Energy-Efficient Split Learning for Fine-Tuning Large Language Models in Edge Networks

Liang Li; Shaohua Wu; Songge Zhang; Zuguang Li

REVIEW 3 major objections 5 minor 14 references

Energy-Efficient Split Learning for Fine-Tuning Large Language Models in Edge Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Adaptive split learning can cut LLM fine-tuning delay by 70.8 percent and server energy by 53.1 percent versus separate baselines.

desk verdict Clean per-device optimization, but the headline efficiency numbers ignore server contention and are not supported as stated. read the letter →

arxiv 2412.00090 v2 pith:PZN23MRC submitted 2024-11-27 cs.LG cs.CLcs.DC

Zuguang Li , Shaohua Wu , Liang Li , Songge Zhang This is my paper · ORCID

classification cs.LGcs.CLcs.DC

keywords cutlayerselectionfine-tuninglargelanguagemodelsedgenetworkssplitlearningLoRAenergyefficiencyresourceallocation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper aims to show that split learning, which runs the first layers of a large language model on mobile devices and the remaining layers on an edge server, can fine-tune models on private edge data without the full device burden of federated learning or the full server energy bill of centralized training. The proposed controller, CARD, decides two things per device and per training round: where to cut the model and how fast the server GPU should run, using a weighted, normalized cost that trades delay against server energy. Its reported result is a 70.8 percent average training-delay reduction over the device-only baseline and a 53.1 percent server-energy reduction over the server-only baseline, on a five-device testbed with a 32-layer LLM. If correct, this gives edge operators a simple per-round knob for balancing latency and energy as channels and device capabilities change.

What carries the argument

The load-bearing object is the CARD algorithm, which decomposes a mixed-integer nonlinear program into two nested decisions. Given a cut layer $c_{m,n}$, the upper layer solves a convex problem in the server GPU frequency $f^S_{m,n}$, whose optimal value is the closed-form cubic root $Q = \sqrt[3]{w(E_{\max}-E_{\min})/(2\xi(1-w)(D_{\max}-D_{\min}))}$ clipped to the feasible frequency interval. The lower layer then brute-forces the cut layer over the $I$ transformer layers, since the transmitted-data-size functions make the cost non-convex, giving $O(I)$ complexity per device per round. The objective $U$ is a weighted sum of min-max normalized delay and energy, so CARD is explicitly selecting points on a weakly Pareto-optimal trade-off curve between latency and server power.

What would settle it

Run two or more edge devices through the same split fine-tuning workload on one server, all with the same optimal cut layer, and measure the per-round wall-clock time as additional devices start training at the same moment. If the per-round delay rises with the number of concurrent devices, due to server GPU serialization or uplink contention, the independence-based delay model in Eq. (10) is falsified, and the 70.8 percent saving should be re-derived with a shared-resource term.

Watch

Extended reading notes

Core claim

The central claim is that jointly adapting the cut layer and the server's GPU frequency on a per-device, per-round basis makes LoRA-based split fine-tuning of LLMs substantially cheaper in both time and server energy than static splits. The paper derives a per-round delay model (device and server computation plus smashed-data, gradient, and LoRA-adapter transmission) and a server energy model cubic in GPU frequency, then minimizes a weakly Pareto-optimal weighted sum of normalized delay and energy. For a fixed cut layer the frequency subproblem is convex and yields a closed-form cubic-root solution; over the finite number of transformer layers the cut layer is found by exhaustive search. Simulations on five heterogeneous edge devices show the optimal cut is always all layers or none, with weaker devices offloading more to the server, and report the 70.8 percent and 53.1 percent savings against the two baselines.

Load-bearing premise

The load-bearing premise is that a device's per-round training delay is only its own local computation plus its own transmission time, with no queueing or contention for the shared server GPU or uplink; if several devices train simultaneously, that independence breaks and the reported delay reduction no longer follows directly.

Editorial extensions

If this is right

A per-device, per-round adaptation rule lets the system track wireless channel changes without solving a global joint problem.
Weak devices will systematically offload the entire transformer stack, and the server will assign them higher GPU frequencies, reducing the impact of device heterogeneity on fine-tuning time.
The weighting factor $w$ gives an operator a direct knob: raise it to favor low delay, lower it to favor low server energy, without changing the algorithm.
On a 32-layer model the optimal cut collapses to layer 0 or layer 32, so the practical control decision becomes a simple all-or-nothing offload rule in the simulated regime.

Reading between the lines

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

If multiple devices share one edge server in the same round, the delay model needs a queueing or serialization term for the server GPU and uplink; without it, the reported 70.8 percent figure is best read as a per-device bound rather than a system-level guarantee.
The all-or-nothing structure of the optimal cut suggests the cost may be monotone in the cut layer when every transformer layer has equal FLOPs and equal smashed-data size, which would let a threshold rule replace the exhaustive scan.
The same two-level CARD machinery transfers to split inference or to other parameter-efficient fine-tuning methods, provided per-layer computation and transmitted-data sizes are known.
A direct stress test is to increase the number of participating devices while fixing total data volume; the model predicts unchanged per-round delay, whereas contention in a real deployment would make delay grow.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: the paper is a clean, small optimization contribution to split learning for LLM fine-tuning, but the headline 70.8%/53.1% numbers are not supported as stated because the model treats devices as independent when they share one server GPU and one wireless channel.

What's new: the CARD algorithm jointly selects the cut layer and server GPU frequency for LoRA-based split fine-tuning. The closed-form frequency optimum in Eq. (16) follows from a correct convexity check, and the brute-force scan over cut layers is valid. For a single device, the problem is solved correctly, and the derivation is self-contained. This is a modest but honest contribution to edge-AI resource allocation.

The main soft spot is the per-device independence assumption. A single physical server cannot run at different frequencies for different devices at the same time, and if devices are time-shared, later devices inherit queueing delay. The stress-test note is right: Eq. (10) has no contention term, so the reported delay reductions are computed under a model that may be infeasible for the described single-server deployment. This is a correctness risk for the headline claims, not a missing optional extension. The paper should either switch to a time-division framing where total system time is the sum of per-device slots, or add a contention-aware delay model.

Two smaller issues. There is no static-cut baseline, so it's unclear how much of the gain comes from adapting the cut layer rather than simply tuning the server frequency. And no fine-tuning quality metric is reported — no loss or accuracy — so we don't know whether the chosen split preserves model quality. For a letters paper, one of these alone is acceptable; together they weaken the evaluation.

Proportionate bottom line: the optimization core is sound, the writing is clear, and the problem is worth solving. The paper deserves a serious referee. I would send it back for a revision addressing the contention issue before treating the efficiency claims as robust, and I wouldn't cite it until then. I'd bring it to a reading group; the single-GPU frequency issue is a good prompt for discussing what modeling abstractions are acceptable in a short-format venue.

Referee Report

3 major / 5 minor

Summary. The paper proposes an energy-efficient split learning (SL) framework for fine-tuning large language models (LLMs) in edge networks. Multiple devices with heterogeneous GPUs collaborate with a single edge server; each device fine-tunes the first part of the model with LoRA adapters up to a cut layer, and the server handles the remainder. The authors formulate an optimization problem P1 that minimizes a weighted sum of normalized training delay and server energy consumption, with variables being the per-device-per-round cut layer and server GPU frequency. They propose the CARD algorithm, which decomposes P1 into independent per-device subproblems P2, solves the frequency decision in closed form via convexity (Eq. (16)), and finds the cut layer by brute-force scanning over the finite number of transformer layers. Simulations with a 1B LLaMA 3.2 model and 5 Jetson devices report a 70.8% reduction in average training delay relative to the device-only baseline and a 53.1% reduction in server energy consumption relative to the server-only baseline.

Significance. If the system model were faithful, the result would be a useful and simple mechanism for jointly choosing the split layer and server frequency. The derivation is self-contained: the convexity argument for Eq. (16) is correct, the brute-force search over the finite cut-layer set is valid, and the reported delay/energy comparisons are not fitted to the model (the parameters are fixed a priori). The paper also clearly identifies a real problem—static split strategies and fixed server resources are suboptimal under heterogeneity and channel dynamics. However, the central modeling assumption that devices can be optimized independently is questionable for a single shared edge server and access point, and the paper provides no validation of fine-tuning quality. These issues significantly reduce the confidence in the headline numbers.

major comments (3)

[Section IV, Eq. (10)-Eq. (14)] The decomposition of P1 into independent per-device subproblems P2 is not valid for the described system with a single edge server and a single AP. Eq. (10) defines D_{m,n} as the sum of local computation, server computation, and transmission times, with no queueing or contention. If several devices transmit smashed data concurrently or wait on the same server GPU, D_{m,n} depends on the cut layers and schedules of all other devices. Moreover, the variable f^S_{m,n} in Eq. (13) implies each device has its own server GPU frequency, but one physical GPU cannot run at several different frequencies simultaneously. If the server processes devices sequentially instead, the total training delay should include waiting for earlier devices, and the objective in P1 is not separable. The reported reductions of 70.8% and 53.1% (Section V-B, Fig. 4) are therefore not directly supported by the model as stated.
[Section V] No fine-tuning performance (training loss, validation loss, or downstream task accuracy) is reported. The paper defines a global loss objective in Eq. (1) but the experiments in Section V only show delay and energy consumption. It is possible that dynamically changing the cut layer and server frequency affects convergence or final model quality. Without evidence that the proposed scheme actually fine-tunes the model to a satisfactory level, the claim of an "energy-efficient fine-tuning framework" is incomplete. A comparison of the training loss or task accuracy with the two baselines is needed.
[Section V-B, Fig. 3] The observation that the optimal cut layer is always either 32 or 0, and never an intermediate value, indicates that the proposed optimization is essentially a binary offloading decision under the assumed LLaMA model. The paper's statement that "the optimal cut layer of each device dynamically changes" is misleading because it changes only between the two extremes. This should be explained (e.g., by the assumed equal per-layer FLOPs and smashed-data sizes) and its implications for the generality of the CARD algorithm should be discussed.

minor comments (5)

[Abstract and Section V-B] The abstract states that the proposed approach reduces delay and server energy by 70.8% and 53.1%, but it does not specify that the delay reduction is relative to the device-only baseline and the energy reduction is relative to the server-only baseline. This should be clarified.
[Eq. (9)] The symbol \tilde{S}(c_{m,n}) in Eq. (9) is used for the gradient data size but is not defined in the text; please define it explicitly, e.g., alongside the smashed data size S(c_{m,n}).
[Fig. 3 and accompanying text] The sentence "the optimal cut layer of each device dynamically changes with training rounds" in Section V-B conflicts with the immediately following sentence that the optimal cut is either 32 or 0; please rephrase to avoid implying more than one intermediate cut layer is ever selected.
[Section III-A-2] The notation for transmission rates R^D_{m,n} and R^S_{m,n} is introduced, but the subscript/superscript convention is not consistently used later (e.g., in Eq. (9) the uplink and downlink rates appear as R^D and R^S, which matches, but the text at the end of the paragraph uses R_{m,n} generically). Please ensure consistent notation.
[Section V-A, Table II] No sensitivity analysis is provided for the weighting factor w (set to 0.2). Since w controls the trade-off between delay and energy, a short discussion of its influence (or a supplementary figure) would strengthen the evaluation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the CARD optimization is derived from convexity and exhaustive cut-layer search with no fitted inputs or self-citation load-bearing steps.

full rationale

The paper's derivation chain is self-contained. Problem P1 minimizes a normalized weighted sum of per-device training delay and server energy over two decision variables: the cut layer c_{m,n} (bounded integer) and the server GPU frequency f^S_{m,n} (bounded continuous). For a fixed cut layer, the optimal frequency in Eq. (16) is obtained by differentiating the cost U, verifying convexity via the second derivative, and clamping the stationary point to the feasible frequency interval. The closed-form Q depends only on constants: the weight w, the power coefficient xi, and the boundary normalization values D_min, D_max, E_min, E_max, which are computed from the extreme cut layers (c=0 and c=I) and frequency limits, not from measured delay or energy results. The cut layer is then selected by exhaustive search over the finite set of transformer layers in Algorithm 1, so no optimization output is fed back into the model as a fitted input. The reported 70.8% delay reduction and 53.1% server-energy reduction are simulation results comparing this derived policy against two fixed baselines (server-only and device-only), not predictions of quantities used to fit the policy. I find no self-definitional step, no fitted input renamed as a prediction, and no load-bearing self-citation: the cited references provide standard models (e.g., FLOP counts, CQI-to-MCS mapping, weakly Pareto optimality) but the core optimization is derived analytically within the paper. The skeptical concern about omitted server contention among multiple devices is a modeling or correctness risk about the independence decomposition in Section IV, not a circularity, because it does not make the derivation's outputs equivalent to its inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The ledger lists the simulation constants and modeling assumptions that the CARD solution and its headline reductions depend on. No new physical entities are postulated. The most consequential assumptions are per-device independence and the uniform per-layer workload model.

free parameters (4)

w = 0.2
Weighting factor in the normalized cost function Eq. (12), chosen by hand. It controls the delay versus server energy tradeoff and therefore affects the reported reductions and the frequency formula Eq. (16).
xi = 1e-25 W/(cycle/s)^3
GPU power coefficient in Eq. (11), assumed for the server architecture. It enters the closed-form optimum Q in Eq. (16) and is not measured or justified from hardware data.
phi = 0.1
Compression ratio for smashed data and gradients in Eq. (9), set in Table II. It directly scales the transmission delay for every device.
T_m,n = 5
Number of local epochs per training round, set in Table II. It scales both computation delay and server energy linearly in the model.

assumptions (5)

domain assumption Devices are independent and can be optimized separately with no server contention.
Section IV decomposes P1 into per-device subproblems. Eq. (10) defines D_m,n without any queueing or shared-server term, so the reported per-device delays ignore waiting for other devices.
domain assumption Lossless or adequately compressed transmission of smashed data and gradients preserves centralized LoRA training equivalence.
The paper never reports fine-tuning loss or task accuracy. The claim that the framework is a valid fine-tuning method rests on this standard but unverified assumption, especially with compression ratio phi = 0.1.
domain assumption Each transformer layer has equal computation workload and equal smashed-data size.
Section V-B concludes that every optimal cut layer is either 0 or 32 because layers are uniform. This assumption drives the result that dynamic cut selection degenerates to a binary offload decision.
domain assumption Server GPU power consumption scales with the cube of the GPU frequency.
Eq. (11) uses P = xi f^3. This is a standard empirical model but is not validated for the RTX 4060Ti used in the simulation.
domain assumption The CQI-to-MCS mapping y(.) exactly determines the achievable transmission rate from the SNR.
Eq. (9) uses R = B y(SNR) with no uncertainty, overhead, or retransmission model. This makes the transmission delay deterministic in the optimization.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Energy-Efficient Split Learning for Fine-Tuning Large Language Models in Edge Networks." pith.science (2026). https://pith.science/paper/PZN23MRC

@misc{pith2026241200090,
  author       = {Pith},
  title        = {Pith review of: Energy-Efficient Split Learning for Fine-Tuning Large Language Models in Edge Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PZN23MRC}},
  note         = {Machine review of arXiv:2412.00090}
}

read the original abstract

In this letter, we propose an energy-efficient split learning (SL) framework for fine-tuning large language models (LLMs) using geo-distributed personal data at the network edge, where LLMs are split and alternately across massive mobile devices and an edge server. Considering the device heterogeneity and channel dynamics in edge networks, a \underline{C}ut l\underline{A}yer and computing \underline{R}esource \underline{D}ecision (CARD) algorithm is developed to minimize training delay and energy consumption. Simulation results demonstrate that the proposed approach reduces the average training delay and server's energy consumption by 70.8% and 53.1%, compared to the benchmarks, respectively.

Figures

Figures reproduced from arXiv: 2412.00090 by the authors.

**Figure 2.** LoRA-based fine-tuning. For the edge server deployed in the AP, a server-side pretrained model and its corresponding set of trainable LoRA adapters are denoted by WS m and R S m, respectively. Similarly, R S m = {A cm+1 m , B cm+1 m , . . . , A I m, B I m}, where I is the total number of transform layers in the LLM. The global pretrained LLM is denoted by W = {WD m; WS m}, ∀m ∈ M. The objective of split LoRA-based… view at source ↗

**Figure 1.** The proposed SL framework for fine-tuning LLMs. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 4.** Comparison of performance in different methods. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 11 canonical work pages

[1]

Large language model-based wireless network design,

K. Qiu, S. Bakirtzis, I. Wassell, H. Song, J. Zhang, and K. Wang, “Large language model-based wireless network design,” IEEE Wireless Commun. Lett., 2024

work page 2024
[2]

LLM-based edge intelligence: A compreh ensive survey on architectures, applications, security and trust worthiness,

O. Friha, M. Amine Ferrag, B. Kantarci, B. Cakmak, A. Ozgu n, and N. Ghoualmi-Zine, “LLM-based edge intelligence: A compreh ensive survey on architectures, applications, security and trust worthiness,” IEEE Open J. Commun. Soc. , vol. 5, pp. 5799–5856, 2024

work page 2024
[3]

Gradient-based parameter selection for efﬁcien t ﬁne-tuning,

Z. Zhang, Q. Zhang, Z. Gao, R. Zhang, E. Shutova, S. Zhou, a nd S. Zhang, “Gradient-based parameter selection for efﬁcien t ﬁne-tuning,” in Proc. IEEE/CVF CVPR , 2024, pp. 28 566–28 577

work page 2024
[4]

Knowledge-d riven deep learning paradigms for wireless network optimization in 6G,

R. Sun, N. Cheng, C. Li, F. Chen, and W. Chen, “Knowledge-d riven deep learning paradigms for wireless network optimization in 6G,” IEEE Netw., vol. 38, no. 2, pp. 70–78, 2024

work page 2024
[5]

To talk or to work: Flexible communication compression for energy efﬁcient fe derated learn- ing over heterogeneous mobile edge devices,

L. Li, D. Shi, R. Hou, H. Li, M. Pan, and Z. Han, “To talk or to work: Flexible communication compression for energy efﬁcient fe derated learn- ing over heterogeneous mobile edge devices,” in Proc. IEEE INFOCOM , 2021, DOI:10.1109/INFOCOM42981.2021.9488839

arXiv 2021
[6]

ChatGPT in the age of generative AI and large language model s: A concise survey,

S. Mohamadi, G. Mujtaba, N. Le, G. Doretto, and D. A. Adjer oh, “ChatGPT in the age of generative AI and large language model s: A concise survey,” arXiv:2307.04251, 2023

arXiv 2023
[7]

Plut o and Charon: A time and memory efﬁcient collaborative edge AI fra mework for personal LLMs ﬁne-tuning,

B. Ouyang, S. Y e, L. Zeng, T. Qian, J. Li, and X. Chen, “Plut o and Charon: A time and memory efﬁcient collaborative edge AI fra mework for personal LLMs ﬁne-tuning,” in Proc. ACM ICPP , 2024, pp. 762–771

work page 2024
[8]

Split learning over wireless networks: Parallel design an d resource management,

W. Wu, M. Li, K. Qu, C. Zhou, X. Shen, W. Zhuang, X. Li, and W. Shi, “Split learning over wireless networks: Parallel design an d resource management,” IEEE J. Sel. Areas Commun. , vol. 41, no. 4, pp. 1051– 1066, 2023

work page 2023

Show all 14 references

[9]

Device-edge cooperative ﬁn e-tuning of foundation models as a 6G service,

H. Wu, X. Chen, and K. Huang, “Device-edge cooperative ﬁn e-tuning of foundation models as a 6G service,” IEEE Wireless Commun. , vol. 31, no. 3, pp. 60–67, 2024

2024
[10]

Resource allocation for stable LLM t raining in mobile edge computing,

C. Liu and J. Zhao, “Resource allocation for stable LLM t raining in mobile edge computing,” in Proc. ACM MobiHoc , 2024, pp. 81–90

2024
[11]

Federated ﬁne-tuning for pre- trained foundation models over wireless networks,

Z. Wang, Y . Zhou, Y . Shi, K. Letaief et al., “Federated ﬁne-tuning for pre- trained foundation models over wireless networks,” arXiv:2407.02924, 2024

2024 arXiv
[12]

NR; Physical layer procedures for data,

3GPP, “NR; Physical layer procedures for data,” docume nt TS 38.214 V17.3.0, Sep. 2022

2022
[13]

D elay-aware microservice coordination in mobile edge computing: A rein forcement learning approach,

S. Wang, Y . Guo, N. Zhang, P . Y ang, A. Zhou, and X. Shen, “D elay-aware microservice coordination in mobile edge computing: A rein forcement learning approach,” IEEE Trans. Mobile Comput. , vol. 20, no. 3, pp. 939–951, 2019

2019
[14]

Spinquant: LLM qua ntization with learned rotations,

Z. Liu, C. Zhao, I. Fedorov, B. Soran, D. Choudhary, R. Kr ishnamoorthi, V . Chandra, Y . Tian, and T. Blankevoort, “Spinquant: LLM qua ntization with learned rotations,” arXiv preprint arXiv:2405.16406 , 2024

2024 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Large language model-based wireless network design,

K. Qiu, S. Bakirtzis, I. Wassell, H. Song, J. Zhang, and K. Wang, “Large language model-based wireless network design,” IEEE Wireless Commun. Lett., 2024

work page 2024

[2] [2]

LLM-based edge intelligence: A compreh ensive survey on architectures, applications, security and trust worthiness,

O. Friha, M. Amine Ferrag, B. Kantarci, B. Cakmak, A. Ozgu n, and N. Ghoualmi-Zine, “LLM-based edge intelligence: A compreh ensive survey on architectures, applications, security and trust worthiness,” IEEE Open J. Commun. Soc. , vol. 5, pp. 5799–5856, 2024

work page 2024

[3] [3]

Gradient-based parameter selection for efﬁcien t ﬁne-tuning,

Z. Zhang, Q. Zhang, Z. Gao, R. Zhang, E. Shutova, S. Zhou, a nd S. Zhang, “Gradient-based parameter selection for efﬁcien t ﬁne-tuning,” in Proc. IEEE/CVF CVPR , 2024, pp. 28 566–28 577

work page 2024

[4] [4]

Knowledge-d riven deep learning paradigms for wireless network optimization in 6G,

R. Sun, N. Cheng, C. Li, F. Chen, and W. Chen, “Knowledge-d riven deep learning paradigms for wireless network optimization in 6G,” IEEE Netw., vol. 38, no. 2, pp. 70–78, 2024

work page 2024

[5] [5]

To talk or to work: Flexible communication compression for energy efﬁcient fe derated learn- ing over heterogeneous mobile edge devices,

L. Li, D. Shi, R. Hou, H. Li, M. Pan, and Z. Han, “To talk or to work: Flexible communication compression for energy efﬁcient fe derated learn- ing over heterogeneous mobile edge devices,” in Proc. IEEE INFOCOM , 2021, DOI:10.1109/INFOCOM42981.2021.9488839

arXiv 2021

[6] [6]

ChatGPT in the age of generative AI and large language model s: A concise survey,

S. Mohamadi, G. Mujtaba, N. Le, G. Doretto, and D. A. Adjer oh, “ChatGPT in the age of generative AI and large language model s: A concise survey,” arXiv:2307.04251, 2023

arXiv 2023

[7] [7]

Plut o and Charon: A time and memory efﬁcient collaborative edge AI fra mework for personal LLMs ﬁne-tuning,

B. Ouyang, S. Y e, L. Zeng, T. Qian, J. Li, and X. Chen, “Plut o and Charon: A time and memory efﬁcient collaborative edge AI fra mework for personal LLMs ﬁne-tuning,” in Proc. ACM ICPP , 2024, pp. 762–771

work page 2024

[8] [8]

Split learning over wireless networks: Parallel design an d resource management,

W. Wu, M. Li, K. Qu, C. Zhou, X. Shen, W. Zhuang, X. Li, and W. Shi, “Split learning over wireless networks: Parallel design an d resource management,” IEEE J. Sel. Areas Commun. , vol. 41, no. 4, pp. 1051– 1066, 2023

work page 2023