Latent Space Alignment for AI-Native MIMO Semantic Communications

Emilio Calvanese Strinati; Mario Edoardo Pandolfo; Paolo Di Lorenzo; Simone Fiorellino

REVIEW 5 major objections 5 minor 39 references

Latent Space Alignment for AI-Native MIMO Semantic Communications

T0 review · 5 major / 5 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 Jointly optimizing a semantic MIMO precoder and decoder to align mismatched latent spaces while equalizing the wireless channel outperforms the disjoint approach, with the linear method ahead by about ten accuracy points at small…

desk verdict New problem formulation with sound derivations, but the headline joint-vs-disjoint comparison is confounded because the baseline alignment is trained on clean latent pairs while the proposed methods train through the channel. read the letter →

arxiv 2507.16680 v2 pith:QIXVSHPP submitted 2025-07-22 cs.LG cs.ITcs.NImath.IT

Mario Edoardo Pandolfo , Simone Fiorellino , Emilio Calvanese Strinati , Paolo Di Lorenzo This is my paper

classification cs.LGcs.ITcs.NImath.IT

keywords semanticcommunicationslatentspacealignmentMIMOprecodinganddecodingchannelequalizationjointsource-channelcodingbiconvexoptimizationADMMsparseneuralnetworks

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

Semantic communication systems that use pre-trained neural encoders on different devices suffer from latent space mismatch: the transmitter's representation and the receiver's representation are not the same, so even a perfect physical channel can fail to convey meaning. This paper proposes to fix that mismatch with the MIMO link itself, learning a semantic precoder at the transmitter and a semantic decoder at the receiver that jointly compress the latent representation, equalize the wireless channel, and align the two latent spaces. The linear version is a biconvex optimization solved by ADMM; the nonlinear version is a neural precoder/decoder trained under a power constraint and weight-sparsity regularization. On CIFAR-10 image classification with mismatched ViT encoders, the paper reports that the joint methods outperform a disjoint pipeline that first equalizes the MIMO channel and then aligns latent spaces, with the neural model reaching about 90% accuracy while transmitting only two complex symbols. If this holds, semantic links can be made robust across heterogeneous AI models without retraining either model to use a shared representation.

What carries the argument

The carrying object is the composed semantic MIMO channel $\hat{y}=g(Hf(x)+v)$, where the complex latent vector $x$ is formed by pairing real semantic features, $f$ is a (linear or neural) semantic precoder with a power constraint, $H$ is the Kronecker-block MIMO fading matrix, and $g$ is the semantic decoder. The linear method solves the biconvex problem (6) by alternating closed-form ADMM steps for $G$ and $F$, with a projection step that enforces $\operatorname{tr}(FF^{H})\le P_T$; the neural method minimizes the loss (27) with $\ell^0$ sparsity penalties via proximal gradient descent with hard thresholding, and enforces the power budget by $\ell^2$-normalizing the encoder output. This machinery lets compression and equalization share one optimization, which is what produces the reported accuracy gains.

What would settle it

Measure the same two methods with an imperfect channel estimate $\hat H = H + E$ at increasing estimation-error power; if the disjoint Eigen-$\kappa$ baseline overtakes the jointly optimized methods before the error reaches realistic levels, the central claim that joint design is preferable under practical conditions fails.

Watch

Extended reading notes

Core claim

The central claim is that the physical MIMO channel can be repurposed as a semantic equalizer: instead of separately undoing channel effects and then mapping the transmitter's latent space onto the receiver's, one should learn the map $\hat{y}=g(Hf(x)+v)$ as a single end-to-end transformation. The paper's numerical evidence is that the jointly optimized linear method beats the Eigen-$\kappa$ baseline by roughly ten accuracy points at small compression factors and matches it after $\zeta\approx12\%$, while the neural method reaches near-90% accuracy at $\zeta\approx1\%$ and remains the best across SNR values. The same experiments show that dropping channel knowledge (channel-unaware training) degrades performance markedly, which the paper reads as confirmation that the physical channel matrix is doing real work in the alignment.

Load-bearing premise

The method assumes the transmitter knows the MIMO channel matrix $H$ exactly; when that knowledge is missing, the channel-unaware variants in the experiments degrade sharply, so imperfect channel knowledge is the premise that could sink the claimed gains.

Editorial extensions

If this is right

Joint semantic precoding/decoding makes very high compression viable: about 90% classification accuracy from only two complex symbols ($\zeta \approx 1\%$).
The linear ADMM solution is competitive with the neural one at small and moderate compression, and superior once the neural model is sparsified down to comparable FLOPs.
Channel state information is load-bearing: the channel-unaware variants of both methods degrade sharply, so the semantic equalization gain is tied to knowing $H$.
The neural method's advantage over the linear one is consistent across SNR from -20 to 30 dB at $\zeta\approx4\%$, but costs about 113 times more FLOPs at $\zeta\approx3\%$.

Reading between the lines

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

One consequence the paper leaves implicit: the same joint optimization should be trainable end-to-end with a channel estimator, so robustness to imperfect CSI becomes a tunable property rather than an assumption; a natural test is accuracy versus channel estimation error variance.
Because the method works with two different pre-trained encoders, the framework is a candidate for zero-retraining interoperability among heterogeneous edge models, not just the two architectures tested.
The sparsification results suggest a smooth accuracy-complexity frontier; an obvious extension is to compare against structured sparsity or quantization, which would shift the FLOP curve reported in the paper.

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

Worth a look. The paper formulates a genuinely under-explored problem—joint MIMO precoder/decoder design for latent-space alignment in semantic communications—and backs it with correct ADMM derivations and a clean train/test split. The linear biconvex updates check out, the power constraint handling is sensible, and the FLOPs analysis is careful. Code is released, which helps.

The soft spots are in the evaluation. The headline claim, that joint optimization beats disjoint design by roughly ten points at small ζ, rests on a comparison that is not apples-to-apples. The Eigen-κ baseline's alignment map Q is learned from clean TX/RX semantic pairs in eq. (34), while the proposed methods are trained on the channel-corrupted output G H F x + G v. In deployment, the receiver only sees noisy equalized estimates, so the baseline is fit to a distribution it never encounters at inference. A fairer disjoint baseline would fix the source precoder and SVD equalizer as in (32)-(33), then fit Q to the actual equalized outputs r_i = G_MMSE(H x_i + v_i) on the same pilot set. Until that is run, the claimed advantage could be an artifact of training the baselines on a cleaner oracle signal rather than evidence that joint design beats a well-trained disjoint design.

Smaller issues: results are six-seed averages without error bars, so the ten-point gap has no dispersion estimate. The hard-thresholding routine in Algorithm 2 is a heuristic that does not provably solve the stated ℓ0-regularized objective in (27). And the perfect-CSI assumption is explicit but limits practical conclusions; the paper says so itself. Missing a stronger DeepJSCC-based baseline is a real but not fatal limitation.

Bottom line: the formulation is a solid contribution, the math is sound, and the paper is clearly written. But the central numerical comparison has a training-distribution mismatch that a competent referee would flag. I would send it to peer review, with a request to rerun the disjoint baselines on equalized channel outputs and to add error bars. The method is new enough and the derivations careful enough to deserve referee time.

Referee Report

5 major / 5 minor

Summary. The paper proposes a joint MIMO semantic precoder/decoder design to align mismatched TX/RX latent spaces in semantic communications. It develops a linear solution, formulated as a biconvex problem and solved with ADMM, and a neural solution, trained with a proximal-gradient-with-hard-thresholding routine under an ℓ0 sparsity penalty. Experiments on CIFAR-10 with ViT-based TX/RX encoders evaluate classification accuracy as a function of compression factor, SNR, and FLOPs, and claim that joint optimization outperforms disjoint SVD equalization followed by least-squares latent alignment.

Significance. The problem is timely for AI-native semantic communications, and the paper contains a clean, algebraically correct ADMM derivation for the linear case together with a useful FLOP-counting analysis. If the joint-versus-disjoint advantage were established with a fair comparison, this would be a valuable contribution. However, the current experimental protocol is confounded because the disjoint baseline is trained on clean latent pairs while the proposed methods are trained on channel-corrupted outputs, and the claimed literature benchmarks are not actually evaluated. The quantitative central claims are therefore not yet established.

major comments (5)

[Section IV, Baselines and Fig. 2, Eq. (34)] The disjoint baseline is trained on data it never sees at inference. The alignment map Q in (34) is fitted to clean TX/RX latent pairs (s_T^(i), s_R^(i)), whereas the proposed decoders are optimized on channel-corrupted outputs G H F x + G v in (6) and g_psi(H f_theta(x)+v) in (27). At test time, the baseline's Q receives equalized noisy estimates, so the comparison in Fig. 2 conflates joint optimization with training on the deployment distribution. Please add a fair disjoint baseline that fixes the source and equalizer modules as in (32)-(33), forms r_i = G_MMSE(H x_i + v_i) on the same training set, and fits Q to (r_i, s_R^(i)); report this variant in Figs. 2 and 3.
[Section III.B, Eq. (27)-(29) and Algorithm 2] The hard-thresholding routine is not the proximal operator for the ℓ0 norm appearing in (27). For a composite objective L(w)+λ||w||_0, a proximal gradient step with step size η applies the threshold sqrt(2λη) to w - η∇L(w), whereas Algorithm 2 uses τ_i = βη or γη in (29). The paper therefore does not actually solve (27), and the sparsity-accuracy trade-off in Fig. 4 is not connected to the stated objective. Please either use the correct proximal operator or explicitly present Algorithm 2 as a heuristic sparsification routine, and report the achieved sparsity levels and objective values for the points in Fig. 4.
[Section IV, Figs. 2-4] The paper states that all results are averaged over six seeds but reports no error bars, standard deviations, or confidence intervals. With only six seeds and small semantic-pilot sets (n=420), the reported differences—such as the roughly ten-percentage-point linear-versus-Eigen-κ gap at small compression factors—could lie within seed-to-seed variability. Please report mean ± std or an equivalent variability measure, and indicate the seed-level variance for the key comparisons.
[Section IV, Baselines; Introduction] The Introduction claims the results demonstrate better performance with respect to 'available benchmarks from the literature,' but the only baselines in Section IV are the custom Eigen-κ, Top-κ, and First-κ schemes. The prior semantic equalization methods cited in the Introduction, such as the relative-representation baselines of [27], [31], and [32], are not evaluated. Please add at least one literature baseline or revise the claim to specify 'custom disjoint baselines.'
[Section II and Section IV, Eqs. (32)-(33)] There is a dimension inconsistency in the channel model. The text defines H ∈ C^{NT×NR} and then states H = I_K ⊗ H ∈ C^{KNR×KNT}, but for the model y = Hx + v, the channel must map C^{KNT} to C^{KNR}; the stated orientation is reversed. This propagates to the SVD baselines in (32)-(33), where F = 1_K ⊗ V does not have the required dimension KNT×(d/2) and G = 1_K^T ⊗ (...) is not dimensionally compatible with the rest of the system. Please clarify the dimension conventions and correct the baseline definitions so they are consistent with Eq. (2).

minor comments (5)

[Algorithm 2 and Fig. 4] The abbreviation 'PDG' should be 'PGD' in the Algorithm 2 caption and in the Fig. 4 legend.
[Eq. (32)] The phrase 'the K-dimensional unitary vector' is misleading; 1_K is an all-ones vector, not a unitary vector.
[Fig. 2 and Section IV] The 'channel-unaware' variants are not defined; please specify what optimization is performed when channel knowledge is absent.
[Section II and Section IV] The SNR is used without a definition connecting it to the power budget P_T and the noise covariance Σ_v; please define it explicitly.
[Fig. 2] The accuracy axis begins at 0.52, which can visually exaggerate differences between methods; consider starting the axis at 0.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the joint MIMO semantic equalization is trained and tested on disjoint data, and self-citations are background only.

full rationale

The paper's central claim is an empirical comparison, not a derivation from its own input. Objectives (6) and (27) minimize a semantic MSE over a training set of semantic pilots with a channel-corrupted forward model, and the reported accuracy is evaluated on held-out CIFAR-10 test images, so the result is not fitted to the test set and is not equivalent to the objective by construction. The disjoint baselines in (32)-(34) fit the alignment map Q to clean TX/RX latent pairs, whereas the proposed decoders are optimized on channel-corrupted outputs; if this is confirmed, it is a benchmark fairness concern, not circular reasoning. Self-references ([15], [32], [6]) are used as background on relative representations and goal-oriented communications and are not load-bearing for the MIMO optimization or the numerical comparison. No equation, theorem, or defined quantity in the paper is shown to reduce to its own input, so the circularity score is 0.

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

The central claim rests on standard signal processing assumptions (perfect CSI, Euclidean semantic distance) and on the availability of paired semantic pilots; no invented entities are introduced.

free parameters (4)

ADMM penalty parameter ρ = 100
Set in Section IV for all linear model experiments; a standard hyperparameter that affects convergence speed but not the optimal solution.
Sparsity regularization strengths β = γ = 0 in Figs. 2-3; varied over {0,...,70} in Fig. 4
Hand-chosen to trade off accuracy and FLOPs; they control the neural model's sparsity but are not fit to data.
Learning rate η = 10^-3
Used for the neural model training; standard optimizer hyperparameter.
Training length = 20 ADMM iterations; 50 NN epochs
Chosen to ensure convergence in the reported experiments.

assumptions (4)

domain assumption Perfect knowledge of the MIMO channel matrix H at both transmitter and receiver.
Stated in Section III before the linear method; all training objectives use H, and channel-unaware variants perform much worse.
domain assumption Semantic mismatch can be minimized with mean squared error in the complex latent space.
Used to define the objective in (4)-(6) and (27); the paper notes other metrics are possible but does not test them.
standard math The TX and RX latent spaces are pre-whitened so the power constraint depends only on F.
Stated as w.l.o.g. in Section III.A; a standard linear algebra normalization.
domain assumption Both TX and RX feature extractors are fixed pretrained DNNs; only the semantic precoder/decoder are learnable.
The system model in Section II treats the DNN encoders as given and the precoder/decoder as learnable transformations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Latent Space Alignment for AI-Native MIMO Semantic Communications." pith.science (2026). https://pith.science/paper/QIXVSHPP

@misc{pith2026250716680,
  author       = {Pith},
  title        = {Pith review of: Latent Space Alignment for AI-Native MIMO Semantic Communications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QIXVSHPP}},
  note         = {Machine review of arXiv:2507.16680}
}

read the original abstract

Semantic communications focus on prioritizing the understanding of the meaning behind transmitted data and ensuring the successful completion of tasks that motivate the exchange of information. However, when devices rely on different languages, logic, or internal representations, semantic mismatches may occur, potentially hindering mutual understanding. This paper introduces a novel approach to addressing latent space misalignment in semantic communications, exploiting multiple-input multiple-output (MIMO) communications. Specifically, our method learns a MIMO precoder/decoder pair that jointly performs latent space compression and semantic channel equalization, mitigating both semantic mismatches and physical channel impairments. We explore two solutions: (i) a linear model, optimized by solving a biconvex optimization problem via the alternating direction method of multipliers (ADMM); (ii) a neural network-based model, which learns semantic MIMO precoder/decoder under transmission power budget and complexity constraints. Numerical results demonstrate the effectiveness of the proposed approach in a goal-oriented semantic communication scenario, illustrating the main trade-offs between accuracy, communication burden, and complexity of the solutions.

Figures

Figures reproduced from arXiv: 2507.16680 by the authors.

**Figure 1.** System model: The semantic precoder f of TX encodes a compressed latent space si into a transmitter vector of dimension NT . It is transmitted through the MIMO channel H with noise v. Then, the received vector is decoded by the semantic decoder g and decompressed to match the RX latent space. enabling SCs schemes [7], [25] while performing both semantic compression and channel equalization (source-channel coding).… view at source ↗

**Figure 3.** Accuracy versus SNR, with ζ ≈ 4% and 4,200 SPs per class. 105 106 FLOPs 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy Case Neural Semantic ζ ≈ 5% PDG with Hard Thresholding Neural Semantic ζ ≈ 3% PDG with Hard Thresholding Linear Semantic ζ ≈ 5% Linear Semantic ζ ≈ 3% [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Accuracy versus FLOPs, with ζ ≈ 3% and ζ ≈ 5%, SNR = 20 dB and 4,200 SPs per class. parameters β = γ across the set {0, 10, 20, 30, 40, 50, 60, 70} to enforce different level of weight sparsity and, consequently, different computational complexity (cf. 31). The results are reported in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 26 canonical work pages

[27]

Relative representations enable zero-shot latent space communication,

L. Moschella, V . Maiorca, M. Fumero, A. Norelli, F. Locatello, and E. Rodol `a, “Relative representations enable zero-shot latent space communication,” arXiv preprint arXiv:2209.15430 , 2022

arXiv 2022
[31]

Latent space translation via inverse relative projection,

V . Maiorca, L. Moschella, M. Fumero, F. Locatello, and E. Rodol `a, “Latent space translation via inverse relative projection,” arXiv preprint arXiv:2406.15057, 2024

arXiv 2024
[32]

Relative representations of latent spaces enable efficient semantic channel equalization,

T. H ¨uttebr¨aucker, S. Fiorellino, M. Sana, P. Di Lorenzo, and E. Cal- vanese Strinati, “Relative representations of latent spaces enable efficient semantic channel equalization,” in GLOBECOM 2024-2024 IEEE Global Communications Conference . IEEE, 2024

work page 2024
[1]

Survey on 6g frontiers: Trends, applications, requirements, technologies and future research,

C. De Alwis, A. Kalla, Quoc-Viet Pham, P. Kumar, K. Dev, Won-Joo Hwang, and M. Liyanage, “Survey on 6g frontiers: Trends, applications, requirements, technologies and future research,” IEEE Open Journal of the Communications Society , vol. 2, pp. 836–886, 2021

work page 2021
[2]

A new communication paradigm: From bit accuracy to semantic fidelity,

G. Shi, D. Gao, X. Song, J. Chai, M. Yang, X. Xie, L. Li, and X. Li, “A new communication paradigm: From bit accuracy to semantic fidelity,” arXiv preprint arXiv:2101.12649 , 2021

arXiv 2021
[3]

Towards a new internet for the year 2030 and beyond,

R. Li, “Towards a new internet for the year 2030 and beyond,” Proc. 3rd Annu. ITU IMT-2020/5G Workshop Demo Day , pp. 1–21, 2018

2020
[4]

6G networks: Beyond shan- non towards semantic and goal-oriented communications,

E. Calvanese Strinati and S. Barbarossa, “6G networks: Beyond shan- non towards semantic and goal-oriented communications,” Computer Networks, vol. 190, pp. 107930, 2021

work page 2021
[5]

Semantic communication: A survey on research landscape, challenges, and future directions,

T. Getu, G. Kaddoum, and M. Bennis, “Semantic communication: A survey on research landscape, challenges, and future directions,” Authorea Preprints, 2023

2023

Show all 39 references

[6]

Goal-oriented and semantic communication in 6G AI-native networks: The 6G-goals approach,

E. Calvanese Strinati, P. Di Lorenzo, et al., “Goal-oriented and semantic communication in 6G AI-native networks: The 6G-goals approach,” arXiv preprint arXiv:2402.07573 , 2024

2024 arXiv
[7]

Beyond transmitting bits: Context, semantics, and task-oriented communications,

D. G ¨und¨uz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, Kai Kit Wong, and Chan-Byoung Chae, “Beyond transmitting bits: Context, semantics, and task-oriented communications,” IEEE Journal on Selected Areas in Communications , vol. 41, no. 1, pp. 5–41, 2022

2022
[8]

Semantic communications: Overview, open issues, and future research directions,

X. Luo, Hsiao-Hwa Chen, and Q. Guo, “Semantic communications: Overview, open issues, and future research directions,” IEEE Wireless Communications, vol. 29, no. 1, pp. 210–219, 2022

2022
[9]

The semantic communication game,

B. G ¨uler, A. Yener, and A. Swami, “The semantic communication game,” IEEE Transactions on Cognitive Communications and Network- ing, vol. 4, no. 4, pp. 787–802, 2018

2018
[10]

Towards a theory of semantic communication,

J. Bao, P. Basu, M. Dean, C. Partridge, A. Swami, W. Leland, and J. A. Hendler, “Towards a theory of semantic communication,” in 2011 IEEE Network Science Workshop. IEEE, 2011, pp. 110–117

2011
[11]

Semantics-empowered communication for networked intelligent systems,

M. Kountouris and N. Pappas, “Semantics-empowered communication for networked intelligent systems,” IEEE Communications Magazine , vol. 59, no. 6, pp. 96–102, 2021

2021
[12]

The role of fidelity in goal- oriented semantic communication: A rate distortion approach,

P. A. Stavrou and M. Kountouris, “The role of fidelity in goal- oriented semantic communication: A rate distortion approach,” IEEE Transactions on Communications , 2023

2023
[13]

Adaptive resource optimization for edge inference with goal-oriented communi- cations,

F. Binucci, P. Banelli, P. Di Lorenzo, and S. Barbarossa, “Adaptive resource optimization for edge inference with goal-oriented communi- cations,” EURASIP J. on Advances in Signal Proc. , , no. 1, pp. 123, 2022

2022
[14]

Goal-oriented communications for the IoT: System design and adaptive resource optimization,

P. Di Lorenzo, M. Merluzzi, F. Binucci, C. Battiloro, P. Banelli, E. Cal- vanese Strinati, and S. Barbarossa, “Goal-oriented communications for the IoT: System design and adaptive resource optimization,” IEEE Internet of Things Magazine , vol. 6, no. 4, pp. 26–32, 2023

2023
[15]

Dynamic relative representations for goal-oriented semantic communi- cations,

S. Fiorellino, C. Battiloro, E. Calvanese Strinati, and P. Di Lorenzo, “Dynamic relative representations for goal-oriented semantic communi- cations,” arXiv preprint arXiv:2403.16986 , 2024

2024 arXiv
[16]

Rea- soning with the theory of mind for pragmatic semantic communication,

C. Kurisummoottil Thomas, E. Calvanese Strinati, and W. Saad, “Rea- soning with the theory of mind for pragmatic semantic communication,” arXiv preprint arXiv:2311.18224 , 2023

2023 arXiv
[17]

Semantic communications based on adaptive generative models and information bottleneck,

S. Barbarossa, D. Comminiello, E. Grassucci, F. Pezone, S. Sardellitti, and P.o Di Lorenzo, “Semantic communications based on adaptive generative models and information bottleneck,” IEEE Communications Magazine, vol. 61, no. 11, pp. 36–41, 2023

2023
[18]

Deep learning enabled semantic communication systems,

H. Xie, Z. Qin, G. Ye Li, and Biing-Hwang Juang, “Deep learning enabled semantic communication systems,” IEEE Transactions on Signal Processing, vol. 69, pp. 2663–2675, 2021

2021
[19]

A mathematical theory of communication,

C. E. Shannon, “A mathematical theory of communication,” The Bell system technical journal , vol. 27, no. 3, pp. 379–423, 1948

1948
[20]

Deep joint source- channel coding for wireless image transmission,

E. Bourtsoulatze, D. Burth Kurka, and D. G ¨und¨uz, “Deep joint source- channel coding for wireless image transmission,” IEEE Transactions on Cognitive Communications and Networking , vol. 5, no. 3, pp. 567–579, 2019

2019
[21]

Ofdm-guided deep joint source channel coding for wireless multipath fading channels,

M. Yang, C. Bian, and Hun-Seok Kim, “Ofdm-guided deep joint source channel coding for wireless multipath fading channels,” IEEE Transactions on Cognitive Communications and Networking , vol. 8, no. 2, pp. 584–599, 2022

2022
[22]

Wireless image transmission using deep source channel coding with attention modules,

J. Xu, B. Ai, W. Chen, A. Yang, P. Sun, and M. Rodrigues, “Wireless image transmission using deep source channel coding with attention modules,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 4, pp. 2315–2328, 2022

2022
[23]

Nonlinear transform source-channel coding for semantic communications,

J. Dai, S. Wang, K. Tan, Z. Si, X. Qin, K. Niu, and P. Zhang, “Nonlinear transform source-channel coding for semantic communications,” IEEE Journal on Selected Areas in Communications , vol. 40, no. 8, pp. 2300– 2316, 2022

2022
[24]

Deep joint source-channel coding for adaptive image transmission over mimo channels,

H. Wu, Y . Shao, C. Bian, K. Mikolajczyk, and D. G ¨und¨uz, “Deep joint source-channel coding for adaptive image transmission over mimo channels,” IEEE Transactions on Wireless Communications, vol. 23, no. 10, pp. 15002–15017, 2024

2024
[25]

Deep joint source-channel coding for semantic communications,

J. Xu, Tze-Yang Tung, B. Ai, W. Chen, Y . Sun, and D. G ¨und¨uz, “Deep joint source-channel coding for semantic communications,” IEEE Communications Magazine, vol. 61, no. 11, pp. 42–48, 2023

2023
[26]

Towards optimal transport with global invariances,

D. Alvarez-Melis, S. Jegelka, and T. S Jaakkola, “Towards optimal transport with global invariances,” in The 22nd International Conference on Artificial Intelligence and Statistics . PMLR, 2019, pp. 1870–1879

2019
[28]

Stitchable neural networks,

Z. Pan, J. Cai, and B. Zhuang, “Stitchable neural networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 16102–16112

2023
[29]

Semantic channel equalizer: Modelling language mismatch in multi-user semantic communications,

M. Sana and E. Calvanese Strinati, “Semantic channel equalizer: Modelling language mismatch in multi-user semantic communications,” arXiv preprint arXiv:2308.03789 , 2023

2023 arXiv
[30]

On the direct alignment of latent spaces,

Z. L ¨ahner and M. Moeller, “On the direct alignment of latent spaces,” in Proceedings of UniReps: the First Workshop on Unifying Represen- tations in Neural Models . PMLR, 2024, pp. 158–169

2024
[33]

Distributed optimization and statistical learning via the alternating direction method of multipliers,

S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein, et al., “Distributed optimization and statistical learning via the alternating direction method of multipliers,” Foundations and Trends® in Machine learning , vol. 3, no. 1, pp. 1–122, 2011

2011
[34]

Algorithm 432 [c2]: Solution of the matrix equation ax + xb = c [f4],

R. H. Bartels and G. W. Stewart, “Algorithm 432 [c2]: Solution of the matrix equation ax + xb = c [f4],” Communications of the ACM , vol. 15, pp. 820 – 826, 1972

1972
[35]

Kkt conditions, first-order and second-order optimization, and distributed optimization: Tutorial and survey,

B. Ghojogh, A. Ghodsi, F. Karray, and M. Crowley, “Kkt conditions, first-order and second-order optimization, and distributed optimization: Tutorial and survey,” 2021

2021
[36]

Deep complex networks,

C. Trabelsi, O. Bilaniuk, Y . Zhang, D. Serdyuk, S. Subramanian, J. Felipe Santos, S. Mehri, N. Rostamzadeh, Y . Bengio, and C. J. Pal, “Deep complex networks,” 2018

2018
[37]

Proximal algorithms,

N. Parikh, S. Boyd, et al., “Proximal algorithms,” Foundations and trends® in Optimization , vol. 1, no. 3, pp. 127–239, 2014

2014
[38]

Learning multiple layers of features from tiny images,

A. Krizhevsky, G. Hinton, et al., “Learning multiple layers of features from tiny images,” 2009

2009
[39]

Pytorch image models,

R. Wightman, “Pytorch image models,” https://github.com/rwightman/ pytorch-image-models, 2019

2019

Pith tools

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

[1] [27]

Relative representations enable zero-shot latent space communication,

L. Moschella, V . Maiorca, M. Fumero, A. Norelli, F. Locatello, and E. Rodol `a, “Relative representations enable zero-shot latent space communication,” arXiv preprint arXiv:2209.15430 , 2022

arXiv 2022

[2] [31]

Latent space translation via inverse relative projection,

V . Maiorca, L. Moschella, M. Fumero, F. Locatello, and E. Rodol `a, “Latent space translation via inverse relative projection,” arXiv preprint arXiv:2406.15057, 2024

arXiv 2024

[3] [32]

Relative representations of latent spaces enable efficient semantic channel equalization,

T. H ¨uttebr¨aucker, S. Fiorellino, M. Sana, P. Di Lorenzo, and E. Cal- vanese Strinati, “Relative representations of latent spaces enable efficient semantic channel equalization,” in GLOBECOM 2024-2024 IEEE Global Communications Conference . IEEE, 2024

work page 2024

[4] [1]

Survey on 6g frontiers: Trends, applications, requirements, technologies and future research,

C. De Alwis, A. Kalla, Quoc-Viet Pham, P. Kumar, K. Dev, Won-Joo Hwang, and M. Liyanage, “Survey on 6g frontiers: Trends, applications, requirements, technologies and future research,” IEEE Open Journal of the Communications Society , vol. 2, pp. 836–886, 2021

work page 2021

[5] [2]

A new communication paradigm: From bit accuracy to semantic fidelity,

G. Shi, D. Gao, X. Song, J. Chai, M. Yang, X. Xie, L. Li, and X. Li, “A new communication paradigm: From bit accuracy to semantic fidelity,” arXiv preprint arXiv:2101.12649 , 2021

arXiv 2021

[6] [3]

Towards a new internet for the year 2030 and beyond,

R. Li, “Towards a new internet for the year 2030 and beyond,” Proc. 3rd Annu. ITU IMT-2020/5G Workshop Demo Day , pp. 1–21, 2018

2020

[7] [4]

6G networks: Beyond shan- non towards semantic and goal-oriented communications,

E. Calvanese Strinati and S. Barbarossa, “6G networks: Beyond shan- non towards semantic and goal-oriented communications,” Computer Networks, vol. 190, pp. 107930, 2021

work page 2021

[8] [5]

Semantic communication: A survey on research landscape, challenges, and future directions,

T. Getu, G. Kaddoum, and M. Bennis, “Semantic communication: A survey on research landscape, challenges, and future directions,” Authorea Preprints, 2023

2023