Pith. sign in

REVIEW 4 major objections 5 minor 30 references

A shared super-network makes split learning up to 20x cheaper in communication and 2–5x faster to converge.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 12:36 UTC pith:LLMT4YV6

load-bearing objection Clever combination of supernet, split learning, and local heads that reads honestly, but the headline speedups rest on an uneven baseline and hand-set simulation coefficients. the 4 major comments →

arxiv 2601.02092 v2 pith:LLMT4YV6 submitted 2026-01-05 cs.DC

SuperSFL: Resource-Heterogeneous Federated Split Learning with Weight-Sharing Super-Networks

Abdullah Al Asif , Sixing Yu , Juan Pablo Munoz , Arya Mazaheri , Ali Jannesari This is my paper
classification cs.DC
keywords federated learningsplit learningsuper-networkmodel heterogeneitygradient fusionfault tolerancecommunication efficiencyedge computing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper is trying to establish that a single weight-sharing super-network, combined with a three-phase gradient-fusion rule and a fault-tolerant local classifier, can make split federated learning practical for real edge environments where devices differ in memory and network speed. The claim is that this framework converges two to five times faster in communication rounds than standard split-federated learning, reaches higher accuracy, transmits up to 20 times less data, and trains up to 13 times faster in wall-clock time. If true, this would remove two of the main obstacles to deploying federated learning at the edge: device heterogeneity and intermittent connectivity. The framework also reports better energy efficiency when measured as power per accuracy point.

Core claim

The central discovery is that heterogeneity can be absorbed by the architecture itself. A centrally hosted weight-sharing super-network contains all layers of the global model, and each client is assigned a contiguous prefix of that network—a subnetwork whose depth is determined by the client’s memory and communication latency. Training proceeds with Three-Phase Gradient Fusion (TPGF): the client computes a local loss and gradient through a lightweight classifier, the server computes a deeper loss and returns a gradient through the smashed data, and the client merges the two gradients using a weight that combines a depth ratio with an inverse-loss reliability term. A fault-tolerant client-si

What carries the argument

A weight-sharing super-network: one global model from which every client extracts a contiguous prefix (subnetwork) whose depth is set by a resource-scoring formula over memory and latency. The mechanism that makes it work is Three-Phase Gradient Fusion (TPGF): local and server branches each produce a gradient for the same encoder, and they are merged by a weight that combines a depth ratio with an inverse-loss reliability term. Fault tolerance comes from a client-side classifier that generates local updates when no server response arrives; structure-aware, per-layer aggregation with a server-consistency regularizer (λ=0.01) combines the heterogeneously deep updates.

Load-bearing premise

The resource-allocation rule in Eq. (1), with hand-set coefficients α=0.5 and β=4, together with the simulated memory/latency profiles, is assumed to faithfully represent real device heterogeneity and to be general enough to deliver the reported gains; the paper admits these coefficients were chosen empirically and offers no sensitivity analysis, so if the mapping is unrepresentative or poorly tuned, the speed-ups may be artifacts of the simulation.

What would settle it

Run the same comparison on a physical testbed with real devices spanning the claimed 2–16 GB memory and 20–200 ms latency range, sweeping α and β; if the 2–5x convergence advantage over fixed-split SFL collapses under any realistic profile mix or optimal tuning, the central claim fails. A second check: give the SFL baseline a per-client-tuned split depth; if it matches SuperSFL's speed-up, then the contribution is just depth adaptation, not gradient fusion.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single global model can serve clients with very different hardware; no per-device retraining is needed, only a different contiguous slice of the same super-network.
  • Training can survive server outages and slow networks: clients fall back to a local classifier and continue learning, then resynchronize through aggregation.
  • Communication cost and wall-clock training time drop by up to 20x and 13x respectively compared to fixed-split SFL, at least in the simulated settings evaluated.
  • The per-layer, performance-weighted aggregation rule allows the global encoder to stay coherent even when some clients were serverless during a round.

Where Pith is reading between the lines

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

  • The subnetwork-allocation idea could naturally extend from depth to width (channels per layer), giving a two-dimensional resource-accuracy frontier that the paper does not explore.
  • Part of the reported speed-up may come from adaptive depth assignment rather than gradient fusion per se; comparing against a baseline that also tunes split depth per client would isolate the fusion contribution.
  • The fault-tolerant local classifier suggests a broader asynchronous training mode where clients train locally for variable durations and sync opportunistically—an extension the paper motivates but does not optimize.
  • A stress test with mismatched local and server loss scales (e.g., very different class distributions) would reveal whether the inverse-loss weighting in TPGF remains stable beyond the reported settings.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SuperSFL, a split federated learning framework that combines a weight-sharing super-network with resource-aware subnetwork allocation, a Three-Phase Gradient Fusion (TPGF) rule, a fault-tolerant client-side classifier, and collaborative client–server aggregation. Clients are assigned contiguous prefixes of a global model according to memory and latency profiles; TPGF blends local and server-side encoder gradients; a lightweight local classifier enables progress during server outages. The authors report on CIFAR-10 and CIFAR-100 with 50 and 100 simulated heterogeneous clients that SuperSFL converges 2–5× faster in communication rounds than baseline SFL, with up to 20× lower communication cost and 13× shorter training time, while also improving accuracy and energy efficiency. The paper includes algorithm pseudocode, an ablation of the TPGF weighting, and a fault-tolerance study with varying server-gradient availability.

Significance. If the empirical claims survive a fair comparison, SuperSFL would be a useful contribution to heterogeneous split learning: it directly addresses device heterogeneity, intermittent connectivity, and communication cost in a single framework, and the algorithmic description is sufficiently clear to reproduce. The ablation in Section IV (Fig. 6) gives constructive evidence that both the depth term and the loss-reliability term in the TPGF weight are needed, and the fault-tolerance table (Table III) reports standard deviations, which is a strength. The main significance is currently limited by the simulation-only evaluation, the absence of a compute-parity baseline, and several reporting inconsistencies that make the headline numbers difficult to verify.

major comments (4)
  1. [§II-B (last paragraph) and §III-D] The paper states that TPGF incurs 'no additional communication beyond sending smashed data, identical to standard split learning' and concludes 'without increasing computational or communication overhead.' This is internally inconsistent. Algorithm 2, Phase 1, requires each client to run a forward pass through the local classifier h_φ_i, compute CrossEntropy, update φ_i, and compute the clipped encoder gradient ∇θ_i L_client; Phase 3 then fuses two encoder gradients with extra arithmetic. Section III-D explicitly admits 'additional local computation from TPGF and client-side classifiers.' Standard SFL does not maintain a local classifier and performs one client-side forward/backward per round. The reported 2–5× round reduction could therefore be an artifact of per-round compute asymmetry rather than of the super-network or TPGF. Please add a baseline that gives SFL an equivalent local cl
  2. [§II-A, Eq. (1) and §VI] The resource-allocation rule uses hand-set coefficients α=0.5 [layers/GB] and β=4, while λ=0.01 and τ=0.5 are fixed without analysis. Section VI admits that these values 'were selected empirically' and that heterogeneity was evaluated with simulated device profiles (memory uniformly sampled from [2,16] GB, latency from [20,200] ms). No sensitivity analysis is provided, so the central speed-up claims may be artifacts of the chosen coefficients and the synthetic distribution of profiles. Please provide a sensitivity sweep over α, β, λ, τ, and over the simulated resource ranges, and ideally evaluate on at least one real-device or trace-driven setting. If the coefficients are genuinely robust, show that the conclusions do not change under plausible perturbations.
  3. [Table I] Table I, which supports the headline 2–5×, 20×, and 13× claims, is garbled in the manuscript: column alignment is lost and numeric fields run together (e.g., '2305466' and '2650595' appear to be '2305 466' and '2650 595' from the text). The table also lacks error bars or standard deviations, the target accuracy changes per row, and the stopping criterion for baselines is not stated. This makes the central quantitative claims impossible to verify. Please re-render the table with clear columns, report mean±std over multiple seeds, and state the exact condition under which training was stopped for each method.
  4. [§III-B vs Table II] The text in Section III-B says that 'SFL performs significantly worse, struggling to exceed 70% accuracy even after 100 rounds,' but Table II reports SFL final accuracies of 78.84% (CIFAR-10, 50 clients), 78.25% (CIFAR-100, 50), and 77.81% (CIFAR-100, 100). These statements are inconsistent. If the text refers to a specific early stage or a different setting, that should be stated explicitly. Additionally, the DFL baseline (Dynamic Federated Learning) is not described in the experimental section, so the reader cannot tell whether the comparison is fair. Please clarify and, ideally, provide a brief description of DFL and its hyperparameters.
minor comments (5)
  1. [§IV] The ablation text refers to 'Section 2B (Eq. 3)' and later says 'These results support the design of Eq. (6).' The first should be Section II-B, and the second should likely be Eq. (3), not Eq. (6). Please correct the cross-references.
  2. [§II-A and §III-A] The symbol α is used both for the memory coefficient in Eq. (1) and for the Dirichlet concentration parameter in the non-IID partitioning. This overloading is confusing; rename one of them.
  3. [References] Reference [8] cites a mathematics paper on Dirichlet L-functions, which is not the standard reference for the Dirichlet distribution used to simulate non-IID federated data. Please replace it with a proper citation for Dirichlet-based data partitioning, e.g., the original reference used in the federated learning literature.
  4. [§III-A and Table II] Most main results are reported without variance or number of seeds, while Table III includes ± values. Please add confidence intervals or standard deviations to the headline comparisons so the reader can assess stability.
  5. [§III-C and Figure 4] The caption of Figure 4 reports 'Communication efficiency Speed-up' and 'Training efficiency Speed-up' but does not define the baseline or the formula used. Also, 'communication cost' in Table I is in MB but it is unclear whether this is total transferred data across all clients or per-client aggregated over rounds. Please define these quantities precisely.

Circularity Check

0 steps flagged

No significant circularity: all central claims are empirical measurements, not consequences of fitted inputs or self-citations.

full rationale

SuperSFL's central claims (convergence speed, communication cost, training time, energy) are empirical outcomes measured on CIFAR-10/CIFAR-100 against SFL and DFL baselines (Section III, Tables I-II), not quantities derived from its own fitted parameters. Eq. (1) is an explicit resource-allocation heuristic with hand-set α and β; Eqs. (3)-(4) and (6)-(8) are explicit fusion/aggregation rules whose behavior is tested by ablation (Figure 6), not assumed. No prediction is constructed to equal its input by definition. The only self-citation ([28], a super-network generation paper co-authored by J.P. Muñoz) appears in related work as background and is not load-bearing. Section VI openly notes that α, β, λ, τ were 'selected empirically' and that heterogeneity used simulated profiles; this is an external-validity limitation, not circularity. The inconsistency between TPGF's extra local classifier computation and the claim of 'no computational overhead' is a fairness/correctness concern about the baseline comparison, not a case of a result reducing to its inputs. No specific circular reduction can be exhibited.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The paper does not introduce new physical entities. It contributes an architectural/algorithmic pipeline built from existing concepts (super-networks, split learning, local classifiers). The free parameters are the hand-set coefficients and evaluation thresholds that directly influence the reported speed-ups.

free parameters (8)
  • α (memory coefficient) = 0.5 layers/GB
    Controls depth assignment in Eq. (1); hand-set. Authors state in Section VI that α/β were selected empirically; no sensitivity analysis.
  • β (latency coefficient) = 4
    Controls latency contribution in Eq. (1); hand-set; no sensitivity analysis.
  • λ (server consistency weight) = 0.01
    Used in aggregation Eq. (8) to pull aggregated layers toward server-side parameters; set for all experiments; no ablation.
  • τ (gradient clip threshold) = 0.5
    ℓ2-norm clip for local encoder gradients in TPGF (Algorithm 2); chosen as 'small enough not to distort learning dynamics'; no ablation.
  • Fallback timeout = 5 s
    Threshold for declaring server unreachable (Section II-C); fixed throughout; no analysis of sensitivity.
  • Dirichlet α = 0.5
    Concentration parameter for non-IID data partition; a dataset-simulation choice that affects task difficulty.
  • Simulated resource ranges = memory ∈ [2,16] GB, latency ∈ [20,200] ms
    Uniformly sampled device profiles used to define heterogeneity; hand-chosen in Section III-A.
  • Target accuracy thresholds in Table I = 70–80% depending on dataset/client count
    Defines 'convergence' for speed-up measurement; chosen per setting; affects all round/communication/time claims.
axioms (6)
  • domain assumption Contiguous-prefix subnetworks of a ViT can be trained jointly as a weight-sharing super-network without instability.
    The entire method relies on this; no convergence theory, only empirical evidence on CIFAR with ViT-16 (Section II-A, II-B).
  • domain assumption Inverse-loss weighting in Eqs. (3) and (6) is a valid measure of gradient reliability across the local and server branches.
    TPGF and aggregation weight signals by inverse losses; no theoretical justification that cross-branch losses are comparable (Section II-B, II-D).
  • domain assumption Uniformly sampled memory and latency ranges [2,16] GB and [20,200] ms are representative of real edge heterogeneity.
    Experiments simulate heterogeneity rather than using physical devices; authors list real-device evaluation as future work (Section III-A, VI).
  • domain assumption The 5-second RPC timeout is a reasonable rule for detecting server failure without harming training.
    Fallback behavior depends on this threshold; it is asserted as common practice, not validated (Section II-C).
  • domain assumption Standard FedAvg-style aggregation of prefixes is sufficient to keep a global super-network coherent across heterogeneous depths.
    Collaborative aggregation assumes layer-aligned averaging is valid for ViT prefixes (Section II-D).
  • standard math The aggregation objective Eq. (7) is convex and minimized by the closed-form Eq. (8).
    A standard least-squares computation; no issue.

pith-pipeline@v1.3.0-alltime-deepseek · 13029 in / 16156 out tokens · 143857 ms · 2026-08-03T12:36:56.330111+00:00 · methodology

0 comments
read the original abstract

SplitFed Learning (SFL) combines federated learning and split learning to enable collaborative training across distributed edge devices; however, it faces significant challenges in heterogeneous environments with diverse computational and communication capabilities. This paper proposes \textit{SuperSFL}, a federated split learning framework that leverages a weight-sharing super-network to dynamically generate resource-aware client-specific subnetworks, effectively mitigating device heterogeneity. SuperSFL introduces Three-Phase Gradient Fusion (TPGF), an optimization mechanism that coordinates local updates, server-side computation, and gradient fusion to accelerate convergence. In addition, a fault-tolerant client-side classifier and collaborative client--server aggregation enable uninterrupted training under intermittent communication failures. Experimental results on CIFAR-10 and CIFAR-100 with up to 100 heterogeneous clients show that SuperSFL converges $2$--$5\times$ faster in terms of communication rounds than baseline SFL while achieving higher accuracy, resulting in up to $20\times$ lower total communication cost and $13\times$ shorter training time. SuperSFL also demonstrates improved energy efficiency compared to baseline methods, making it a practical solution for federated learning in heterogeneous edge environments.

Figures

Figures reproduced from arXiv: 2601.02092 by Abdullah Al Asif, Ali Jannesari, Arya Mazaheri, Juan Pablo Munoz, Sixing Yu.

Figure 1
Figure 1. Figure 1: SuperSFL architecture overview. Heterogeneous clients process data through customized local layers, send optimized representations to the main server, and participate in collaborative model aggregation via the FedServer. The system dynamically adapts to each client’s capabilities while maintaining model consistency through a weight-sharing super-network. Algorithm 1 Resource-Aware Subnetwork Allocation 1: … view at source ↗
Figure 2
Figure 2. Figure 2: The Model Aggregation Process in SuperSFL involves [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparative analysis of CIFAR-100 accuracy with 50 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Communication and Training Efficiency Comparison (speed-up) Across Models and dataset with different number of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparative analysis of models based on consumption per accuracy and carbon footprint across CIFAR-10 and [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation results of the TPGF fusion rule on CIFAR-10 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

30 extracted references · 7 linked inside Pith

  1. [1]

    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,” inProceedings of the 20th International Conference on Artificial Intelligence and Statistics. Proceedings of Machine Learning Research, 2017, pp. 1273–1282. [Online]. Available: https://proceedings.mlr.press/v54/mcm...

    2017

  2. [2]

    Splitfed: When federated learning meets split learning,

    C. Thapa, P. C. Mahawaga Arachchige, S. Camtepe, and L. Sun, “Splitfed: When federated learning meets split learning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 8. AAAI Press, 2022, pp. 8485–8493

    2022

  3. [3]

    Advances and open problems in federated learning,

    P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummings, R. G. L. D’Oliveira, H. Eichner, S. E. Rouayheb, D. Evans, J. Gardner, Z. Garrett, A. Gasc ´on, B. Ghazi, P. B. Gibbons, M. Gruteser, Z. Harchaoui, C. He, L. He, Z. Huo, B. Hutchinson, J. Hsu, M. Jaggi, T. Javidi, G. Joshi, M. Khodak...

    Pith/arXiv arXiv 2021

  4. [4]

    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 Processing Magazine, vol. 37, no. 3, pp. 50–60, 2020

    2020

  5. [5]

    Splitfed: When federated learning meets split learning,

    C. Thapa, M. A. P. Chamikara, and S. Camtepe, “Splitfed: When federated learning meets split learning,”arXiv preprint arXiv:2004.12088, 2020

    Pith/arXiv arXiv 2004

  6. [6]

    Client selection for federated learning with heterogeneous resources in mobile edge,

    T. Nishio and R. Yonetani, “Client selection for federated learning with heterogeneous resources in mobile edge,” inICC 2019 - 2019 IEEE International Conference on Communications (ICC). IEEE, May 2019. [Online]. Available: http://dx.doi.org/10.1109/ICC.2019.8761315

    arXiv 2019

  7. [7]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inInternational Conference on Learning Representations (ICLR), 2021. [Online]. Available: https://openreview.net/forum?...

    2021

  8. [8]

    Distribution of dirichlet l-functions,

    Z. Dong, W. Wang, and H. Zhang, “Distribution of dirichlet l-functions,” Mathematika, vol. 69, no. 3, pp. 719–750, 2023

    2023

  9. [9]

    Exploring the accuracy – energy trade-off in machine learning,

    A. E. Brownlee, J. Adair, S. O. Haraldsson, and J. Jabbo, “Exploring the accuracy – energy trade-off in machine learning,” in2021 IEEE/ACM International Workshop on Genetic Improvement (GI), 2021, pp. 11–18

    2021

  10. [10]

    Federated learning for internet of things: A comprehensive survey,

    D. C. Nguyen, M. Ding, P. N. Pathirana, A. Seneviratne, J. Li, and H. V . Poor, “Federated learning for internet of things: A comprehensive survey,”IEEE Communications Surveys & Tutorials, vol. 23, no. 3, pp. 1622–1658, 2021

    2021

  11. [11]

    Federated machine learning: Concept and applications,

    Q. Yang, Y . Liu, T. Chen, and Y . Tong, “Federated machine learning: Concept and applications,”ACM Transactions on Intelligent Systems and Technology (TIST), vol. 10, no. 2, pp. 1–19, 2019

    2019

  12. [12]

    Federated learning for resource-constrained iot devices: Panoramas and state-of- the-art,

    A. Imteaj, U. Thakker, S. Wang, J. Li, and M. H. Amini, “Federated learning for resource-constrained iot devices: Panoramas and state-of- the-art,”arXiv preprint arXiv:2002.10610, 2020

    Pith/arXiv arXiv 2002

  13. [13]

    Split learning for health: Distributed deep learning without sharing raw patient data,

    P. Vepakomma, O. Gupta, T. Swedish, and R. Raskar, “Split learning for health: Distributed deep learning without sharing raw patient data,” in Proceedings of the International Conference on Learning Representations Workshop, 2018

    2018

  14. [14]

    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

    2018

  15. [15]

    Split computing and early exiting for deep learning applications: Survey and research challenges,

    Y . Matsubara, M. Levorato, and F. Restuccia, “Split computing and early exiting for deep learning applications: Survey and research challenges,” arXiv preprint arXiv:2103.04505, 2021

    Pith/arXiv arXiv 2021

  16. [16]

    Fedsl: Federated split learning on dis- tributed sequential data in recurrent neural networks,

    A. Abedi and S. U. Khan, “Fedsl: Federated split learning on dis- tributed sequential data in recurrent neural networks,”arXiv preprint arXiv:2011.03180, 2020

    Pith/arXiv arXiv 2011

  17. [17]

    Federated or split? a performance and privacy analysis of hybrid split and federated learning architectures,

    V . Turina, Z. Zhang, F. Esposito, and I. Matta, “Federated or split? a performance and privacy analysis of hybrid split and federated learning architectures,” in2021 IEEE 14th International Conference on Cloud Computing (CLOUD). IEEE, 2021, pp. 250–260

    2021

  18. [18]

    Dfl: Dynamic federated split learning in heterogeneous iot,

    E. Samikwa, A. D. Maio, and T. Braun, “Dfl: Dynamic federated split learning in heterogeneous iot,”IEEE Internet of Things Journal, vol. 2, 2024

    2024

  19. [19]

    Grouping synchronous to eliminate stragglers with edge computing in distributed deep learning,

    Z. Gui, S. Shi, B. Li, and X. Chu, “Grouping synchronous to eliminate stragglers with edge computing in distributed deep learning,” inIEEE International Conference on Parallel and Distributed Processing with Applications. IEEE, 2021, pp. 429–436

    2021

  20. [20]

    Fedmask: Joint computation and communication-efficient personalized federated learning via heterogeneous masking,

    A. Li, J. Sun, X. Zeng, M. Zhang, H. Li, and Y . Chen, “Fedmask: Joint computation and communication-efficient personalized federated learning via heterogeneous masking,” inProceedings of the ACM Conference on Embedded Networked Sensor Systems, 2021, pp. 42–55

    2021

  21. [21]

    Clustered federated learning: Model-agnostic distributed multitask optimization under privacy con- straints,

    F. Sattler, K.-R. M ¨uller, and W. Samek, “Clustered federated learning: Model-agnostic distributed multitask optimization under privacy con- straints,”IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 8, pp. 3710–3722, 2020

    2020

  22. [22]

    Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach,

    A. Fallah, A. Mokhtari, and A. Ozdaglar, “Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach,” inAdvances in Neural Information Processing Systems, vol. 33, 2020

    2020

  23. [23]

    Improving federated learning personalization via model agnostic meta learning,

    Y . Jiang, J. Koneˇcn`y, K. Rush, and S. Kannan, “Improving federated learning personalization via model agnostic meta learning,”arXiv preprint arXiv:1909.12488, 2019

    Pith/arXiv arXiv 1909

  24. [24]

    Personalized federated learning using hypernetworks,

    A. Shamsian, A. Navon, E. Fetaya, and G. Chechik, “Personalized federated learning using hypernetworks,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 9489–9502

    2021

  25. [25]

    Heterofl: Computation and com- munication efficient federated learning for heterogeneous clients,

    E. Diao, J. Ding, and V . Tarokh, “Heterofl: Computation and com- munication efficient federated learning for heterogeneous clients,” in International Conference on Learning Representations, 2021

    2021

  26. [26]

    Universally slimmable networks and improved training techniques,

    J. Yu and T. S. Huang, “Universally slimmable networks and improved training techniques,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 1803–1811

    2019

  27. [27]

    Once for all: Train one network and specialize it for efficient deployment,

    H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han, “Once for all: Train one network and specialize it for efficient deployment,”arXiv preprint arXiv:1908.09791, 2019

    Pith/arXiv arXiv 1908

  28. [28]

    Enabling nas with automated super-network generation,

    J. P. Mu ˜noz, N. Lyalyushkin, Y . Akhauri, A. Senina, A. Kozlov, and N. Jain, “Enabling nas with automated super-network generation,”1st International Workshop on Practical Deep Learning in the Wild at AAAI, 2022. [Online]. Available: https://practical-dl.github.io/2022/short paper/21.pdf

    2022

  29. [29]

    Alphanet: Improved training of supernets with alpha-divergence,

    D. Wang, C. Gong, M. Li, and V . Chandra, “Alphanet: Improved training of supernets with alpha-divergence,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 10 760–10 771

    2021

  30. [30]

    Attentivenas: Improving neural architecture search via attentive sampling,

    D. Wang, M. Li, C. Gong, and V . Chandra, “Attentivenas: Improving neural architecture search via attentive sampling,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 6418–6427

    2021