Sheaf-Based Decentralized Multimodal Learning for Next-Generation Wireless Communication Systems

Abdulmomen Ghalkha; Chaouki Ben Issaid; Mehdi Bennis; Zhuojun Tian

REVIEW 4 major objections 6 minor 27 references

Sheaf-Based Decentralized Multimodal Learning for Next-Generation Wireless Communication Systems

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that decentralized multimodal federated learning converges to a stationary point at O(1/R) when task relationships are modeled by a learnable sheaf structure, and shows the resulting algorithms outperform decentralized…

desk verdict The paper's advertised O(1/R) convergence guarantee does not follow from its own proof, and the empirical work, while genuine, is under-reported. read the letter →

arxiv 2506.22374 v1 pith:SAY5TUKU submitted 2025-06-27 cs.LG cs.AI

Abdulmomen Ghalkha , Zhuojun Tian , Chaouki Ben Issaid , Mehdi Bennis This is my paper · ORCID

classification cs.LGcs.AI

keywords decentralizedlearningmultimodalfederatedsheaftheoryLaplacianattention-basedfusionconvergenceanalysislinkblockagepredictionmmWavebeamforming

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 devices with different sensor modalities can train a shared model together over a peer-to-peer network, without a central server, and that the training is provably convergent. It places a sheaf structure—vector spaces on communication links with learnable projections between clients—on top of each client's task-specific layer, so clients with different modality combinations learn how much to align with each other. The main formal claim is Theorem 1: under smoothness, boundedness, and connectivity assumptions, the attention-based variant Sheaf-DMFL-Att reaches a stationary point of a global objective at rate $O(1/R)$. If true, this gives decentralized multimodal federated learning the same style of convergence guarantee already available for centralized and single-modality decentralized methods, which matters in wireless systems where a parameter server is a single point of failure.

What carries the argument

The load-bearing object is a cellular sheaf placed on the communication graph: each client's task-specific head $\omega_i$ is a stalk over a node, each edge carries a lower-dimensional comparison space, and learnable restriction maps $P_{ij}$ project $\omega_i$ and $P_{ji}\omega_j$ into that space. The sheaf Laplacian regularizer $\frac{\lambda}{2}\sum_{(i,j)\in E}\|P_{ij}\omega_i - P_{ji}\omega_j\|^2$ penalizes disagreement between neighboring tasks after projection, so the system learns not only the models but also how tasks should be compared. Around this sits the partially shared architecture: modality encoders are averaged across clients using Metropolis-Hastings mixing matrices $W_k$, and attention weights $\alpha_{i,k}$ fuse modalities locally. The convergence proof's key device is the modified parameter vector $\tilde{\theta}_i^r$ in which shared encoders are replaced by their network averages, which decouples the encoder consensus dynamics (Lemma 1) from the head and attention updates (Lemma 2) and lets the whole system telescope into the $O(1/R)$ bound.

What would settle it

Run Sheaf-DMFL-Att on a network where one modality is split across two disconnected clusters, with all other assumptions satisfied, and measure the average squared gradient norm of $\Psi$ over $R$ rounds: if the bound still decays at $O(1/R)$, the connectivity assumption is not necessary; if the encoders drift apart or the bound fails, the assumption is doing the work.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that multimodal heterogeneity in a decentralized network can be modeled as multi-task learning with a learnable task-relationship structure, and that the resulting algorithm carries a worst-case convergence guarantee. Sheaf-DMFL-Att trains shared modality encoders by gossip averaging over modality-specific subgraphs, fuses the encoders' outputs through per-client attention weights, and then aligns the task-specific heads through a sheaf Laplacian regularizer. The proof tracks a modified global parameter vector in which each client uses the average encoder for each modality rather than its local encoder; Lemma 1 shows these averages move like gradient descent because the mixing matrices are doubly stochastic, and Lemma 2 turns L-smoothness into a one-step descent inequality. Theorem 1 then yields $\frac{1}{R}\sum_{r=0}^{R-1}\|\nabla\Psi(\tilde{\theta}^r,P^r)\|^2 \le \frac{\Psi(\tilde{\theta}^0,P^0)-\Psi^*}{\rho R}$, i.e. the average squared gradient norm of the global objective shrinks at the standard $O(1/R)$ non-convex rate.

Load-bearing premise

The load-bearing premise is Assumption 2: for every modality, the clients that possess it must form a connected subgraph of the communication network, because the convergence proof needs a strictly positive spectral gap for each modality-specific mixing matrix.

Editorial extensions

If this is right

Under the stated assumptions, Sheaf-DMFL-Att converges to a stationary point of the global objective at rate $O(1/R)$, matching the standard non-convex decentralized SGD rate.
No central parameter server is required; training runs over peer-to-peer links, with each client exchanging only projected head parameters and encoder aggregates rather than raw data.
The attention mechanism gives a concrete fix for negative transfer: in the beam prediction task, GPS-only clients improve when multimodal neighbors use attention-weighted fusion, because the gradients they receive are no longer dominated by a single strong modality.
The experiments report faster convergence and higher accuracy than DSGD, DMML-KD, and local-only training on blockage prediction and mmWave beamforming, approaching the centralized upper bound.
The design is tunable: larger projection dimension $\gamma$ and structured initialization of the restriction maps improve accuracy at the cost of memory, so the framework admits a memory-accuracy trade-off.

Reading between the lines

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

The convergence proof is modular: the sheaf regularizer only touches the task-specific heads, so the $O(1/R)$ bound likely survives if the attention layer is replaced by any local fusion mechanism that keeps the loss L-smooth and the attention parameters bounded.
Because Assumption 2 requires each modality's client subgraph to be connected, deployments where a sensor type appears in isolated clusters should either split that modality into separate tasks or add communication links before applying this algorithm.
The theorem's bound is stated for the averaged-encoder objective $\Psi(\tilde{\theta},P)$, not for the actual local encoders each client deploys; measuring the gap between the averaged model's stationary point and per-client personalized performance would be a natural next step.
The learned attention weights $\alpha_{i,k}$ could be read as online estimates of modality reliability: if one sensor is occluded or noisy, the weights should shift toward the reliable modality, and tracking that shift could turn into a sensor-quality monitoring tool.

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

You should know two things. First, the core theoretical claim, Theorem 1, is not established by the proof as written. Second, the empirical evaluation on two real 6G datasets is genuine but lacks the reporting standard that would make the performance claims fully credible.

What is new: the paper applies the sheaf-Laplacian regularizer from Ben Issaid et al. to task-specific heads in decentralized multimodal federated learning, adds a partially shared encoder architecture and a local attention fusion variant. That combination is not in the cited prior work. The attention variant is a sensible extension, and the experiments on DeepSense blockage prediction and drone beam prediction show consistent gains over DSGD, a distillation baseline, and local training, with an ablation of projection compression factor and initialization. The problem is well motivated and the paper is clearly written.

The soft spots are real and load-bearing. In Eq. (42) the P-gradient term is evaluated at (θ̃^{r+1}, P^r), but the final bound (44) claims a norm at (θ̃^r, P^r). There is no Lipschitz bound on ∇_P Ψ with respect to θ, so that index shift cannot be absorbed. Equation (33) identifies the global gradient ∇_ω Ψ(θ̃^r, P^r) with the actual local update (11), which uses gradients at θ^r before encoder averaging; that requires a consensus-error bound that the proof never provides. Lemma 2 says 'the detailed derivation is omitted due to space limitations,' which is exactly the step that would need to be checked. These are not presentation issues; they are missing steps in the central advertised contribution.

The experimental reporting also falls short: accuracy curves are single-seed with no error bars (aside from Table II), no code is provided, and the baseline name switches between DMML-KL and DMML-KD. Assumption 2, connectivity of each modality subgraph, is untested because the experimental graphs connect modality groups by construction.

That said, this is not a paper to ignore. The problem setup is legitimate, the combination is new, and the real-data evaluation is a serious attempt. My recommendation: send it to a serious referee, but instruct them to focus on the convergence proof. As written, Theorem 1 does not go through. If the authors can repair the proof or honestly weaken the claim, the paper is salvageable.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes Sheaf-DMFL and Sheaf-DMFL-Att, decentralized multimodal federated learning algorithms in which clients with heterogeneous modality sets share per-modality feature encoders via gossip aggregation and couple their task-specific heads through a learnable cellular-sheaf Laplacian regularizer. The central theoretical claim (Theorem 1) is an O(1/R) bound on the average squared gradient norm of the global objective Ψ under smoothness and connectedness assumptions. The experimental section evaluates the methods on link blockage prediction and mmWave beamforming benchmarks, reporting accuracy improvements over DSGD, DMML-KD, and local training baselines.

Significance. If the convergence guarantee were established, the paper would make a useful contribution by extending sheaf-based multi-task learning to decentralized multimodal settings with partial modality overlap. The problem is well motivated and the empirical comparisons on real-world datasets are a strength. However, the advertised theoretical guarantee, which is the paper's main novelty beyond the earlier conference version [22], is not proven: the proof of Theorem 1 contains an invalid identification of gradient terms, an unproved and generally false matrix inequality, an index mismatch in the final telescoping argument, and the central Lemma 2 is asserted with an omitted proof. These are load-bearing defects, not presentation issues. The empirical evaluation also lacks multi-seed statistics for the main accuracy curves and the code is not released, which limits the confidence one can place in the reported gains.

major comments (4)

[Appendix B, Eq. (33)] The equality ⟨∇ωΨ(θ̃^r, P^r), ω^{r+1}−ω^r⟩ = −α‖∇ωΨ(θ̃^r, P^r)‖² is not justified by the update rule (29), which uses ∇f(ω^r) evaluated at the un-averaged local parameters. Identifying ∇f(ω^r) with ∇ωΨ(θ̃^r, P^r) requires a consensus-error bound between θ^r and θ̃^r that is never stated or proved. Consequently the descent inequality (35) does not follow.
[Appendix A, Lemma 2] The proof of Lemma 2 is explicitly truncated ('The detailed derivation is omitted due to space limitations'). This lemma is the foundation of the theorem, since it gives the descent inequality (20) for the modified parameter vector. The sketch in (27)–(28) does not control the difference between the local gradients ∇_{ϕ_i,k} f_i(θ_i^r) used in the updates (15)–(16) and the gradients ∇_{ϕ̄_k} f(θ̃^r) that appear in the bound, so the lemma is unsubstantiated.
[Appendix B, Eq. (37)] Inequality (37) asserts that zeroing out entries of the matrix M = P^r − ηλP^rω^{r+1}(ω^{r+1})^T through the Hadamard product with H yields a Gram matrix dominated in the PSD order by M^T M. This is false in general: entrywise masking can increase the largest eigenvalue of the Gram matrix. Because this inequality is used to obtain (41), the claimed descent for the restriction-map update is not established.
[Appendix B, Eqs. (42)–(44)] The telescoping argument exhibits an index mismatch: the negative P-gradient term in (42) and (43) is evaluated at (θ̃^{r+1}, P^r), while the final averaged gradient norm in (44) is defined with ‖∇PΨ(θ̃^r, P^r)‖²_F. No Lipschitz bound on ∇PΨ with respect to θ is stated, so the two terms cannot be equated or bounded by one another. Thus Theorem 1's O(1/R) claim does not follow from the proof.

minor comments (6)

[Section V] The baseline is introduced as 'DMML-KL [27]' but Figures 5, 7, 9 and Table I label it 'DMML-KD'; the naming should be unified.
[Section V-A and V-B] Figures 5 and 7 show single-run accuracy curves without error bars or a statement about the number of seeds; Table II reports means and standard deviations, so the same reporting should be used for the main results.
[Section V-A and V-B] The hyperparameters α, ηφ, ηβ, η, λ, and the attention parameters are not listed in the experimental settings; without these values the experiments are difficult to reproduce.
[Appendix B, Eqs. (42) and (44)] In Eq. (42) the coefficient of the P-gradient descent term contains ηλN D_ω²/2, whereas the definition of ρ in Eq. (44) uses ηλD_ω²/2; this inconsistency should be resolved.
[Section III-B] The statement after Eq. (8) defines restriction maps P_{ij} with dimensions R^{d_{ij}×d_i}, but the compression factor γ in Section III-B gives dimensions ⌊γ·(d_i+d_j)/2⌋×d_i; the relation between the two definitions should be clarified.
[Assumption 2 and Section V] Assumption 2 requires each modality subgraph to be connected; the graphs in Figures 4 and 6 satisfy this by construction, so the experiments do not probe the behavior of the algorithm when this assumption fails. The paper should at least comment on this limitation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the convergence theorem is a genuine (though gappy) derivation from stated assumptions, and self-citations are provenance rather than load-bearing inputs.

full rationale

I find no constructional circularity in the claimed derivation chain. Theorem 1 (Section IV, Eq. 22) is stated to follow from Assumptions 1-4 via a descent argument built on the paper's own update rules (11), (12), (15)-(17); the bound on the average squared gradient norm of Ψ is not assumed in the hypotheses and is not taken as a conclusion from refs. [20] or [25]. The sheaf Laplacian and the task-specific update rule are indeed imported from the authors' prior work ([20], and 'as follows [20]' in Section III-B), and the authors note that their convergence analysis 'is similar to that in [25]' (Section IV), but these citations are used as provenance and design lineage, not as the evidence that establishes Theorem 1. The empirical evaluation is run on external DeepSense datasets against decentralized and local baselines, so the reported accuracy improvements are not tautological. I do flag two weaknesses located in the manuscript, but they are correctness risks rather than circularity: (i) Appendix A says 'The detailed derivation is omitted due to space limitations' immediately after Lemma 2, even though Lemma 2 is the core descent inequality feeding Theorem 1; (ii) in Appendix B, the P-gradient term is bounded at (θ̃^{r+1}, P^r) in Eqs. (42)-(43), while the final statement (44) and the definition of ‖∇Ψ(θ̃^r, P^r)‖² use (θ̃^r, P^r), with no stated Lipschitz bound on ∇_P Ψ with respect to θ to bridge the mismatch. These are missing steps in the written proof, not a reduction of the theorem's conclusion to its own inputs, so they do not raise the circularity score.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claims rest on standard smoothness and boundedness assumptions, a connectivity requirement on each modality subgraph, and a matrix inequality in the convergence proof that is not justified. The sheaf machinery is imported from prior work with overlapping authorship; no new physical entities are introduced.

free parameters (4)

lambda (sheaf regularization strength) = not reported
Controls how strongly task-specific heads are pulled toward neighbors; appears in Algorithm 1 and Eq (10). The reported gains depend on its choice.
gamma (projection compression factor) = 0.1 and 0.23 (ablated)
Sets the row dimension of restriction maps P_ij; Table II shows it changes accuracy by up to about 2 points.
Step sizes alpha, eta_phi, eta_beta, eta = not reported
Learning rates in updates (11), (15)-(17); Theorem 1 requires bounds on them but the paper does not give the values used in experiments.
Projection initialization spread sigma^2 = not reported
Random init P^R_ij is drawn from N(0, sigma^2); Table II shows structured vs random init changes accuracy.

assumptions (6)

domain assumption Assumption 1: each modality mixing matrix W_k is doubly stochastic.
Required for consensus averaging in Eq (6) and Lemma 1; not guaranteed by any protocol in the paper.
domain assumption Assumption 2: each modality subgraph G_k is connected.
Needed for spectral gap > 0; the weakest assumption on network topology and untested on disconnected modality groups.
domain assumption Assumption 3: each local loss f_i is differentiable and L-smooth.
Standard in nonconvex optimization; used throughout Lemma 2 and Theorem 1.
domain assumption Assumption 4: parameter vectors are uniformly bounded by D_omega, D_phi_k, D_beta_k.
Necessary for the restriction-map step size condition eta < 2/(lambda N D_omega^2).
standard math Cellular sheaf and sheaf Laplacian formalism from [20].
Definitions in Section II-B are taken from algebraic topology; no proof is given, but they are prior definitions.
ad hoc to paper Unproved matrix inequality (P^{r+1})^T P^{r+1} <= (P^r - eta lambda P^r omega^{r+1} (omega^{r+1})^T)^T (...).
Used to bound the descent of Psi in the P-update; follows neither from the Hadamard mask H nor from standard PSD calculus as written.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Sheaf-Based Decentralized Multimodal Learning for Next-Generation Wireless Communication Systems." pith.science (2026). https://pith.science/paper/SAY5TUKU

@misc{pith2026250622374,
  author       = {Pith},
  title        = {Pith review of: Sheaf-Based Decentralized Multimodal Learning for Next-Generation Wireless Communication Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SAY5TUKU}},
  note         = {Machine review of arXiv:2506.22374}
}

read the original abstract

In large-scale communication systems, increasingly complex scenarios require more intelligent collaboration among edge devices collecting various multimodal sensory data to achieve a more comprehensive understanding of the environment and improve decision-making accuracy. However, conventional federated learning (FL) algorithms typically consider unimodal datasets, require identical model architectures, and fail to leverage the rich information embedded in multimodal data, limiting their applicability to real-world scenarios with diverse modalities and varying client capabilities. To address this issue, we propose Sheaf-DMFL, a novel decentralized multimodal learning framework leveraging sheaf theory to enhance collaboration among devices with diverse modalities. Specifically, each client has a set of local feature encoders for its different modalities, whose outputs are concatenated before passing through a task-specific layer. While encoders for the same modality are trained collaboratively across clients, we capture the intrinsic correlations among clients' task-specific layers using a sheaf-based structure. To further enhance learning capability, we propose an enhanced algorithm named Sheaf-DMFL-Att, which tailors the attention mechanism within each client to capture correlations among different modalities. A rigorous convergence analysis of Sheaf-DMFL-Att is provided, establishing its theoretical guarantees. Extensive simulations are conducted on real-world link blockage prediction and mmWave beamforming scenarios, demonstrate the superiority of the proposed algorithms in such heterogeneous wireless communication systems.

Figures

Figures reproduced from arXiv: 2506.22374 by the authors.

**Figure 2.** Illustration of a cellular sheaf over a communication [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** An illustration of three multi-modal BSs and their model architectures: The proposed sheaf-enabled decentralized [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Client groups (BSs) with available modalities over a [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Test accuracy as a function of the number of communication rounds for different modality combinations: (a) LiDAR [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Client groups (BSs) with available modalities over a [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Test accuracy as a function of the number of communication rounds for different modality combinations: (a) GPS only, [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

**Figure 8.** Figure 8: Camera views of clients 2 (a), 3 (b), and 6 (c) for the [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

**Figure 9.** Figure 9: Comparison of test accuracy for different modality [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 25 canonical work pages

[22]

A sheaf-theoretic ap- proach to decentralized multimodal federated learning for next-generation communication systems,

A. Ghalkha, Z. Tian, C. B. Issaid, and M. Bennis, “A sheaf-theoretic ap- proach to decentralized multimodal federated learning for next-generation communication systems,” in PIMRC, Istanbul, September 2025, accepted

work page 2025
[1]

A vision of 6g wireless systems: Applications, trends, technologies, and open research problems,

W. Saad, M. Bennis, and M. Chen, “A vision of 6g wireless systems: Applications, trends, technologies, and open research problems,” IEEE network, vol. 34, no. 3, pp. 134–142, 2019

2019
[2]

Proactively predicting dynamic 6G link blockages using LiDAR and in-band signatures,

S. Wu, C. Chakrabarti, and A. Alkhateeb, “Proactively predicting dynamic 6G link blockages using LiDAR and in-band signatures,” IEEE Open Journal of the Commun. Society , vol. 4, pp. 392–412, 2023

work page 2023
[3]

Proactive received power pre- diction using machine learning and depth images for mmwave networks,

T. Nishio, H. Okamoto, K. Nakashima, Y . Koda, K. Yamamoto, M. Morikura, Y . Asai, and R. Miyatake, “Proactive received power pre- diction using machine learning and depth images for mmwave networks,” IEEE Journal on Selected Areas in Commun. , vol. 37, no. 11, pp. 2413– 2427, 2019

work page 2019
[4]

Millimeter-wave networking in the sky: A machine learning and mean field game approach for joint beamforming and beam-steering,

L. Li, H. Ren, Q. Cheng, K. Xue, W. Chen, M. Debbah, and Z. Han, “Millimeter-wave networking in the sky: A machine learning and mean field game approach for joint beamforming and beam-steering,” IEEE Trans. on Wireless Commun. , vol. 19, no. 10, pp. 6393–6408, 2020

work page 2020
[5]

Hybrid beamforming/combining for millimeter wave mimo: A machine learning approach,

J. Chen, W. Feng, J. Xing, P. Yang, G. E. Sobelman, D. Lin, and S. Li, “Hybrid beamforming/combining for millimeter wave mimo: A machine learning approach,” IEEE Trans. Vehicular Technology , vol. 69, no. 10, pp. 11 353–11 368, 2020

work page 2020
[6]

Towards real-world 6g drone communication: Position and camera aided beam prediction,

G. Charan, A. Hredzak, C. Stoddard, B. Berrey, M. Seth, H. Nunez, and A. Alkhateeb, “Towards real-world 6g drone communication: Position and camera aided beam prediction,” in GLOBECOM 2022-2022 IEEE Global Commun. Conference . IEEE, 2022, pp. 2951–2956

work page 2022
[7]

Deep multimodal learning: Merging sensory data for massive mimo channel prediction,

Y . Yang, F. Gao, C. Xing, J. An, and A. Alkhateeb, “Deep multimodal learning: Merging sensory data for massive mimo channel prediction,” IEEE Journal on Selected Areas in Commun. , vol. 39, no. 7, pp. 1885– 1898, 2020

work page 2020

Show all 27 references

[8]

Din: A decentral- ized inexact newton algorithm for consensus optimization,

A. Ghalkha, C. B. Issaid, A. Elgabli, and M. Bennis, “Din: A decentral- ized inexact newton algorithm for consensus optimization,” IEEE Trans. Machine Learning in Commun. and Networking , 2024

2024
[9]

Scalable and resource- efficient second-order federated learning via over-the-air aggregation,

A. Ghalkha, C. B. Issaid, and M. Bennis, “Scalable and resource- efficient second-order federated learning via over-the-air aggregation,” IEEE Wireless Commun. Letters , 2024

2024
[10]

Distributed machine learning based downlink channel estimation for ris assisted wireless communications,

L. Dai and X. Wei, “Distributed machine learning based downlink channel estimation for ris assisted wireless communications,” IEEE Trans. Commun., vol. 70, no. 7, pp. 4900–4909, 2022

2022
[11]

Federated learning for wireless communications: Motivation, opportunities, and challenges,

S. Niknam, H. S. Dhillon, and J. H. Reed, “Federated learning for wireless communications: Motivation, opportunities, and challenges,” IEEE Commun. Magazine , vol. 58, no. 6, pp. 46–51, 2020

2020
[12]

Distributed machine learning for uav swarms: Computing, sensing, and semantics,

Y . Ding, Z. Yang, Q.-V . Pham, Y . Hu, Z. Zhang, and M. Shikh-Bahaei, “Distributed machine learning for uav swarms: Computing, sensing, and semantics,” IEEE Internet of Things Journal , vol. 11, no. 5, pp. 7447– 7473, 2023

2023
[13]

Tackling modality- heterogeneous client drift holistically for heterogeneous multimodal fed- erated learning,

H. Song, J. Wang, J. Zhou, and L. Wang, “Tackling modality- heterogeneous client drift holistically for heterogeneous multimodal fed- erated learning,” IEEE Trans. Medical Imaging , 2024

2024
[14]

Multi-modality sensing in mmwave beamforming for connected vehicles using deep learning,

M. B. Mollah, H. Wang, M. A. Karim, and H. Fang, “Multi-modality sensing in mmwave beamforming for connected vehicles using deep learning,” IEEE Trans. Cognitive Commun. and Networking , 2025

2025
[15]

Wireless interference recognition with multimodal learning,

P. Wang, K. Ma, Y . Bai, C. Sun, Z. Wang, and S. Chen, “Wireless interference recognition with multimodal learning,” IEEE Trans. Wireless Commun., 2024

2024
[16]

Multimodal feder- ated learning via contrastive representation ensemble,

Q. Yu, Y . Liu, Y . Wang, K. Xu, and J. Liu, “Multimodal feder- ated learning via contrastive representation ensemble,” arXiv preprint arXiv:2302.08888, 2023

2023 arXiv
[17]

Harmony: Heterogeneous multi-modal federated learning through disentangled model training,

X. Ouyang, Z. Xie, H. Fu, S. Cheng, L. Pan, N. Ling, G. Xing, J. Zhou, and J. Huang, “Harmony: Heterogeneous multi-modal federated learning through disentangled model training,” in Proc. 21st Annual International Conference on Mobile Systems, Applications and Services, 2023, p...

2023
[18]

Towards optimal multi-modal federated learning on non-iid data with hierarchical gradient blending,

S. Chen and B. Li, “Towards optimal multi-modal federated learning on non-iid data with hierarchical gradient blending,” in IEEE INFOCOM 2022-IEEE conference on computer Commun. . IEEE, 2022, pp. 1469– 1478

2022
[19]

Fedmsplit: Correlation-adaptive federated multi- task learning across multimodal split networks,

J. Chen and A. Zhang, “Fedmsplit: Correlation-adaptive federated multi- task learning across multimodal split networks,” in Proc. 28th ACM SIGKDD conference on knowledge discovery and data mining , 2022, pp. 87–96

2022
[20]

Tackling feature and sample heterogeneity in decentralized multi-task learning: A sheaf- theoretic approach,

C. Ben Issaid, P. Vepakomma, and M. Bennis, “Tackling feature and sample heterogeneity in decentralized multi-task learning: A sheaf- theoretic approach,” Trans. Machine Learning Research , 2025

2025
[21]

Robinson, Topological signal processing

M. Robinson, Topological signal processing. Springer, 2014, vol. 81

2014
[23]

Federated multi-task learning,

V . Smith, C.-K. Chiang, M. Sanjabi, and A. S. Talwalkar, “Federated multi-task learning,” Proc. Adv. Neural Inf. Process. Syst. (NIPS), vol. 30, 2017

2017
[24]

Exploiting shared representations for personalized federated learning,

L. Collins, H. Hassani, A. Mokhtari, and S. Shakkottai, “Exploiting shared representations for personalized federated learning,” in Interna- tional conference on machine learning . PMLR, 2021, pp. 2089–2099

2021
[25]

Distributed learning over networks with graph-attention-based personalization,

Z. Tian, Z. Zhang, Z. Yang, R. Jin, and H. Dai, “Distributed learning over networks with graph-attention-based personalization,” IEEE Trans. Signal Processing, vol. 71, pp. 2071–2086, 2023

2023
[26]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Proc. Adv. Neural Inf. Process. Syst. (NIPS) , vol. 30, 2017

2017
[27]

Knowledge distillation and training bal- ance for heterogeneous decentralized multi-modal learning over wireless networks,

B. Yin, Z. Chen, and M. Tao, “Knowledge distillation and training bal- ance for heterogeneous decentralized multi-modal learning over wireless networks,” IEEE Trans. Mobile Computing , vol. 23, no. 10, pp. 9629– 9644, 2024

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [22]

A sheaf-theoretic ap- proach to decentralized multimodal federated learning for next-generation communication systems,

A. Ghalkha, Z. Tian, C. B. Issaid, and M. Bennis, “A sheaf-theoretic ap- proach to decentralized multimodal federated learning for next-generation communication systems,” in PIMRC, Istanbul, September 2025, accepted

work page 2025

[2] [1]

A vision of 6g wireless systems: Applications, trends, technologies, and open research problems,

W. Saad, M. Bennis, and M. Chen, “A vision of 6g wireless systems: Applications, trends, technologies, and open research problems,” IEEE network, vol. 34, no. 3, pp. 134–142, 2019

2019

[3] [2]

Proactively predicting dynamic 6G link blockages using LiDAR and in-band signatures,

S. Wu, C. Chakrabarti, and A. Alkhateeb, “Proactively predicting dynamic 6G link blockages using LiDAR and in-band signatures,” IEEE Open Journal of the Commun. Society , vol. 4, pp. 392–412, 2023

work page 2023

[4] [3]

Proactive received power pre- diction using machine learning and depth images for mmwave networks,

T. Nishio, H. Okamoto, K. Nakashima, Y . Koda, K. Yamamoto, M. Morikura, Y . Asai, and R. Miyatake, “Proactive received power pre- diction using machine learning and depth images for mmwave networks,” IEEE Journal on Selected Areas in Commun. , vol. 37, no. 11, pp. 2413– 2427, 2019

work page 2019

[5] [4]

Millimeter-wave networking in the sky: A machine learning and mean field game approach for joint beamforming and beam-steering,

L. Li, H. Ren, Q. Cheng, K. Xue, W. Chen, M. Debbah, and Z. Han, “Millimeter-wave networking in the sky: A machine learning and mean field game approach for joint beamforming and beam-steering,” IEEE Trans. on Wireless Commun. , vol. 19, no. 10, pp. 6393–6408, 2020

work page 2020

[6] [5]

Hybrid beamforming/combining for millimeter wave mimo: A machine learning approach,

J. Chen, W. Feng, J. Xing, P. Yang, G. E. Sobelman, D. Lin, and S. Li, “Hybrid beamforming/combining for millimeter wave mimo: A machine learning approach,” IEEE Trans. Vehicular Technology , vol. 69, no. 10, pp. 11 353–11 368, 2020

work page 2020

[7] [6]

Towards real-world 6g drone communication: Position and camera aided beam prediction,

G. Charan, A. Hredzak, C. Stoddard, B. Berrey, M. Seth, H. Nunez, and A. Alkhateeb, “Towards real-world 6g drone communication: Position and camera aided beam prediction,” in GLOBECOM 2022-2022 IEEE Global Commun. Conference . IEEE, 2022, pp. 2951–2956

work page 2022

[8] [7]

Deep multimodal learning: Merging sensory data for massive mimo channel prediction,

Y . Yang, F. Gao, C. Xing, J. An, and A. Alkhateeb, “Deep multimodal learning: Merging sensory data for massive mimo channel prediction,” IEEE Journal on Selected Areas in Commun. , vol. 39, no. 7, pp. 1885– 1898, 2020

work page 2020