objective in Eq. (4), which is minimized jointly over x_c and the auxiliary network a_c, and Eq. (8) updates both variables. The proof, however, treats F_c as a function of x_c alone: the L-smoothness inequality (18) is applied to F_c(x^{t+1}_c) - F_c(x^t_c) even though a^{t+1}_c differs from a^t_c, and the auxiliary-network updates never enter the descent calculation. As written, the proof establishes a partial-gradient bound under a fixed auxiliary network, not convergence of the joint variable (x_c, a_c). The statement should either bound the joint gradient norm or explicitly explain why the auxiliary-network updates can be ignored in the descent argument.
minor comments (4)
- [Eq. (17) and Appendix B, Eq. (48)] The first term on the right-hand side contains F_c(x*_s); this should presumably read F_s(x*_s).
- [Appendix B, Eq. (42)] The displayed derivation of the (B4) bound has a typesetting artifact ('}') and the notation switches between g_{x*_c}(z) and P*_{c,i}(z); please check the identity and the role of G_2 in the second inequality.
- [Section VI-A and Algorithm 1] The text says 'global aggregation is performed once per epoch (C = 1)', but Algorithm 1 defines c as the number of mini-batches between aggregations, so C = 1 would mean aggregation after every batch; the parameter meaning should be clarified.
- [Figure 5 caption] The caption says 'partial clients (5 clients) participation', but the text says clients are randomly chosen in each round; please specify the number of sampled clients per round.
Circularity Check
No circular derivation: the convergence proofs are self-contained, though the abstract overstates the server-side guarantee by omitting the drift-term condition.
full rationale
I walked the claimed derivation chain. Proposition 1 (Eq. 16) is a standard smooth non-convex SGD bound; its right-hand side depends only on L, h, G_1, the initial suboptimality, and T, with no hidden reuse of the claimed conclusion. Proposition 2 (Eq. 17) is explicit that the bound contains the drift term (4 G_2^2 / ((2n-1)T)) sum_t sum_i d^t_{c,i}; the paper then states convergence of the server-side model only "if we make another assumption that d^t_{c,i} decays with t," and otherwise acknowledges "a gap away from the stationary point." Therefore the server-side convergence claim is conditional on a distribution-drift property, not circularly forced: nothing in Proposition 1 or in the definition of d^t_{c,i} makes the drift term vanish by construction. Assumption 3 (sum_i d^t_{c,i} < infinity) is indeed weak, since d^t_{c,i} is an L1 distance between probability distributions and hence automatically finite for finite n; however, the paper does not use Assumption 3 to eliminate the residual term, only to bound the mismatch in Eq. (42). The conditional decay assumption is an additional unproved condition, so the abstract's phrase "rigorously guaranteeing its convergence" is an overclaim relative to the theorem, but this is a correctness/conditional-gap concern rather than a circularity. Self-citations such as [1] (the preliminary ICC version) and [51]-[52] in the standard-assumption citation block are not load-bearing. Hence no circular step can be exhibited from the paper's own equations.
Assumptions & free parameters
free parameters (2)
- h (batch upload frequency) =
h in {1,2,4,5,10,25,50} in experiments
- c (aggregation frequency) =
1 epoch in experiments, C=1 in theory
assumptions (5)
- domain assumption Assumption 1: client-side and server-side loss functions are L-smooth.
- domain assumption Assumption 2: expected squared norm of stochastic gradients is uniformly bounded by G1^2 and G2^2.
- ad hoc to paper Assumption 3: sum_i d^t_{c,i} is finite, where d^t_{c,i} is the L1 distance between current and optimal client-side output distributions.
- domain assumption Full client participation and equal local dataset sizes (Section III-A).
- domain assumption IID local datasets for convergence analysis (Section V-A).
Cite this review
Pith. "Pith review of Federated Split Learning with Improved Communication and Storage Efficiency." pith.science (2026). https://pith.science/paper/DY4RRBW2
@misc{pith2026250715816,
author = {Pith},
title = {Pith review of: Federated Split Learning with Improved Communication and Storage Efficiency},
year = {2026},
howpublished = {\url{https://pith.science/paper/DY4RRBW2}},
note = {Machine review of arXiv:2507.15816}
}
read the original abstract
Federated learning (FL) is one of the popular distributed machine learning (ML) solutions but incurs significant communication and computation costs at edge devices. Federated split learning (FSL) can train sub-models in parallel and reduce the computational burden of edge devices by splitting the model architecture. However, it still requires a high communication overhead due to transmitting the smashed data and gradients between clients and the server in every global round. Furthermore, the server must maintain separate partial models for every client, leading to a significant storage requirement. To address these challenges, this paper proposes a novel communication and storage efficient federated split learning method, termed CSE-FSL, which utilizes an auxiliary network to locally update the weights of the clients while keeping a single model at the server, hence avoiding frequent transmissions of gradients from the server and greatly reducing the storage requirement of the server. Additionally, a new model update method of transmitting the smashed data in selected epochs can reduce the amount of smashed data sent from the clients. We provide a theoretical analysis of CSE-FSL, rigorously guaranteeing its convergence under non-convex loss functions. The extensive experimental results further indicate that CSE-FSL achieves a significant communication reduction over existing FSL solutions using real-world FL tasks.
Figures
Figures from the paper (4 more)
Reference graph
Works this paper leans on
-
[9]
Accelerating federated learning with split learning on locally generated losses,
D.-J. Han, H. I. Bhatti, J. Lee, and J. Moon, “Accelerating federated learning with split learning on locally generated losses,” in ICML 2021 Workshop on Federated Learning for User Privacy and Data Confidentiality. ICML Board, 2021
work page 2021
-
[1]
Communication and storage efficient federated split learning,
Y . Mu and C. Shen, “Communication and storage efficient federated split learning,” in IEEE International Conference on Communications (ICC), May 2023, pp. 1–6
work page 2023
-
[2]
Communication-efficient learning of deep networks from decentralized data,
B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial intelligence and statistics . PMLR, 2017, pp. 1273– 1282
2017
-
[3]
Distributed learning of deep neural network over multiple agents,
O. Gupta and R. Raskar, “Distributed learning of deep neural network over multiple agents,” Journal of Network and Computer Applications , vol. 116, pp. 1–8, 2018
work page 2018
-
[5]
Reducing leakage in distributed deep learning for sensitive health data,
P. Vepakomma, O. Gupta, A. Dubey, and R. Raskar, “Reducing leakage in distributed deep learning for sensitive health data,” arXiv preprint arXiv:1812.00564, vol. 2, 2019
arXiv 2019
-
[6]
Splitfed: When federated learning meets split learning,
C. Thapa, M. A. P. Chamikara, S. Camtepe, and L. Sun, “Splitfed: When federated learning meets split learning,” arXiv preprint arXiv:2004.12088, 2020
arXiv 2004
-
[7]
Fedsl: Federated split learning on distributed sequential data in recurrent neural networks,
A. Abedi and S. S. Khan, “Fedsl: Federated split learning on distributed sequential data in recurrent neural networks,” Multimedia Tools and Applications, pp. 1–21, 2023
work page 2023
-
[8]
Group knowledge transfer: Federated learning of large cnns at the edge,
C. He, M. Annavaram, and S. Avestimehr, “Group knowledge transfer: Federated learning of large cnns at the edge,” Advances in Neural Information Processing Systems , vol. 33, pp. 14 068–14 080, 2020
work page 2020
Show all 57 references
-
[10]
Federated optimization: Distributed machine learning for on-device intelligence,
J. Kone ˇcn`y, H. B. McMahan, D. Ramage, and P. Richt ´arik, “Federated optimization: Distributed machine learning for on-device intelligence,” arXiv preprint arXiv:1610.02527 , 2016
2016 arXiv
-
[11]
Federated learning: Challenges, methods, and future directions,
T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated learning: Challenges, methods, and future directions,” IEEE Signal Process. Mag., vol. 37, no. 3, pp. 50–60, 2020
2020
-
[12]
Advances and open problems in federated learning,
P. Kairouz et al., “Advances and open problems in federated learning,” arXiv preprint arXiv:1912.04977 , 2019
1912 arXiv
-
[13]
Deep gradient compression: Reducing the communication bandwidth for distributed training,
Y . Lin, S. Han, H. Mao, Y . Wang, and W. J. Dally, “Deep gradient compression: Reducing the communication bandwidth for distributed training,” arXiv preprint arXiv:1712.01887 , 2017
2017 arXiv
-
[14]
Acceleration for com- pressed gradient descent in distributed and federated optimization,
Z. Li, D. Kovalev, X. Qian, and P. Richt ´arik, “Acceleration for com- pressed gradient descent in distributed and federated optimization,” arXiv preprint arXiv:2002.11364 , 2020
2002 arXiv
-
[15]
Atomo: Communication-efficient learning via atomic spar- sification,
H. Wang, S. Sievert, S. Liu, Z. Charles, D. Papailiopoulos, and S. Wright, “Atomo: Communication-efficient learning via atomic spar- sification,” Advances in neural information processing systems , vol. 31, 2018
2018
-
[16]
Error feedback fixes SignSGD and other gradient compression schemes,
S. P. Karimireddy, Q. Rebjock, S. U. Stich, and M. Jaggi, “Error feedback fixes SignSGD and other gradient compression schemes,” in Proceedings of the 36th International Conference on Machine Learning , June 2019, pp. 3252–3261
2019
-
[17]
Federated learning with heterogeneous quantization bit allocation and aggregation for Internet of Things,
S. Chen, L. Li, G. Wang, M. Pang, and C. Shen, “Federated learning with heterogeneous quantization bit allocation and aggregation for Internet of Things,” IEEE Internet Things J., vol. 11, no. 2, pp. 3132–3143, January 2023
2023
-
[18]
Design and analysis of uplink and downlink communications for federated learning,
S. Zheng, C. Shen, and X. Chen, “Design and analysis of uplink and downlink communications for federated learning,” IEEE J. Select. Areas Commun., vol. 39, no. 7, pp. 2150–2167, July 2021
2021
-
[19]
Dynamic aggregation for heterogeneous quantization in federated learning,
S. Chen, C. Shen, L. Zhang, and Y . Tang, “Dynamic aggregation for heterogeneous quantization in federated learning,” IEEE Trans. Wireless Commun., vol. 20, no. 10, pp. 6804–6819, October 2021
2021
-
[20]
Joint optimal quantization and aggregation of federated learning scheme in V ANETs,
Y . Li, Y . Guo, M. Alazab, S. Chen, C. Shen, and K. Yu, “Joint optimal quantization and aggregation of federated learning scheme in V ANETs,” IEEE Transactions on Intelligent Transportation Systems , pp. 1–12, 2022
2022
-
[21]
Resource rationing for wireless federated learning: Concept, benefits, and challenges,
C. Shen, J. Xu, S. Zheng, and X. Chen, “Resource rationing for wireless federated learning: Concept, benefits, and challenges,” IEEE Commun. Mag., vol. 59, no. 5, pp. 82–87, May 2021
2021
-
[22]
Random orthogonalization for federated learning in massive MIMO systems,
X. Wei, C. Shen, J. Yang, and H. V . Poor, “Random orthogonalization for federated learning in massive MIMO systems,” IEEE Trans. Wireless Commun., vol. 23, no. 3, pp. 2469–2485, March 2024
2024
-
[23]
Efficient model compression for hierarchical federated learning,
X. Zhu, S. Yu, J. Wang, and Q. Yang, “Efficient model compression for hierarchical federated learning,” arXiv preprint arXiv:2405.17522, 2024
2024 arXiv
-
[24]
Fedsl: A communication-efficient federated learning with split layer aggregation,
W. Zhang, T. Zhou, Q. Lu, Y . Yuan, A. Tolba, and W. Said, “Fedsl: A communication-efficient federated learning with split layer aggregation,” IEEE Internet of Things Journal, vol. 11, no. 9, pp. 15 587–15 601, 2024
2024
-
[25]
Can we use split learning on 1d cnn models for privacy preserving training?
S. Abuadbba, K. Kim, M. Kim, C. Thapa, S. A. Camtepe, Y . Gao, H. Kim, and S. Nepal, “Can we use split learning on 1d cnn models for privacy preserving training?” in Proceedings of the 15th ACM Asia Conference on Computer and Communications Security , 2020, pp. 305– 318
2020
-
[26]
Split learning for collaborative deep learning in health- care,
M. G. Poirot, P. Vepakomma, K. Chang, J. Kalpathy-Cramer, R. Gupta, and R. Raskar, “Split learning for collaborative deep learning in health- care,” arXiv preprint arXiv:1912.12115 , 2019
1912 arXiv
-
[27]
Multiple classification with split learning,
J. Kim, S. Shin, Y . Yu, J. Lee, and K. Lee, “Multiple classification with split learning,” in The 9th International Conference on Smart Media and Applications, 2020, pp. 358–363
2020
-
[28]
Unleashing the tiger: Inference attacks on split learning,
D. Pasquini, G. Ateniese, and M. Bernaschi, “Unleashing the tiger: Inference attacks on split learning,” in Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security , 2021, pp. 2113–2129
2021
-
[29]
Neural compression and filtering for edge-assisted real-time object detection in challenged networks,
Y . Matsubara and M. Levorato, “Neural compression and filtering for edge-assisted real-time object detection in challenged networks,” in 2020 25th International Conference on Pattern Recognition (ICPR) . IEEE, 2021, pp. 2272–2279. 12
2020
-
[30]
Spliteasy: A practical approach for training ml models on mobile devices,
K. Palanisamy, V . Khimani, M. H. Moti, and D. Chatzopoulos, “Spliteasy: A practical approach for training ml models on mobile devices,” in Proceedings of the 22nd International Workshop on Mobile Computing Systems and Applications , 2021, pp. 37–43
2021
-
[31]
End-to-end evaluation of feder- ated learning and split learning for internet of things,
Y . Gao, M. Kim, S. Abuadbba, Y . Kim, C. Thapa, K. Kim, S. A. Camtepe, H. Kim, and S. Nepal, “End-to-end evaluation of feder- ated learning and split learning for internet of things,” arXiv preprint arXiv:2003.13376, 2020
2003 arXiv
-
[32]
Detailed com- parison of communication efficiency of split learning and federated learning,
A. Singh, P. Vepakomma, O. Gupta, and R. Raskar, “Detailed com- parison of communication efficiency of split learning and federated learning,” arXiv preprint arXiv:1909.09145 , 2019
1909 arXiv
-
[33]
Learning the optimal partition for collaborative dnn training with privacy requirements,
L. Zhang and J. Xu, “Learning the optimal partition for collaborative dnn training with privacy requirements,” IEEE Internet of Things Journal , vol. 9, no. 13, pp. 11 168–11 178, 2021
2021
-
[34]
Hivemind: Towards cellular native machine learning model splitting,
S. Wang, X. Zhang, H. Uchiyama, and H. Matsuda, “Hivemind: Towards cellular native machine learning model splitting,” IEEE Journal on Selected Areas in Communications , vol. 40, no. 2, pp. 626–640, 2021
2021
-
[35]
Federated split vision transformer for covid-19 cxr diagnosis using task-agnostic training,
S. Park, G. Kim, J. Kim, B. Kim, and J. C. Ye, “Federated split vision transformer for covid-19 cxr diagnosis using task-agnostic training,” arXiv preprint arXiv:2111.01338 , 2021
2021 arXiv
-
[36]
Dfl: Dynamic federated split learning in heterogeneous iot,
E. Samikwa, A. Di Maio, and T. Braun, “Dfl: Dynamic federated split learning in heterogeneous iot,” IEEE transactions on machine learning in communications and networking , 2024
2024
-
[37]
Privacy and efficiency of communications in federated split learning,
Z. Zhang, A. Pinto, V . Turina, F. Esposito, and I. Matta, “Privacy and efficiency of communications in federated split learning,” IEEE Transactions on Big Data , 2023
2023
-
[38]
Exploring the privacy-energy consumption tradeoff for split federated learning,
J. Lee, M. Seif, J. Cho, and H. V . Poor, “Exploring the privacy-energy consumption tradeoff for split federated learning,” IEEE Network, 2024
2024
-
[39]
Optimizing privacy and latency tradeoffs in split federated learn- ing over wireless networks,
——, “Optimizing privacy and latency tradeoffs in split federated learn- ing over wireless networks,” IEEE Wireless Communications Letters , 2024
2024
-
[40]
Ppsfl: Privacy-preserving split federated learning for heterogeneous data in edge-based internet of things,
J. Zheng, Y . Chen, and Q. Lai, “Ppsfl: Privacy-preserving split federated learning for heterogeneous data in edge-based internet of things,” Future Generation Computer Systems , vol. 156, pp. 231–241, 2024
2024
-
[41]
Efficient parallel split learning over resource-constrained wireless edge networks,
Z. Lin, G. Zhu, Y . Deng, X. Chen, Y . Gao, K. Huang, and Y . Fang, “Efficient parallel split learning over resource-constrained wireless edge networks,” IEEE Transactions on Mobile Computing , vol. 23, no. 10, pp. 9224–9239, 2024
2024
-
[42]
Esfl: Ef- ficient split federated learning over resource-constrained heterogeneous wireless devices,
G. Zhu, Y . Deng, X. Chen, H. Zhang, Y . Fang, and T. F. Wong, “Esfl: Ef- ficient split federated learning over resource-constrained heterogeneous wireless devices,” IEEE Internet of Things Journal , 2024
2024
-
[43]
Communication-efficient split learning via adaptive feature-wise compression,
Y . Oh, J. Lee, C. G. Brinton, and Y .-S. Jeon, “Communication-efficient split learning via adaptive feature-wise compression,”IEEE Transactions on Neural Networks and Learning Systems , 2025
2025
-
[44]
Asynchronous federated learning over wireless communication networks,
Z. Wang, Z. Zhang, Y . Tian, Q. Yang, H. Shan, W. Wang, and T. Q. Quek, “Asynchronous federated learning over wireless communication networks,” IEEE Transactions on Wireless Communications , vol. 21, no. 9, pp. 6961–6978, 2022
2022
-
[45]
Efficient asyn- chronous federated learning with prospective momentum aggregation and fine-grained correction,
Y . Zang, Z. Xue, S. Ou, L. Chu, J. Du, and Y . Long, “Efficient asyn- chronous federated learning with prospective momentum aggregation and fine-grained correction,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 15, 2024, pp. 16 642–16 650
2024
-
[46]
Energy-efficient dynamic asynchronous federated learning in mobile edge computing net- works,
G. Xu, X. Li, H. Li, Q. Fan, X. Wang, and V . C. Leung, “Energy-efficient dynamic asynchronous federated learning in mobile edge computing net- works,” in ICC 2023-IEEE international conference on communications. IEEE, 2023, pp. 160–165
2023
-
[47]
Digital twin enabled asynchronous splitfed learning in e-healthcare systems,
V . Stephanie, I. Khalil, and M. Atiquzzaman, “Digital twin enabled asynchronous splitfed learning in e-healthcare systems,” IEEE Journal on Selected Areas in Communications , vol. 41, no. 11, pp. 3650–3661, 2023
2023
-
[48]
Semi-asynchronous federated split learning for computing-limited devices in wireless net- works,
H. Ao, H. Tian, W. Ni, G. Nie, and D. Niyato, “Semi-asynchronous federated split learning for computing-limited devices in wireless net- works,” IEEE Transactions on Wireless Communications , 2025
2025
-
[49]
Boyd and L
S. Boyd and L. Vandenberghe, Convex optimization . Cambridge University Press, 2004
2004
-
[50]
On the convergence of FedAvg on non-IID data,
X. Li, K. Huang, W. Yang, S. Wang, and Z. Zhang, “On the convergence of FedAvg on non-IID data,” in International Conference on Learning Representations, 2020
2020
-
[51]
Federated learning over noisy channels: Con- vergence analysis and design examples,
X. Wei and C. Shen, “Federated learning over noisy channels: Con- vergence analysis and design examples,” IEEE Trans. Cogn. Commun. Netw., vol. 8, no. 2, pp. 1253–1268, 2022
2022
-
[52]
On the convergence of hybrid server- clients collaborative training,
K. Yang, S. Chen, and C. Shen, “On the convergence of hybrid server- clients collaborative training,” IEEE J. Select. Areas Commun. , vol. 41, no. 3, pp. 802–819, 2023
2023
-
[53]
Decoupled greedy learning of cnns,
E. Belilovsky, M. Eickenberg, and E. Oyallon, “Decoupled greedy learning of cnns,” in International Conference on Machine Learning . PMLR, 2020, pp. 736–745
2020
-
[54]
Learning multiple layers of features from tiny images,
A. Krizhevsky, “Learning multiple layers of features from tiny images,” University of Toronto, Tech. Rep., April 2009
2009
-
[55]
LEAF: A benchmark for federated settings,
S. Caldas et al. , “LEAF: A benchmark for federated settings,” arXiv preprint arXiv:1812.01097, 2018
2018 arXiv
-
[56]
On the difficulty of training recurrent neural networks,
R. Pascanu, T. Mikolov, and Y . Bengio, “On the difficulty of training recurrent neural networks,” in International conference on machine learning. PMLR, 2013, pp. 1310–1318
2013
-
[57]
Adaptive federated optimization,
S. Reddi, Z. Charles, M. Zaheer, Z. Garrett, K. Rush, J. Kone ˇcn`y, S. Kumar, and H. B. McMahan, “Adaptive federated optimization,” in International Conference on Learning Representations , 2021. Yujia Mu received her B.E. degree from Tsinghua University in 2017 and M.S. degr...
2021
-
[58]
(46) where the last inequality holds because n2 nX i=1 E h ∇Fs(xt+1 s,i ) 2i ≥ nX i=1 i iX j=1 E h ∇Fs(xt+1 s,j ) 2i
(45) Therefore, Equation (36) can be further derived as: E Fs(xt+1 s ) ≤ E Fs(xt s) − 2n − 1 4 ηtE h ∇Fs(xt s) 2i − nX i=1 ηt 2 E h ∇Fs(xt+1 s,i ) 2i 16 − L2η3 t i 2 iX j=1 E h ∇Fs(xt+1 s,j ) 2i ! + ηtG2 2 nX i=1 dt c,i + L 2 η2 t n2G2 2 ≤ E Fs(xt s) − 2n − 1 4 ηtE h ∇Fs(xt s)...
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.