Pith. sign in

REVIEW 2 major objections 5 minor 31 references

CSI compression and aging can be solved together by predicting latent codes, not raw channel matrices, inside a 3GPP pipeline.

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

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

T0 review · grok-4.5

2026-07-12 11:27 UTC pith:QETFFRVT

load-bearing objection Solid 3GPP-aligned engineering that cleanly puts CPC on latents inside the standardized compressor; the age-aware systems claim is still unclosed. the 2 major comments →

arxiv 2607.05419 v1 pith:QETFFRVT submitted 2026-06-26 cs.IT cs.AIeess.SPmath.IT

Contrastive Predictive Coding with Compression for Enhanced Channel State Feedback in Wireless Networks

Ahmed Y. Radwan , Hina Tabassum , Fahad Syed Muhammad , Matthew Baker This is my paper
classification cs.IT cs.AIeess.SPmath.IT
keywords contrastive predictive codingCSI feedbackchannel aging3GPPjoint compression and predictionInfoNCESGCSlatent forecasting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Wireless systems need fresh channel state information (CSI) for beamforming and scheduling, but today’s standards treat compression of CSI for uplink feedback and prediction of future CSI as two separate problems. Channel aging—the CSI becoming outdated before it is used—therefore remains unaddressed inside the standardized feedback pipeline. This paper shows that both problems can be solved jointly by inserting Contrastive Predictive Coding into a 3GPP-compliant autoencoder: instead of forecasting the high-dimensional CSI matrix itself, the model forecasts future latent representations and trains with a combined reconstruction-plus-contrastive loss. Two placements of the predictive module are offered—one before quantization at the user device and one after decompression at the base station—so that feedback overhead stays fixed at 64 bits while decoder compute can drop by a factor of thirty-two. On company-provided 3GPP datasets the before-compression variant routinely exceeds 90 percent reconstruction accuracy, demonstrating that age-aware feedback is achievable without enlarging the air interface.

Core claim

Integrating Contrastive Predictive Coding into a 3GPP-compliant quantized CSI autoencoder, and jointly optimizing a 1-SGCS reconstruction loss with an InfoNCE temporal loss, yields latent representations that remain predictive across future horizons while preserving or improving reconstruction fidelity at the same 64-bit feedback budget. The CPC-before-Compression placement achieves SGCS scores above 0.90 on multiple company datasets and reduces decoder GFLOPs by roughly 32 imes relative to the pure-compression 3GPP baseline.

What carries the argument

A combined loss L_total = α(1-SGCS) + (1-α)InfoNCE applied to future latent vectors produced by a GRU (or LSTM/Transformer) inside the standardized encoder–quantizer–decoder pipeline; the two architectural variants place the autoregressive module either before the quantized bottleneck (UE side) or after reconstruction (BS side).

Load-bearing premise

That high reconstruction similarity and low contrastive loss on company CSI tensors are enough to prove that channel aging is actually mitigated in real multi-user beamforming or scheduling.

What would settle it

Replace the SGCS/InfoNCE tables with an end-to-end MU-MIMO simulation that measures spectral efficiency or block-error rate under a realistic CSI delay; if the CPC variants show no gain over the pure-compression baseline once the delay is introduced, the age-aware claim fails.

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

If this is right

  • Standardized CSI feedback can become age-aware without increasing the 64-bit uplink payload.
  • Base-station decoders can be made an order of magnitude lighter when temporal modeling is moved to the encoder side.
  • User devices with strict power budgets can keep the original lightweight encoder by shifting CPC to the base station after decompression.
  • Reconstruction quality remains stable across prediction horizons of 2–20 steps, so the base station can obtain multi-step future CSI at no extra reconstruction cost.
  • Pruning and low-rank factorization of the GRU further reduce inference cost while retaining most of the SGCS gain.

Where Pith is reading between the lines

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

  • Because the method never reports link-level metrics, its practical value for 3GPP will be settled only after spectral-efficiency or BLER trials under delayed CSI are published.
  • The near-zero InfoNCE observed when CPC is applied after decompression suggests that the compression bottleneck itself may destroy the temporal structure the contrastive loss needs; this points to a possible information-theoretic limit on post-bottleneck prediction.
  • The same latent-forecast idea could be tried on other 3GPP AI use-cases (e.g., beam prediction or positioning) that also suffer from aging under a fixed feedback budget.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes integrating Contrastive Predictive Coding (CPC) into a 3GPP-compliant CSI compression pipeline so that compression and temporal prediction are jointly optimized. Two variants are introduced: CPC-before-Compression (GRU-based autoregressive modeling on encoder features before the quantized linear bottleneck) and CPC-after-Compression (temporal modeling deferred to the BS decoder). Training uses a combined loss α L_SGCS + (1−α) L_InfoNCE. On proprietary 3GPP company datasets (Nokia, Oppo, CATT and a mixed set), CPC-before-Compression reports SGCS often exceeding 0.90 at the same 64-bit feedback overhead while reducing decoder GFLOPs by roughly 32 imes relative to the 3GPP baseline; CPC-after-Compression keeps the encoder footprint identical. Ablations cover backbone choice, prediction horizon T, bottleneck size, GRU hidden dimension, reconstruction loss, decoder capacity and structured pruning. Source code is released.

Significance. If the latent-level gains translate to usable future CSI under delay, the work would be a practical contribution to 3GPP AI/ML CSI feedback: it unifies two currently separate Rel-18 use-cases inside a standardized quantized autoencoder, keeps the 64-bit overhead fixed, and supplies concrete complexity numbers plus open code. The multi-company evaluation and systematic ablations are strengths relative to much of the CSI-compression literature. The significance is currently limited by the absence of any end-to-end MU-MIMO metric (SE, beamforming gain or BLER under aged CSI), so the “age-aware” claim remains an extrapolation from representation quality.

major comments (2)
  1. The abstract, §I and §V repeatedly claim an “age-aware” CSI feedback solution that mitigates channel aging. The entire evaluation (§IV, Tables IV–VIII, Figs. 3–4) reports only SGCS on reconstructed/predicted latents, InfoNCE, parameter/GFLOP counts and pruning curves. No spectral efficiency, beamforming gain, or BLER under delayed CSI is shown, nor is there a comparison against a pure-compression baseline that simply re-uses the last reconstructed CSI. Without at least one such closed-loop experiment the systems-level age-aware assertion is not substantiated by the reported evidence.
  2. For CPC-after-Compression, Table IV shows InfoNCE values near zero (0.005–0.074) across all train/test pairs. The text itself notes that “the contrastive learning loses its effectiveness” after the compression bottleneck. This raises the question whether the after-variant actually learns useful temporal structure or merely reconstructs the current frame; the paper should either demonstrate that the near-zero InfoNCE still yields predictive utility or qualify the claim that both variants address channel aging.
minor comments (5)
  1. Table I lists several related works with incomplete architecture or loss entries (e.g., [11]); a short note that those fields were not reported would improve transparency.
  2. Eq. (1) defines SGCS with an additive ε inside the denominator; a brief justification for the particular form (versus plain cosine similarity or NMSE) would help readers unfamiliar with the 3GPP metric choice.
  3. Fig. 2 caption and the surrounding text should clarify whether the “Future Linear / FC” blocks are shared or horizon-specific, as Algorithms 1–2 suggest dedicated FC_k layers.
  4. The pruning study (Fig. 3) is performed only on CPC-before-Compression; a one-sentence statement that the after-variant was left unpruned because its encoder already matches the baseline would avoid an apparent asymmetry.
  5. Minor typographical inconsistencies appear (e.g., “N0KIR4” vs “NOKIA”, “CAT0R4” vs “CATT” in the dataset filenames versus table labels).

Circularity Check

0 steps flagged

No circularity: standard empirical ML training and held-out evaluation of a joint CPC+compression architecture; reported SGCS/InfoNCE are ordinary test metrics, not forced by construction.

full rationale

The paper proposes two architectural variants that insert a GRU-based CPC module either before or after a 3GPP-style quantized ResNet autoencoder, trains end-to-end with the composite loss L_total = α(1-SGCS)+(1-α)InfoNCE (α=0.5), and reports SGCS and InfoNCE on held-out splits of the Nokia/Oppo/CATT 3GPP datasets together with complexity numbers. This is ordinary supervised/contrastive practice: the model is optimized for the same quantities later measured on unseen data; nothing reduces by algebraic identity or by a fitted constant renamed as a prediction. There is no self-definitional loop (SGCS and InfoNCE are independently defined), no uniqueness theorem imported from the authors, no ansatz smuggled via self-citation, and the single self-citation ([31], TinyML pruning) appears only in the conclusion and is not load-bearing for the CSI results. The systems-level claim that high latent SGCS constitutes an age-aware solution is an untested extrapolation, but that is a validity gap, not circularity. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

Empirical systems paper: load-bearing content is architectural choices and hyperparameters chosen for the 3GPP CSI setting, plus domain assumptions that latent CPC coherence tracks useful channel aging mitigation. No new physical entities; free parameters are the usual ML knobs (α, τ, T, L, hidden size, bit width) fixed by hand or grid.

free parameters (5)
  • loss mix α = 0.5
    α=0.5 balances L_SGCS and L_InfoNCE; chosen empirically, not derived.
  • InfoNCE temperature τ = 0.1
    Fixed to 0.1 in Table III; controls contrastive sharpness.
  • prediction horizon T and context window L = T=5, L=10
    Default T=5, L=10; ablated but defaults drive main tables.
  • GRU hidden size / latent bottleneck = hidden=128, z-dim=32, 2-bit
    Hidden=128 and 32-dim 2-bit quantized bottleneck set complexity and rate; ablated but central configs use these.
  • pruning ratio (post-training) = varied 0–0.8
    Structured GRU pruning ratios explored in Fig. 3; performance–size trade-off depends on chosen sparsity.
axioms (4)
  • domain assumption 3GPP Type-1 offline joint training of encoder/decoder at one side then split deployment is an acceptable and representative training paradigm for CSI feedback.
    Stated in §III and used throughout experiments; alternatives (Type-2/3) are reviewed but not evaluated for CPC.
  • domain assumption Squared Generalized Cosine Similarity (SGCS) and InfoNCE on latents are adequate training/evaluation proxies for CSI reconstruction quality and temporal predictive coherence.
    Eqs. (1)–(5); no SE/BLER validation of the aging claim.
  • domain assumption The ResNet CSI autoencoder with quantized linear bottleneck from 3GPP drafts is the correct standardized backbone to extend.
    §III.1 and Fig. 2; all variants build on this architecture.
  • ad hoc to paper InfoNCE contrastive objective on future latents captures channel dynamics relevant to aging without predicting high-dimensional CSI matrices.
    Core design choice of the paper; CPC-after results (near-zero InfoNCE) show this can fail after a hard bottleneck.
invented entities (2)
  • CPC-before-Compression architecture no independent evidence
    purpose: Apply GRU+FC predictive heads on encoded features before the quantized bottleneck so future latents are compressed and fed back.
    New pipeline layout relative to 3GPP baseline and to [16]; independent evidence is empirical SGCS/complexity only.
  • CPC-after-Compression architecture no independent evidence
    purpose: Keep UE encoder identical to 3GPP baseline and attach CPC only at the BS after reconstruction to cut UE complexity.
    Second layout for resource-limited UEs; contrastive signal largely collapses post-bottleneck in reported numbers.

pith-pipeline@v1.1.0-grok45 · 20387 in / 3509 out tokens · 38456 ms · 2026-07-12T11:27:03.547561+00:00 · methodology

0 comments
read the original abstract

Accurate and timely channel state information (CSI) is essential for next-generation wireless systems, yet existing works treat CSI compression and CSI prediction as separate problems, both in academia and in current 3GPP studies. Consequently, channel aging remains insufficiently addressed within standardized CSI feedback pipelines. In this article, we propose a unified compression-prediction framework that integrates Contrastive Predictive Coding (CPC) directly into the 3GPP-compliant CSI compression architecture. Instead of predicting high-dimensional CSI matrices, our approach forecasts future latent representations and jointly optimizes reconstruction fidelity and temporal predictive coherence via a combined 1-SGCS and InfoNCE objective. This design enables temporal representation learning without increasing feedback overhead. We present two variants: CPC-before-Compression, which performs autoregressive modeling on encoded features prior to quantization, and CPC-after-Compression, which shifts temporal modeling to the base-station to reduce the complexity of users' devices. Evaluations on 3GPP-compliant datasets from Nokia, Oppo, and CATT show that CPC-before-Compression achieves over 90% reconstruction accuracy with 32x lower decoder GFLOPs than the 3GPP baseline, while CPC-after-Compression preserves an identical encoder footprint and the same 64-bit feedback overhead. By unifying compression and prediction within a standardized pipeline, the proposed framework provides an age-aware, computationally efficient CSI feedback solution. The source code is publicly available at: https://github.com/AhmedRadwan02/cpc-3gpp

Figures

Figures reproduced from arXiv: 2607.05419 by Ahmed Y. Radwan, Fahad Syed Muhammad, Hina Tabassum, Matthew Baker.

Figure 1
Figure 1. Figure 1: Overview of the model training types studied in 3GPP- Release 18: (1) Joint training of the encoder and decoder [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The CPC-after-Compression model (below) builds on the 3GPP standard backbone by adding CPC to compressed [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Impact of structured pruning on CPC-before [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

31 extracted references · 4 linked inside Pith

  1. [1]

    Molecular absorption- aware user assignment, spectrum, and power allocation in dense thz networks with multi-connectivity,

    M. Amin Saeidi, H. Tabassum, and M. Alizadeh, “Molecular absorption- aware user assignment, spectrum, and power allocation in dense thz networks with multi-connectivity,”IEEE Trans. on Wireless Commun., vol. 23, no. 11, pp. 16 404–16 420, 2024

    2024

  2. [2]

    Power control with QoS guarantees: A differentiable projection-based unsupervised learning framework,

    M. Alizadeh and H. Tabassum, “Power control with QoS guarantees: A differentiable projection-based unsupervised learning framework,”IEEE Trans. on Commun., vol. 71, no. 8, pp. 4605–4619, 2023

    2023

  3. [3]

    Technical Report (TR) 38.843: Study on Artificial Intelligence (AI)/Machine Learning (ML) for NR air interface,

    3GPP, “Technical Report (TR) 38.843: Study on Artificial Intelligence (AI)/Machine Learning (ML) for NR air interface,” 3rd Generation Partnership Project (3GPP), Technical Specification Group Radio Access Network (TSG RAN), 3GPP Technical Report TR 38.843, Sep. 2025, release 19, Version 19.0.0

    2025

  4. [4]

    Learning at the speed of wireless: Online real-time learning for AI-enabled MIMO in NextG,

    J. Xu, S. Jere, Y . Song, Y .-H. Kao, L. Zheng, and L. Liu, “Learning at the speed of wireless: Online real-time learning for AI-enabled MIMO in NextG,”IEEE Commun. Magazine, 2024

    2024

  5. [5]

    Summary#0 for CSI evaluation of [112-R18-AI/ML] (Agenda Item 9.2.2.1),

    3GPP TSG RAN1 Working Group and Moderator (Huawei), “Summary#0 for CSI evaluation of [112-R18-AI/ML] (Agenda Item 9.2.2.1),” https://www.3gpp.org/ftp/tsg ran/WG1 RL1/TSGR1 112/Inbox/drafts/9.2(FS NR AIML air)/9.2.2.1/1st%20round/ R1-221xxxx%20Summary%230%20of%20%5B112-R18-AIML%5D CSI Eva v003 Lenovo vivo.docx, 3GPP TSG-RAN WG1 Meeting #112, Athens, Greece...

    2023

  6. [6]

    On the combination of ai and wireless technologies: 3GPP standardization progress,

    C. Sun, T. Cui, W. Zhang, Y . Bai, S. Wang, and H. Li, “On the combination of ai and wireless technologies: 3GPP standardization progress,” in2024 IEEE/CIC Intl. Conf. on Commun. in China (ICCC Wkshp.). IEEE, 2024, pp. 523–528

    2024

  7. [7]

    An efficient deep learning framework for low rate massive mimo csi reporting,

    Z. Liu, L. Zhang, and Z. Ding, “An efficient deep learning framework for low rate massive mimo csi reporting,”IEEE Transactions on Commun., vol. 68, no. 8, pp. 4761–4772, 2020

    2020

  8. [8]

    Distributed deep convolu- tional compression for massive mimo csi feedback,

    M. B. Mashhadi, Q. Yang, and D. G ¨und¨uz, “Distributed deep convolu- tional compression for massive mimo csi feedback,”IEEE Transactions on Wireless Commun., vol. 20, no. 4, pp. 2621–2633, 2020

    2020

  9. [9]

    Deep learning for CSI feedback: One-sided model and joint multi-module learning perspectives,

    Y . Guo, W. Chen, F. Sun, J. Cheng, M. Matthaiou, and B. Ai, “Deep learning for CSI feedback: One-sided model and joint multi-module learning perspectives,”arXiv preprint arXiv:2405.05522, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Quantized trainable compressed sensing for MIMO CSI feedback,

    H. Shao, H. Zhang, W. Zhang, and X. Zhang, “Quantized trainable compressed sensing for MIMO CSI feedback,”IEEE Trans. on V ehicular Technology, 2024

    2024

  11. [11]

    A deep learning-based approach to lightweight CSI feedback,

    Y . An, S. Lu, H. Cai, and Z. Ji, “A deep learning-based approach to lightweight CSI feedback,”Physical Communication, vol. 68, p. 102538, 2025

    2025

  12. [12]

    Continuous online learning- based CSI feedback in massive MIMO systems,

    X. Zhang, J. Wang, Z. Lu, and H. Zhang, “Continuous online learning- based CSI feedback in massive MIMO systems,”IEEE Commun. Letters, 2024

    2024

  13. [13]

    Channel state information prediction for 5g wireless communications: A deep learning approach,

    C. Luo, J. Ji, Q. Wang, X. Chen, and P. Li, “Channel state information prediction for 5g wireless communications: A deep learning approach,” IEEE Trans. on network science and engineering, vol. 7, no. 1, pp. 227–236, 2018

    2018

  14. [14]

    Trans- former network based channel prediction for csi feedback enhancement in ai-native air interface,

    T. Zhou, X. Liu, Z. Xiang, H. Zhang, B. Ai, L. Liu, and X. Jing, “Trans- former network based channel prediction for csi feedback enhancement in ai-native air interface,”IEEE Trans. on Wireless Commun., vol. 23, no. 9, pp. 11 154–11 167, 2024

    2024

  15. [15]

    Ai for csi prediction in 5g-advanced and beyond,

    J. Chengyong, G. Jiajia, L. Xiangyi, J. Shi, and Z. Jun, “Ai for csi prediction in 5g-advanced and beyond,”China Commun., vol. 22, no. 11, pp. 1–16, 2025

    2025

  16. [16]

    Deep learning based joint csi compression and prediction for beyond- 5g systems,

    S. Kadambar, A. T. Abebe, A. Kumar, A. K. R. Chavva, and H.-J. Ji, “Deep learning based joint csi compression and prediction for beyond- 5g systems,” inGLOBECOM 2023-2023 IEEE Global Commun. Conf. IEEE, 2023, pp. 4792–4797

    2023

  17. [17]

    A survey of model compression and acceleration for deep neural networks,

    Y . Cheng, D. Wang, P. Zhou, and T. Zhang, “A survey of model compression and acceleration for deep neural networks,”arXiv preprint arXiv:1710.09282, 2017

    Pith/arXiv arXiv 2017

  18. [18]

    Learning both weights and con- nections for efficient neural network,

    S. Han, J. Pool, J. Tran, and W. Dally, “Learning both weights and con- nections for efficient neural network,” inAdvances in neural information processing systems, 2015, pp. 1135–1143

    2015

  19. [19]

    Learning the CSI denoising and feed- back without supervision,

    V . Rizzello and W. Utschick, “Learning the CSI denoising and feed- back without supervision,” in2021 IEEE 22nd Intl. Wkshp. on Signal Processing Advances in Wireless Commun. (SPA WC). IEEE, 2021, pp. 16–20

    2021

  20. [20]

    A deep learning method for joint compression and unsupervised denoising of CSI feedback,

    T.-H. Huang, A. Malhotra, and S. Hamidi-Rad, “A deep learning method for joint compression and unsupervised denoising of CSI feedback,” in ICC 2023-IEEE Intl. Conf. on Commun.IEEE, 2023, pp. 4150–4156

    2023

  21. [21]

    Deep learning for joint channel estimation and feedback in massive mimo systems,

    J. Guo, T. Chen, S. Jin, G. Y . Li, X. Wang, and X. Hou, “Deep learning for joint channel estimation and feedback in massive mimo systems,” Digital Commun. and Networks, vol. 10, no. 1, pp. 83–93, 2024

    2024

  22. [22]

    Network-first separate training with raw dataset sharing: A training approach for AI/ML-driven CSI feedback,

    A. Saini, J. H. Kim, A. A. Tehrani, Y . Xing, and W. Gerstacker, “Network-first separate training with raw dataset sharing: A training approach for AI/ML-driven CSI feedback,” in2024 IEEE Intl. Conf. on Commun. Wkshp.IEEE, 2024, pp. 1950–1955

    2024

  23. [23]

    Federated unfolding learning for CSI feedback in distributed edge networks,

    C. Tan, D. Cai, F. Fang, Z. Ding, and P. Fan, “Federated unfolding learning for CSI feedback in distributed edge networks,”IEEE Trans. on Commun., 2024

    2024

  24. [24]

    NR AIML Air Interface CSI Compres- sion Datasets (R4 113),

    3GPP RAN4 Working Group, “NR AIML Air Interface CSI Compres- sion Datasets (R4 113),” https://www.3gpp.org/ftp/tsg ran/WG4 Radio/ Data sharing/NR AIML air/CSI compression/Datasets/R4 113, 2025, accessed: 2025-08-06

    2025

  25. [25]

    Discussions on AI-CSI,

    3GPP TSG RAN WG1, “Discussions on AI-CSI,” 3GPP, Xiamen, China, Tech. Rep. R1-2308873, 2023, 3GPP TSG RAN WG1 #114-bis Meeting

    2023

  26. [26]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  27. [27]

    WF on requirements for AI/ML air interface,

    3GPP TSG-RAN WG4 Meeting #112-bis, “WF on requirements for AI/ML air interface,” Qualcomm, Hefei, Anhui, China, Tech. Rep. R4- 2417212, Oct 14th–Oct 18th 2024

    2024

  28. [28]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,”arXiv preprint arXiv:1412.3555, 2014

    Pith/arXiv arXiv 2014

  29. [29]

    (2025) Cuda c programming guide: Cuda graphs

    NVIDIA. (2025) Cuda c programming guide: Cuda graphs. [Online]. Available: https://docs.nvidia.com/cuda/cuda-c-programming-guide/ index.html#cuda-graphs

    2025

  30. [30]

    On AI/ML Based CSI Com- pression,

    3GPP TSG-RAN WG4 Meeting #113, “On AI/ML Based CSI Com- pression,” Nokia, Orlando, US, Tech. Rep. R4-2419178, Nov 18th–Nov 22nd 2024, agenda item: 7.17.2.1, Document for: Discussion

    2024

  31. [31]

    Tinyml nlp scheme for semantic wireless sentiment classification with privacy preservation,

    A. Y . Radwan, M. Shehab, and M.-S. Alouini, “Tinyml nlp scheme for semantic wireless sentiment classification with privacy preservation,” in 2025 Joint European Conf. on Networks and Commun. & 6G Summit (EuCNC/6G Summit). IEEE, 2025, pp. 133–138

    2025