Energy Efficient Federated Learning with Hyperdimensional Computing over Wireless Communication Networks

Jiaxiang Wang; Mohammad Shikh-Bahaei; Yahao Ding; Yinchao Yang; Zhaohui Yang; Zhu Han

REVIEW 3 major objections 5 minor 1 cited by

Energy Efficient Federated Learning with Hyperdimensional Computing over Wireless Communication Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing neural-network local models with hyperdimensional computing in wireless federated learning, and jointly optimizing the hypervector dimension with bandwidth, transmit power, transmission time, and CPU frequenc

desk verdict The resource-allocation core is mostly sound, but the headline numbers (d=4000, 83.3%) rest on an unvalidated curve fit and a dimension subproblem that drops the rate constraint; keep the framework, don't quote the savings. read the letter →

arxiv 2602.21949 v2 pith:4RL6OWLP submitted 2026-02-25 cs.DC

Yahao Ding , Yinchao Yang , Jiaxiang Wang , Zhaohui Yang , Zhu Han , Mohammad Shikh-Bahaei This is my paper · ORCID

classification cs.DC

keywords federatedlearninghyperdimensionalcomputingdifferentialprivacyenergyefficiencyresourceallocationwirelessnetworksconvergenceroundszCDP

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 tries to establish that the energy cost of wireless federated learning can be cut dramatically by replacing neural-network local models with hyperdimensional computing (HDC) and then treating the HDC dimension itself as an allocatable resource. The authors propose FL-HDC-DP, a framework in which each user trains a high-dimensional bipolar class hypervector with simple vector operations, adds Gaussian noise calibrated by differential privacy, and uploads it over an FDMA link; the paper's optimization jointly chooses the hypervector dimension, transmission time, bandwidth, transmit power, and CPU frequency to minimize total energy under latency and privacy constraints. They report an interior optimum at roughly d=4,000 dimensions and an energy reduction of up to 83.3% against a neural-network baseline, with about 90% accuracy reached in approximately 3.5× fewer communication rounds. A sympathetic reader would care because, if true, resource-constrained edge devices could run private federated learning with a fraction of the energy and without backpropagation. The size of the claimed saving, however, depends on an empirical sigmoid-variant curve that maps dimension to convergence rounds, fitted to the authors' own MNIST simulations with no reported error bars or out-of-sample checks.

What carries the argument

The load-bearing object is the empirical sigmoid-variant function J_d(d)=μ+ν/(1+e^{β(log d−α)}), which connects the hypervector dimension to the number of communication rounds needed to hit a target accuracy. The paper uses it to make the discrete round count a continuous function of d, so that the energy objective can be differentiated and dimension can be searched by solving a one-dimensional stationarity equation and projecting onto the discrete candidate set. Around this sit the energy model (per-dimension CPU cycles for encoding, aggregation, and retraining, plus transmission energy J·t_i·p_i), zCDP noise calibration from sensitivity κ and √2κ, and the alternating optimization that deri

What would settle it

Run the proposed joint optimization with empirically measured convergence-round counts for each candidate dimension (rather than substituting the fitted curve): simulate FL-HDC-DP on MNIST for d=3000,4000,...,10000 with ε=25, δ=10^-5, measuring the rounds to reach 88% accuracy, and plug those measured J_d values into the energy objective. If the measured J_d values differ materially from Eq. (26) with the stated coefficients, the predicted interior optimum d=4,000 and the 83.3% energy saving will not reproduce.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that FL-HDC-DP, a federated learning framework using HDC for local training and Gaussian DP noise for privacy, admits a jointly optimal allocation of HDC dimension d, uplink transmission time, system bandwidth, transmit power, and CPU frequency that minimizes total user energy subject to latency and privacy constraints. The key empirical discovery is that total energy is non-monotonic in d, with a minimum at d=4,000 under the tested settings: raising d from 3,000 to 4,000 cuts required rounds enough to lower energy, while further increases add per-round computation and transmission cost without enough convergence benefit. The paper also claims t

Load-bearing premise

The entire energy-saving claim rests on the four-parameter sigmoid fit J_d(d)=μ+ν/(1+e^{β(log d−α)}) accurately predicting convergence rounds for all candidate dimensions and privacy budgets, yet the fit is against the authors' own MNIST simulations with no error bars or held-out settings; if the true rounds-vs-dimension curve deviates from the fit, the optimal d=4,000 and the 83.3% saving shift.

Editorial extensions

If this is right

Energy optimization should treat model dimension as a radio-style resource: a moderate dimension (≈4,000) beats both too-small and too-large dimensions, so dimension and resource allocation should be co-designed rather than fixed a priori.
Formal privacy is achievable at low overhead: calibrating Gaussian noise from HDC's sensitivity (κ, then √2κ) under zCDP gives an end-to-end (ε,δ) guarantee without backpropagation.
Convergence speed translates directly into energy: reaching about 90% accuracy in roughly 3.5× fewer rounds than a neural baseline reduces both computation and transmission energy.
The resource allocation pattern is systematic: as channel quality degrades, users receive more bandwidth and higher transmit power, with CPU frequency rising to fit computation into the remaining time budget.
Wireless bandwidth acts as a substitute for transmit power and dimension: the energy curves flatten at high bandwidth, so operators can trade spectrum against computation.

Reading between the lines

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

The interior optimum d≈4,000 is likely a structural consequence of saturating convergence gains: any setting where per-round cost grows linearly with d while rounds plateau at high d will have an interior minimum, so the same machinery could be reused after estimating a fresh J_d curve.
Because J_d(d) is fitted to MNIST with fixed privacy budgets, the 83.3% figure should be read as conditional; a testable extension is to refit Eq. (26) on non-IID data, other datasets, and other ε values, and check whether the qualitative interior optimum and savings persist.
The framework suggests a broader design principle: in federated learning, the right model size depends on the wireless environment, so model-radio co-design may matter more than optimizing either layer alone.
A natural extension is to allow the dimension to vary per round or per user, rather than using a single global d for all rounds and all users.

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 worth engaging, but the headline figures are not. What's actually new is the first co-optimization of HDC dimension with bandwidth, power, transmission time, and CPU frequency under differential privacy, plus the sigmoid-variant J_d(d) model for convergence rounds. I checked the core convex derivations myself: Lemma 1's zCDP calibration is correct, the tight-rate argument in Theorem 1 and Prop. 1 is right, the second-derivative signs in (40)-(41) are correct, and Prop. 2's feasibility certificate is sound. That is real and useful work, extending the Yang et al. / Chen et al. energy-minimization program to HDC.

Now the soft spots. The dimension subproblem is not the problem you want to solve: Eq. (27) fixes t_i and p_i while d moves, so it ignores the rate constraint (25b) and the fact that the payload N_i d scales with d. The interior optimum at d=4000 is therefore not established for the stated optimization problem. Bigger still, the empirical J_d(d) curve is four parameters fit to eight simulation points, with no error bars, no held-out settings, and no residual analysis. The stationarity condition depends on J'_d(d), and the claimed optimum sits on the steepest part of the sigmoid, so a small fit error shifts both d* and the 83.3% number. The paper itself says the coefficients depend on Acc_target and epsilon, which is an admission that the fitted instance used in optimization is not universal. The stress-test note lands.

Two smaller issues: the "globally optimal solution" claim for a nonconvex alternating scheme is unsupported; monotone decrease is all they show. And "five times faster" vs "3.5x more rounds" in the contributions are not equivalent—minor, but sloppy in a paper whose numbers matter.

Who is this for? Wireless-FL resource allocation folks and HDC researchers. The framework is legitimate and the math is mostly reproducible from the text, but there is no code or data, so the empirical curves are not independently checkable. A serious referee should engage it, with clear requests to fix the d-subproblem, add robustness and confidence information to the J_d fit, and soften the global optimality claim. I would not cite the 83.3% number in my own work yet.

Referee Report

3 major / 5 minor

Summary. The paper proposes FL-HDC-DP, a federated learning framework in which edge users train hyperdimensional computing (HDC) models with differential privacy and upload their class hypervectors to a base station over a wireless FDMA network. The authors formulate a joint optimization problem that minimizes total computation and transmission energy by choosing the HDC dimension, transmission time, bandwidth, transmit power, and CPU frequency, subject to a completion-time constraint, a privacy budget, and per-user hardware limits. Because the number of convergence rounds J_d depends on the dimension d, they introduce an empirical sigmoid-variant model J_d(d) = μ + ν/(1 + e^{β(log d − α)}), and they solve the original problem by alternating between a one-dimensional d-update and a convex resource-allocation subproblem. A feasibility initialization procedure is also given. Simulation results on MNIST show about 90% accuracy in roughly 3.5× fewer rounds than an NN baseline and an energy reduction of up to 83.3%, with an interior energy-minimizing dimension around d=4000.

Significance. If the claims hold, the paper makes a useful contribution: it is, to my knowledge, the first attempt to co-optimize the HDC model dimension together with wireless transmission resources under a DP constraint, and it provides a complete solvable formulation with closed-form updates. The core mathematical pieces that I checked are sound: Lemma 1's zCDP calibration is correct; Theorem 1 and Proposition 1's tight-rate arguments are valid; the convexity and monotonicity claims in Eqs. (40)–(41) are correct; and Proposition 2 gives a genuine feasibility certificate. The main weakness is that the central quantitative results—the interior optimum at d≈4000 and the 83.3% energy saving—are outputs of an optimization whose only learning-dependent input is an unvalidated four-parameter empirical fit (Eq. (26)). The paper also leaves the dependence of the DP noise calibration on the unknown total number of rounds implicit, which affects the privacy guarantee as stated. These issues are fixable with additional robustness analysis and a clarified privacy accounting, but they are load-bearing for the paper's headline claims.

major comments (3)

[Sec. V-A, Eq. (26); Sec. VII-D] The central quantitative claims—the interior optimum d≈4000, the energy profile in Fig. 6, and the 83.3% saving—are all computed from the four-parameter sigmoid J_d(d)=μ+ν/(1+e^{β(log d−α)}) fitted to the authors' own simulation outputs. The paper reports no residuals, confidence intervals, held-out settings, or robustness checks, and the functional form is a hand-chosen Ansatz. Since the energy objective (23) and the stationarity condition (29) depend on J_d and J'_d, any bias in α or β shifts the predicted optimum and the magnitude of the savings. The paper itself states that the coefficients depend on the selected Acc_target and ε, yet the optimization in Sec. VII-D uses a single fitted instance (μ=15.25, ν=99.99, α=7.80, β=5.69). Please add uncertainty quantification or at least a sensitivity analysis over the fit parameters, or explicitly reframe the headline numbers as conditional
[Sec. III-A, Lemma 1/Eq. (13); Sec. V-A/Eq. (26)] The per-round Gaussian noise in Lemma 1 is calibrated as σ_j = Δ_j√(J/(2ρ_max)), which depends on the total number of rounds J. But in the optimization, J is exactly the convergence-round count J_d(d) that varies with the dimension, and J_d is itself an output of the training dynamics under privacy noise. The paper does not state what J is used to pre-calibrate the noise in the experiments of Sec. VII-B or Sec. VII-D, nor does it explain how the privacy guarantee is maintained when the optimized J_d differs from the calibrated J. If the noise is calibrated for a fixed J_max larger than every possible J_d, that bound must be explicit and the fitted J_d(d) must be measured under that same noise. As written, the privacy accounting and the convergence model are mutually dependent, so the reported (ε,δ) guarantee for the optimized solution is not well-defined.
[Sec. V-B, Eqs. (27)–(29); Algorithm 2] The text claims that the alternating scheme 'efficiently converges to the globally optimal solution', but no global convergence proof is provided. Subproblem 1 minimizes a nonconvex function: J_d is a decreasing sigmoid, the derivative equation (29) may have multiple roots, a single bisection root is not a certificate of global optimality, and the 'boundary values determined by (27a)' are not characterized. Since the admissible set D_HDC is discrete and small (eight values in the simulations), a global solution can be obtained by enumerating d and solving the convex resource-allocation subproblem for each d. Please either prove the needed monotonicity/convexity or revise the algorithm and the global-optimality claim.

minor comments (5)

[Fig. 5] Axis labels contain typos: 'con erge' should be 'converge', 'true alue' should be 'true value', and 'fitting cur e' should be 'fitting curve'.
[Table II and Eq. (25)] The number of users is denoted U in the model but K in Table II and in constraint (25c). Please unify the notation.
[Sec. V-B, Eq. (35)] The closed form for f_i^* does not explicitly enforce the upper bound f_i ≤ f_i^max. Feasibility may be preserved by the initialization certificate, but the paper should state this or include the min(·, f_i^max) operation; otherwise constraint (34b) is not enforced in the closed form.
[Eqs. (17) and (23)] The payload is modeled as N_i d bits, i.e., one bit per dimension per class. After averaging and DP noise, class hypervectors are real-valued; the bit depth or quantization assumption should be stated, since it directly scales transmission energy.
[Sec. VII-D] The statement 'DP noise variance is calibrated in advance' should give the specific J used for calibration in the optimization setting. Without this, the reader cannot reproduce the simulation or verify the privacy guarantee.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the energy-optimal dimension and headline savings are conditional on an openly fitted empirical convergence-round curve, but the optimization itself is not a by-construction restatement of that fit.

full rationale

The paper's central derivation chain is self-contained: the energy model (Eq. 23), the joint optimization problem (Eq. 25), and the alternating resource-allocation algorithms are derived in-paper from explicit computation and communication models. The only learning-dependent input is the convergence-round function J_d(d) in Eq. (26), which the paper transparently introduces as an empirical sigmoid-variant model: 'Since deriving an analytical relationship among dimension, privacy, and convergence remains challenging, we develop an empirical model to precisely estimate J_d.' The coefficients are obtained by nonlinear least squares against the authors' own MNIST simulations and are explicitly reported in Sec. VII-D (mu=15.25, nu=99.99, alpha=7.80, beta=5.69). This is a modeling input rather than a disguised derivation: the paper does not claim Eq. (26) follows from first principles, and the energy-minimizing dimension d=4000 is not a by-construction identity of the fitted curve—it emerges from the full energy objective (23) and constraints, including computation costs, transmission costs, and the product R_i J_d d. The 83.3% energy saving is therefore conditional on the fitted curve being representative; if the fit is inaccurate, the quantitative results shift. That is a robustness/validation risk, not circularity. Self-citations ([4], [5], [11]) appear only as related-work context and do not carry the load of any theorem, lemma, or numerical claim. The convergence and accuracy comparisons against FL-NN-DP are direct simulation results rather than renamed outputs of the fitted model. No load-bearing step reduces, by the paper's own equations or by a self-citation chain, to an input that was already assumed.

Assumptions & free parameters 3 free parameters · 6 assumptions · 1 invented entities

The architecture of the paper is: a standard DP calibration lemma (self-contained), standard convex resource-allocation subproblems (self-contained), and a fitted empirical J_d(d) model that carries the entire dimension-optimization burden. The free parameters of consequence are the sigmoid coefficients, which are fitted to the paper's own simulation outputs; the simulation inputs (error rate, cycle counts) are assumed constants. No new physical entities are postulated; the only invented object is the empirical convergence model.

free parameters (3)

Sigmoid convergence-round coefficients (μ, ν, α, β) = μ=15.25, ν=99.99, α=7.80, β=5.69 for the (ε,δ)=(25,1e−5), Acc≥88% optimization setting; refit per setting
Eq. (26) is fitted by nonlinear least squares to simulated J_d vs d (Fig. 5); the optimal dimension d=4000 and the headline energy savings are conditional on these values.
Inference error rate e_i = 0.4 (Table II)
Fixed per-user misclassification ratio in the retraining energy model (Eq. 20); chosen, not measured per user.
Per-dimension CPU cycle costs (C_enc, C_agg, C_sim, C_up) = C_enc=C_agg=784, C_sim=100, C_up=8 (Table II)
Input constants to the energy model (Eqs. 18-21); not validated against hardware measurements.

assumptions (6)

domain assumption Encoded bipolar HVs are ℓ2-clipped to norm κ before aggregation and transmission (Assumption 1, Eq. 7).
The DP sensitivity bounds Δ1=κ (Eq. 9) and Δj=√2κ (Eq. 11) rest on it; κ is never given a numerical value in the simulation setup.
domain assumption Per-round HDC retraining sensitivity is √2κ even though update decisions depend on the data-dependent broadcast global AM (Eqs. 10-11).
Lemma 1 composes per-round zCDP; the paper does not discuss the adaptive-composition conditions under which conditioning on the broadcast model keeps the per-round sensitivity bound valid.
domain assumption Switched-capacitance energy model E = γ C D d f² with per-dimension cycle counts (Eqs. 19-21).
Standard surrogate from [26]; the headline energy numbers inherit its accuracy and the fixed values of γ and the cycle counts.
standard math FDMA Shannon-rate model r_i = b_i log2(1 + p_i g_i/(N0 b_i)) (Eq. 16).
Standard orthogonal uplink rate model used throughout the wireless-FL literature the paper cites.
standard math zCDP additivity and the (ε,δ) conversion ε = ρ + 2√(ρ ln(1/δ)) (Eq. 6, Lemma 1).
Matches Bun-Steinke zCDP; the derived ρ_max formula is algebraically verified.
ad hoc to paper Sigmoid-variant functional form J_d(d) = μ + ν/(1 + e^{β(log d − α)}) (Eq. 26).
Hand-chosen Ansatz fitted to simulated data; no derivation of this form, no justification beyond visual fit quality in Fig. 5.

invented entities (1)

Sigmoid-variant convergence-round model J_d(d) (Eq. 26)
purpose: Empirical link between HDC dimension and the required number of FL rounds; the coupling that makes dimension optimization possible
A four-parameter empirical model fit to the authors' own simulations; no falsifiable predictions outside the fitted settings are offered, so it has no independent evidential handle.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Energy Efficient Federated Learning with Hyperdimensional Computing over Wireless Communication Networks." pith.science (2026). https://pith.science/paper/4RL6OWLP

@misc{pith2026260221949,
  author       = {Pith},
  title        = {Pith review of: Energy Efficient Federated Learning with Hyperdimensional Computing over Wireless Communication Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4RL6OWLP}},
  note         = {Machine review of arXiv:2602.21949}
}

read the original abstract

In this paper, we investigate a problem of minimizing total energy consumption for secure federated learning (FL) over wireless edge networks. To address the high computational cost and privacy challenges in conventional FL with neural networks (NN) for resource-constrained users, we propose a novel FL with hyperdimensional computing and differential privacy (FL-HDC-DP) framework. In the considered model, each edge user employs hyperdimensional computing (HDC) for local training, which replaces complex neural updates with simple hypervector operations, and applies differential privacy (DP) noise to protect transmitted model information. We optimize the total energy of computation and communication under both latency and privacy constraints. We formulate the problem as an optimization that minimizes the total energy of all users by jointly allocating HDC dimension, transmission time, system bandwidth, transmit power, and CPU frequency. To solve this problem, a sigmoid-variant function is proposed to characterize the relationship between the HDC dimension and the convergence rounds required to reach a target accuracy. Based on this model, we develop two alternating optimization algorithms, where closed-form expressions for time, frequency, bandwidth, and power allocations are derived at each iteration. Since the iterative algorithm requires a feasible initialization, we construct a feasibility problem and obtain feasible initial resource parameters by solving a per round transmission time minimization problem. Simulation results demonstrate that the proposed FL-HDC-DP framework achieves up to 83.3% total energy reduction compared with the baseline, while attaining about 90% accuracy in approximately 3.5X fewer communication rounds than the NN baseline.

Figures

Figures reproduced from arXiv: 2602.21949 by the authors.

**Figure 2.** Illustration of the considered model for FL-HDC over [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Accuracy vs. Epochs of FL-HDC under IID data with [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Accuracy vs. epochs of FL-HDC-DP and FL-NN-DP under [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 5.** Figure 5: Fitting performance of the proposed sigmoid-variant model [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

**Figure 6.** Figure 6: Total energy vs. hypervector dimension under different [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: Total energy vs. maximum transmit power. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

**Figure 8.** Figure 8: Total energy vs. total bandwidth. 0-100 m 100-200 m 200-300 m 300-400 m 400-500 m Distance groups 0 0.05 0.1 0.15 0.2 Bandwidth b (MHz) 0-100 m 100-200 m 200-300 m 300-400 m 400-500 m Distance groups 0.8 0.85 0.9 0.95 1 Power p (mW) 0-100 m 100-200 m 200-300 m 300-400 …

**Figure 9.** Figure 9: Optimized average values of bandwidth, transmit power, [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

CRB-Guided Framework Design and Resource Allocation for Indoor mmWave ISCC Systems
cs.IT 2026-05 unverdicted novelty 5.0 of 10

CRB-guided joint optimization of sensing power and adaptive Mamba depth minimizes pose prediction error in indoor mmWave ISCC under latency and energy limits.

Reference graph

Works this paper leans on

28 extracted references · 1 linked inside Pith · cited by 1 Pith paper

[1]

Energy-eﬀicient feder- ated edge learning with streaming data: A lyapunov optimiza- tion approach,

C.-H. Hu, Z. Chen, and E. G. Larsson, “Energy-eﬀicient feder- ated edge learning with streaming data: A lyapunov optimiza- tion approach,” IEEE Transactions on Communications, vol. 73, no. 2, pp. 1142–1156, 2025

2025
[2]

Multi-stage hybrid federated learning over large-scale d2d-enabled fog networks,

S. Hosseinalipour, S. S. Azam, C. G. Brinton, N. Michelusi, V. Aggarwal, D. J. Love, and H. Dai, “Multi-stage hybrid federated learning over large-scale d2d-enabled fog networks,” IEEE/ACM Transactions on Networking, vol. 30, no. 4, pp. 1569–1584, 2022

2022
[3]

Federated edge network utility maximization for a multi-server system: Algorithm and convergence,

N. Karakoç, A. Scaglione, M. Reisslein, and R. Wu, “Federated edge network utility maximization for a multi-server system: Algorithm and convergence,” IEEE/ACM Transactions on Net- working, vol. 30, no. 5, pp. 2002–2017, 2022

2002
[4]

Distributed machine learning for uav swarms: Comput- ing, sensing, and semantics,

Y. Ding, Z. Yang, Q.-V. Pham, Y. Hu, Z. Zhang, and M. Shikh- Bahaei, “Distributed machine learning for uav swarms: Comput- ing, sensing, and semantics,” IEEE Internet of Things Journal, vol. 11, no. 5, pp. 7447–7473, 2024

2024
[5]

Joint layer selection and differential privacy design for feder- ated learning over wireless networks,

Y. Ding, W. Shang, Y. Yang, W. Ding, and M. Shikh-Bahaei, “Joint layer selection and differential privacy design for feder- ated learning over wireless networks,” IEEE Internet of Things Journal, vol. 11, no. 24, pp. 39 767–39 779, 2024

2024
[6]

Prive-hd: Privacy- preserved hyperdimensional computing,

B. Khaleghi, M. Imani, and T. Rosing, “Prive-hd: Privacy- preserved hyperdimensional computing,” in 2020 57th ACM/IEEE Design Automation Conference (DAC). IEEE, 2020, pp. 1–6

2020
[7]

Energy eﬀicient federated learning over wireless communica- tion networks,

Z. Yang, M. Chen, W. Saad, C. S. Hong, and M. Shikh-Bahaei, “Energy eﬀicient federated learning over wireless communica- tion networks,” IEEE Transactions on Wireless Communica- tions, vol. 20, no. 3, pp. 1935–1949, 2021

1935
[8]

Energy eﬀicient federated learning over heterogeneous mobile devices via joint design of weight quantization and wireless transmis- sion,

R. Chen, L. Li, K. Xue, C. Zhang, M. Pan, and Y. Fang, “Energy eﬀicient federated learning over heterogeneous mobile devices via joint design of weight quantization and wireless transmis- sion,” IEEE Transactions on Mobile Computing, vol. 22, no. 12, pp. 7451–7465, 2023

2023

Show all 28 references

[9]

Fl-hdc: Hyper- dimensional computing design for the application of federated learning,

C.-Y. Hsieh, Y.-C. Chuang, and A.-Y. A. Wu, “Fl-hdc: Hyper- dimensional computing design for the application of federated learning,” in Proc. IEEE Int. Conf. Artif. Intell. Circuits Syst. (AICAS). IEEE, 2021, pp. 1–5

2021
[10]

HyperFeel: An eﬀicient federated learning framework using hyperdimensional comput- ing,

H. Li, F. Liu, Y. Chen, and L. Jiang, “HyperFeel: An eﬀicient federated learning framework using hyperdimensional comput- ing,” in Proc. Asia South Pacific Des. Autom. Conf. (ASP- DAC), 2024, pp. 716–721

2024
[11]

Hyperdimensional computing empowered federated foundation model over wireless networks for meta- verse,

Y. Ding, W. Shang, M. Xu, Z. Yang, Y. Hu, D. Niyato, and M. Shikh-Bahaei, “Hyperdimensional computing empowered federated foundation model over wireless networks for meta- verse,” in 2024 IEEE Smart World Congress (SWC), 2024, pp. 2187–2194

2024
[12]

Private and eﬀicient learning with hyperdimensional computing,

B. Khaleghi, X. Yu, J. Kang, X. Wang, and T. Rosing, “Private and eﬀicient learning with hyperdimensional computing,” IEEE Transactions on Circuits and Systems for Artificial Intelligence, 2024

2024
[13]

Privacy- preserving federated learning with differentially private hyperdi- mensional computing,

F. J. Piran, Z. Chen, M. Imani, and F. Imani, “Privacy- preserving federated learning with differentially private hyperdi- mensional computing,” Computers and Electrical Engineering, vol. 123, p. 110261, 2025

2025
[14]

Hydrea: Utilizing hyperdimensional computing for a more robust and eﬀicient machine learning system,

J. Morris, K. Ergun, B. Khaleghi, M. Imani, B. Aksanli, and T. Simunic, “Hydrea: Utilizing hyperdimensional computing for a more robust and eﬀicient machine learning system,” ACM Trans. Embed. Comput. Syst., vol. 21, no. 6, pp. 1–25, 2022

2022
[15]

A highly energy-eﬀicient hyperdimensional computing processor for biosignal classification,

A. Menon, D. Sun, S. Sabouri, K. Lee, M. Aristio, H. Liew, and J. M. Rabaey, “A highly energy-eﬀicient hyperdimensional computing processor for biosignal classification,” IEEE Trans- actions on Biomedical Circuits and Systems, vol. 16, no. 4, pp. 524–534, 2022

2022
[16]

On hyperdi- mensional computing-based federated learning: A case study,

S. Zhang, D. Ma, S. Bian, L. Yang, and X. Jiao, “On hyperdi- mensional computing-based federated learning: A case study,” in Proc. Int. Joint Conf. Neural Netw.(IJCNN). IEEE, 2023, pp. 1–8

2023
[17]

A new pathway to integrated learning and commu- nication (ilac): Large ai model and hyperdimensional computing for communication,

W. Xu, Z. Yang, D. W. K. Ng, R. Schober, H. V. Poor, Z. Zhang, and X. You, “A new pathway to integrated learning and commu- nication (ilac): Large ai model and hyperdimensional computing for communication,” arXiv preprint arXiv:2506.18432, 2025

2025 arXiv
[18]

Hyperdimensional computing: An introduction to computing in distributed representation with high-dimensional random vectors,

P. Kanerva, “Hyperdimensional computing: An introduction to computing in distributed representation with high-dimensional random vectors,” Cognitive computation, vol. 1, pp. 139–159, 2009

2009
[19]

Classification using hyperdimensional computing: A review,

L. Ge and K. K. Parhi, “Classification using hyperdimensional computing: A review,” IEEE Circuits Sys. Mag., vol. 20, no. 2, pp. 30–47, 2020

2020
[20]

Hyperdimensional computing vs. neural networks: Comparing architecture and learning process,

D. Ma, C. Hao, and X. Jiao, “Hyperdimensional computing vs. neural networks: Comparing architecture and learning process,” in Proc. Int. Symp. Qual. Electron. Des. (ISQED). IEEE, 2024, pp. 1–5

2024
[21]

Hyperdimensional biosignal processing: A case study for EMG- based hand gesture recognition,

A. Rahimi, S. Benatti, P. Kanerva, L. Benini, and J. M. Rabaey, “Hyperdimensional biosignal processing: A case study for EMG- based hand gesture recognition,” in Proc. IEEE Int. Conf. Rebooting Comput. (ICRC). IEEE, 2016, pp. 1–8

2016
[22]

Calibrating noise to sensitivity in private data analysis,

C. Dwork, F. McSherry, K. Nissim, and A. Smith, “Calibrating noise to sensitivity in private data analysis,” in Theory of cryptography conference. Springer, 2006, pp. 265–284

2006
[23]

Concentrated differential privacy: Simplifications, extensions, and lower bounds,

M. Bun and T. Steinke, “Concentrated differential privacy: Simplifications, extensions, and lower bounds,” in Theory of cryptography conference. Springer, 2016, pp. 635–658

2016
[24]

Private and eﬀicient learning with hyperdimensional computing,

B. Khaleghi, X. Yu, J. Kang, X. Wang, and T. Rosing, “Private and eﬀicient learning with hyperdimensional computing,” IEEE Transactions on Circuits and Systems for Artificial Intelligence, vol. 1, no. 2, pp. 204–219, 2024

2024
[25]

Energy-eﬀicient resource allocation for mobile-edge computation offloading,

C. You, K. Huang, H. Chae, and B.-H. Kim, “Energy-eﬀicient resource allocation for mobile-edge computation offloading,” IEEE Transactions on Wireless Communications, vol. 16, no. 3, pp. 1397–1411, 2017

2017
[26]

Dynamic computation offloading for mobile-edge computing with energy harvesting devices,

Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloading for mobile-edge computing with energy harvesting devices,” IEEE Journal on Selected Areas in Communications, vol. 34, no. 12, pp. 3590–3605, 2016

2016
[27]

A tutorial on decomposition methods for network utility maximization,

D. Palomar and M. Chiang, “A tutorial on decomposition methods for network utility maximization,” IEEE Journal on Selected Areas in Communications, vol. 24, no. 8, pp. 1439– 1451, 2006

2006
[28]

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 Proc. Artif. Intell. Stat. (AIS- TATS). PMLR, 2017, pp. 1273–1282

2017

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Energy-eﬀicient feder- ated edge learning with streaming data: A lyapunov optimiza- tion approach,

C.-H. Hu, Z. Chen, and E. G. Larsson, “Energy-eﬀicient feder- ated edge learning with streaming data: A lyapunov optimiza- tion approach,” IEEE Transactions on Communications, vol. 73, no. 2, pp. 1142–1156, 2025

2025

[2] [2]

Multi-stage hybrid federated learning over large-scale d2d-enabled fog networks,

S. Hosseinalipour, S. S. Azam, C. G. Brinton, N. Michelusi, V. Aggarwal, D. J. Love, and H. Dai, “Multi-stage hybrid federated learning over large-scale d2d-enabled fog networks,” IEEE/ACM Transactions on Networking, vol. 30, no. 4, pp. 1569–1584, 2022

2022

[3] [3]

Federated edge network utility maximization for a multi-server system: Algorithm and convergence,

N. Karakoç, A. Scaglione, M. Reisslein, and R. Wu, “Federated edge network utility maximization for a multi-server system: Algorithm and convergence,” IEEE/ACM Transactions on Net- working, vol. 30, no. 5, pp. 2002–2017, 2022

2002

[4] [4]

Distributed machine learning for uav swarms: Comput- ing, sensing, and semantics,

Y. Ding, Z. Yang, Q.-V. Pham, Y. Hu, Z. Zhang, and M. Shikh- Bahaei, “Distributed machine learning for uav swarms: Comput- ing, sensing, and semantics,” IEEE Internet of Things Journal, vol. 11, no. 5, pp. 7447–7473, 2024

2024

[5] [5]

Joint layer selection and differential privacy design for feder- ated learning over wireless networks,

Y. Ding, W. Shang, Y. Yang, W. Ding, and M. Shikh-Bahaei, “Joint layer selection and differential privacy design for feder- ated learning over wireless networks,” IEEE Internet of Things Journal, vol. 11, no. 24, pp. 39 767–39 779, 2024

2024

[6] [6]

Prive-hd: Privacy- preserved hyperdimensional computing,

B. Khaleghi, M. Imani, and T. Rosing, “Prive-hd: Privacy- preserved hyperdimensional computing,” in 2020 57th ACM/IEEE Design Automation Conference (DAC). IEEE, 2020, pp. 1–6

2020

[7] [7]

Energy eﬀicient federated learning over wireless communica- tion networks,

Z. Yang, M. Chen, W. Saad, C. S. Hong, and M. Shikh-Bahaei, “Energy eﬀicient federated learning over wireless communica- tion networks,” IEEE Transactions on Wireless Communica- tions, vol. 20, no. 3, pp. 1935–1949, 2021

1935

[8] [8]

Energy eﬀicient federated learning over heterogeneous mobile devices via joint design of weight quantization and wireless transmis- sion,

R. Chen, L. Li, K. Xue, C. Zhang, M. Pan, and Y. Fang, “Energy eﬀicient federated learning over heterogeneous mobile devices via joint design of weight quantization and wireless transmis- sion,” IEEE Transactions on Mobile Computing, vol. 22, no. 12, pp. 7451–7465, 2023

2023