Pith. sign in

REVIEW 4 major objections 6 minor 4 cited by

BERT4MIMO: A Foundation Model using BERT Architecture for Massive MIMO Channel State Information Prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a BERT-style transformer, trained on masked synthetic massive MIMO channel matrices, reconstructs the full CSI with an average MSE of 0.011035, roughly 30 times lower than linear regression or a multilayer…

desk verdict BERT4MIMO applies BERT-style masking to CSI but its headline MSE is contradicted by its own table and the test protocol is unspecified. read the letter →

arxiv 2501.01802 v1 pith:ECSFY4VN submitted 2025-01-03 cs.IT cs.AIeess.SPmath.IT

Ferhat Ozgur Catak , Murat Kuzlu , Umit Cali This is my paper · ORCID
classification cs.ITcs.AIeess.SPmath.IT
keywords massiveMIMOchannelstateinformationCSIreconstructionBERTtransformerfoundationmodelmaskingwirelessprediction
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

BERT4MIMO is a BERT-inspired transformer that predicts missing entries in massive MIMO channel state information (CSI) matrices. The paper claims it reconstructs masked CSI with an average MSE of 0.011035, about 30 times better than linear regression (0.309207) and an MLP (0.314465) on the same task. If true, this would show that a masked-language-model-style approach transfers to wireless channel estimation, letting one model handle missing data across different mobility and channel conditions. The authors also report that performance stays nearly flat for masking ratios up to 50% and degrades only slightly with Doppler shifts up to 400 Hz, suggesting robustness to realistic channel impairments. The result matters because accurate CSI prediction underpins beamforming, spectral efficiency, and reliability in 5G and 6G systems.

What carries the argument

The key machinery is BERT-style masked autoencoding. A random binary mask blanks out a fraction of the CSI entries, and the transformer must fill them in using self-attention over all subcarrier positions and feature dimensions. The paper combines a learnable temporal embedding (indexing subcarriers or time steps) with a feature embedding produced by a linear projection, and sums them before the encoder. The self-attention mechanism lets each masked position draw on global context from every other position, which is what gives the model its edge over linear regression and MLP, which lack this global contextual mechanism. The architecture is described as a foundation model because the same pretrained encoder can, in principle, be fine-tuned for downstream wireless tasks such as beamforming, spectrum sensing, or channel estimation.

What would settle it

Train BERT4MIMO on the released synthetic dataset and evaluate it on a real measured massive MIMO CSI dataset, or on a held-out portion of the synthetic data generated with different delay spreads and SNRs. If the reconstruction MSE rises to the baseline level of about 0.3 rather than staying near 0.011, then the central performance claim is refuted.

Watch

Extended reading notes

Core claim

The central discovery is that a transformer encoder with BERT-style masked training can reconstruct massive MIMO CSI from partially observed inputs. The model splits each CSI matrix into real and imaginary parts, normalizes them, adds learnable temporal and feature embeddings, and feeds the sequence through a 12-layer, 12-head transformer encoder. A fully connected head then outputs the full CSI, and training minimizes mean squared error. On a synthetic dataset of 6,000 matrices generated with the MATLAB 5G Toolbox tapped-delay-line models (stationary, high-speed mobility, and urban macrocell scenarios), the model reaches an average reconstruction MSE of 0.011035. This is approximately 30 times lower than linear regression and MLP baselines, the paper reports. Scenario-wise, stationary and high-speed channels are reconstructed at MSE about 0.0032, while urban macrocell is harder at 0.0266, and cross-scenario tests show that models trained on one scenario still reconstruct the others at similar errors, with urban macrocell consistently worst.

Load-bearing premise

The load-bearing premise is that the simulated tapped-delay-line channels from the MATLAB 5G Toolbox are representative of real massive MIMO channels, and that the reported MSE is measured on data the model did not train on.

Editorial extensions

If this is right

  • If the claim holds, a single masked-CSI transformer can serve as a general-purpose CSI estimator, replacing task-specific models that need retraining per scenario.
  • The reported robustness to masking ratios up to 0.5 suggests the model could handle realistic partial feedback or packet loss in CSI reporting without retraining.
  • Cross-scenario results imply that a model pretrained on one environment transfers to another with little loss, except for urban macrocell, which may need fine-tuning.
  • A 30x improvement over linear regression and MLP, if reproduced, would make transformer-based CSI reconstruction the clear baseline for future work on this task.
  • The graceful degradation with Doppler shift implies the model remains usable for high-mobility users, though accuracy slowly decreases as velocity increases.

Reading between the lines

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

  • The synthetic TDL-based dataset may not capture real-world non-stationarities, interference, or hardware impairments, so the true performance gain on measured channels remains unknown; a real-channel test would be the natural next step.
  • The strong masking-ratio robustness is likely a direct consequence of the attention mechanism pooling information across all subcarriers, which suggests the model could be adapted to other structured missing-data problems in wireless, such as pilot decontamination or compressed sensing.
  • The comparison to a single 512-unit MLP is a limited baseline; a larger or better-tuned MLP might narrow the gap, though the architectural advantage of attention over fixed-size context windows is likely to persist.
  • Because the reported numbers come from the same simulation pipeline used for training, an independent held-out test set or a measured dataset is needed before the 30x claim can be generalized beyond simulation.
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.

Referee Report

4 major / 6 minor

Summary. The paper proposes BERT4MIMO, a BERT-style transformer encoder for reconstructing masked massive MIMO channel state information (CSI) matrices. CSI data is generated synthetically with the MATLAB 5G Toolbox using TDL channel models for three scenarios: stationary, high-speed mobility, and urban macrocell. The model embeds subcarrier and feature information, passes the combined representation through a 12-layer transformer encoder, and reconstructs the normalized real and imaginary CSI components with an MSE loss. The authors report an overall reconstruction MSE of 0.011035, about 30 times lower than linear regression and MLP baselines, and additionally report scenario-wise, subcarrier-wise, masking-ratio, cross-scenario, and Doppler-shift sensitivity experiments. The central claim is that BERT4MIMO provides superior CSI reconstruction performance across varying mobility scenarios and channel conditions.

Significance. If the reported performance were established on a sound experimental protocol, a BERT-style foundation model for CSI reconstruction could be a useful contribution to AI-native wireless physical-layer processing. The paper's strengths include the release of source code on GitHub, a clearly described synthetic dataset generation process, and a straightforward, reproducible architecture. However, the experimental evaluation as written contains load-bearing gaps: no held-out test split is specified, the aggregate MSE is numerically inconsistent with the per-subcarrier breakdown, and the cross-scenario results are nearly invariant to the training scenario. These issues prevent the reader from assessing whether the claimed superiority over baselines reflects genuine reconstruction ability or artifacts of the evaluation setup. The significance of the claimed result therefore cannot be established from the manuscript in its current form.

major comments (4)
  1. [V.A / Algorithm 1] No train/test split is described anywhere in the paper. Section V.A defines the reconstruction experiment by masking every 10th sample (Eq. 33), but this does not create a holdout set; Algorithm 1 only describes training. All reported MSE values may therefore be computed on training data, in which case the low error and the claimed factor-of-30 advantage over baselines could reflect memorization rather than predictive reconstruction. A proper evaluation must specify how the 6,000 CSI matrices were split into disjoint training, validation, and test sets, and all reported metrics must be computed on the test set.
  2. [Table III vs. V.A/Table V] The reported overall MSE of 0.011035 is numerically inconsistent with the per-subcarrier-group MSEs in Table III. Groups of 8 subcarriers each have MSEs of 0.012956 (group 0-7) and roughly 0.075-0.081 (groups 8-63); since the groups are equal in size, the unweighted average is approximately 0.069, not 0.011035. No normalization or weighting explanation is given. This discrepancy calls into question the reliability of all reported error values, including the baseline comparison in Table V that uses the same 0.011035 figure.
  3. [Table IV / Figure 4] The cross-scenario results are almost invariant to the training scenario. In Table IV, every row has the same entries to at least six decimal places (e.g., 0.003185 for the stationary test scenario regardless of whether training was stationary, high-speed, or urban macrocell), and the urban macrocell test MSE (0.026610/0.026611) also barely changes with the training scenario. Such invariance contradicts the scenario-specific differences reported in Table II and suggests that the experiment either did not actually retrain per scenario, used overlapping test sets, or evaluated a model that does not capture scenario information. The heatmap in Figure 4 also shows different values (e.g., 0.0031844) from Table IV for the same cells, further undermining the reliability of these results.
  4. [Section III.C / Eq. (14)] The preprocessing description states that each CSI matrix is normalized across the dataset using dataset-wide mean and standard deviation. Unless the normalization parameters are estimated on the training portion only, this procedure leaks test-set statistics into the model even before any training. Because no data split is defined, the normalization described in Eq. (14) is a potential additional source of circularity in all reported error metrics; the authors should clarify how normalization statistics were computed and, if a split is introduced, estimate them on the training set only.
minor comments (6)
  1. [Section V.E] There is a typo: 'folows' should be 'follows' in the introduction of the cross-scenario experiment.
  2. [Section V.E / Figure 4] The heatmap in Figure 4 displays values that differ from those in Table IV for the same entries; the authors should ensure consistent reporting between figures and tables.
  3. [Section VI] The phrase 'satisfied performance' should be 'satisfactory performance' in the observations list.
  4. [Section II.B / Eq. (12)] The channel model equation in Eq. (12) uses notation at(θp) and a†r(ϕp) but does not define the antenna array response functions; a brief definition would improve reproducibility.
  5. [References] The original BERT paper (Devlin et al.) is not cited even though the architecture is described as 'BERT-inspired'; the authors should cite the original work in Section II or IV.
  6. [Section V.A / Eq. (33)] The text says 'masking every 10th sample', but Eq. (33) sets entries to zero when i mod 10 = 0; this is consistent, though the wording 'every 10th sample' could be clarified as zeroing the 10th, 20th, etc. entries.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported CSI reconstruction MSE is the training loss on the training data; no held-out split is specified, so the central 'prediction' result reduces to a fit residual.

  1. fitted input called prediction [Algorithm 1; Section V.A, Eqs. (31)-(33); Table V]
    "Generate masks: M ∼ MaskGenerator(B); Compute predictions: Xˆ ← BERT4MIMO(X, M); Compute loss: L ← MSE(X, Xˆ) ... In this experiment, the accuracy of BERT4MIMO is evaluated by masking every 10th sample in the data and measuring the MSE ... According to the results, the model achieves an average MSE of 0.011035"

    The evaluation metric in Eq. (32) is the same MSE used as the training loss in Eq. (31). Algorithm 1 loops over the whole dataset {Hi} with randomly generated masks and updates parameters against MSE(X, Xhat); no train/test split is described anywhere in the paper. The reported reconstruction MSE, and the claimed ~30x advantage over linear regression and MLP in Table V, are therefore computed on the same samples used for fitting. The headline 'prediction' performance is the training objective evaluated on the training data, i.e., a fitted residual relabeled as a predictive result, not evidence of generalization to unseen CSI.

full rationale

The paper's central empirical claim is that BERT4MIMO offers superior CSI reconstruction, quantified by an average MSE of 0.011035. As written, this number is produced by the identical procedure used for training: Algorithm 1 generates masks, computes BERT4MIMO predictions, and minimizes MSE(X, Xhat) over the dataset; Section V.A then masks every 10th sample in 'the data' and reports the same MSE. No held-out split is stated, and no external benchmark is used. Thus the flagship result reduces to the training loss, which is the fitted-input-called-prediction pattern. The cross-scenario and baseline comparisons inherit the same in-sample issue. I do not find load-bearing self-citation or an imported uniqueness theorem: the few self-citations are contextual and not used to justify the architecture's validity. The internal inconsistency between the per-subcarrier-group MSEs (Table III, averaging about 0.069) and the overall 0.011035 is a correctness risk, not a circularity, but it reinforces that the reported number is not a reliable predictive estimate. Overall, the derivation of the architecture is not circular, but the central validation claim is partially circular because it reports the training objective on the training data as if it were a prediction.

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

The empirical claim rests on four unvalidated assumptions: synthetic TDL channels stand in for real channels; normalization does not leak test statistics; masked reconstruction MSE is treated as a meaningful system metric; and the low-rank structure enabling 50 percent masking transfers to real channels. No external benchmark is used to discharge any of these assumptions.

free parameters (4)
  • Transformer depth and heads = 12 layers, 12 heads, d_model unspecified with no ablation
    Hand-chosen architecture parameters that determine capacity and the reported MSE; no sensitivity analysis is given.
  • Evaluation masking ratio = 0.1 for main MSE; swept 0 to 0.5
    The reconstruction task and all reported errors depend on this choice; the main result uses every tenth sample masked.
  • Synthetic dataset parameters = C=10, U=200, Ns=64, Nt=64, Nr=4, delay spreads 100/300/500 ns, SNR 100 dB
    These hand-selected simulation constants generate the only dataset used; changing them would change the reported MSE.
  • MLP baseline hidden size = 512
    A single weak baseline with one hidden layer; no hyperparameter search is reported, so the 30x comparison is not a fair state-of-the-art benchmark.
assumptions (4)
  • domain assumption MATLAB 5G Toolbox TDL channel models (TDL-A, TDL-C, TDL-D) produce CSI representative of real massive MIMO channels.
    Section III generates all data with MATLAB TDL models; the paper never validates against measured channels, yet all conclusions are drawn from this synthetic data.
  • ad hoc to paper Normalization computed across the dataset does not leak test information into training.
    Equation 14 normalizes by global mean and standard deviation; no train/test split is described, so the normalization and any split (if one exists) are ambiguous.
  • domain assumption Reconstruction MSE on masked synthetic CSI is a meaningful proxy for channel prediction quality.
    Section V uses only MSE; no link is established to beamforming gain, spectral efficiency, or other system-level metrics.
  • domain assumption The masked positions can be recovered from unmasked positions in realistic channels at masking ratios up to 50 percent.
    Figure 3 shows almost constant MSE up to 50 percent masking; this assumes strong low-rank structure that may not hold in real channels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BERT4MIMO: A Foundation Model using BERT Architecture for Massive MIMO Channel State Information Prediction." pith.science (2026). https://pith.science/paper/ECSFY4VN

@misc{pith2026250101802,
  author       = {Pith},
  title        = {Pith review of: BERT4MIMO: A Foundation Model using BERT Architecture for Massive MIMO Channel State Information Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ECSFY4VN}},
  note         = {Machine review of arXiv:2501.01802}
}
read the original abstract

Massive MIMO (Multiple-Input Multiple-Output) is an advanced wireless communication technology, using a large number of antennas to improve the overall performance of the communication system in terms of capacity, spectral, and energy efficiency. The performance of MIMO systems is highly dependent on the quality of channel state information (CSI). Predicting CSI is, therefore, essential for improving communication system performance, particularly in MIMO systems, since it represents key characteristics of a wireless channel, including propagation, fading, scattering, and path loss. This study proposes a foundation model inspired by BERT, called BERT4MIMO, which is specifically designed to process high-dimensional CSI data from massive MIMO systems. BERT4MIMO offers superior performance in reconstructing CSI under varying mobility scenarios and channel conditions through deep learning and attention mechanisms. The experimental results demonstrate the effectiveness of BERT4MIMO in a variety of wireless environments.

Figures

Figures reproduced from arXiv: 2501.01802 by the authors.

Figure 1
Figure 1. An illustrative overview of BERT4MIMO’s role in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The architecture of BERT4MIMO [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Effect of masking ratio on reconstruction MSE. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Error distribution across all subcarrier groups. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Impact of Doppler Shift on reconstruction MSE. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Generalizable and Computational Efficient Channel Extrapolation for 6G: A Configurable AI-Driven Framework Built from a Modular Perspective

    eess.SP 2026-08 conditional novelty 6.0 of 10

    A three-stage modular AI framework, pretrain, cluster experts, and learn routing, improves channel extrapolation accuracy and cuts FLOPs in simulated 6G scenarios.

  2. WiFo-2: a generalist foundation model unifies heterogeneous wireless system design

    eess.SP 2025-11 unverdicted novelty 6.0 of 10

    WiFo-2 is a space-time-frequency foundation model pretrained on heterogeneous CSI data that delivers strong zero-shot and few-shot performance across wireless communications and sensing tasks.

  3. Towards channel foundation models (CFMs): Motivations, methodologies and opportunities

    eess.SP 2025-07 conditional novelty 4.0 of 10

    A survey and position paper proposing channel foundation models, with experiments on two pretrained CSI models showing gains over a vanilla ViT baseline.

  4. Foundation Model Empowered Synesthesia of Machines (SoM): AI-native Intelligent Multi-Modal Sensing-Communication Integration

    eess.SP 2025-06 conditional novelty 4.0 of 10

    The paper proposes a systematic classification and two roadmaps for using foundation models (LLMs and wireless foundation models) to design Synesthesia of Machines systems for 6G, with preliminary case-study evidence ...

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages · cited by 4 Pith papers

  1. [30]

    Finding the Missing Data: A BERT-inspired Approach Against Package Loss in Wireless Sensing

    Z. Zhao, T. Chen, F. Meng, H. Li, X. Li, and G. Zhu, “Finding the missing data: A bert-inspired approach against package loss in wireless sensing,” arXiv preprint arXiv:2403.12400 , 2024

    work page Pith review arXiv 2024

  2. [25]

    Large wireless model (lwm): A foundation model for wireless channels,

    S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (lwm): A foundation model for wireless channels,” arXiv preprint arXiv:2411.08872, 2024

    arXiv 2024

  3. [1]

    5g-advanced toward 6g: Past, present, and future,

    W. Chen, X. Lin, J. Lee, A. Toskala, S. Sun, C. F. Chiasserini, and L. Liu, “5g-advanced toward 6g: Past, present, and future,”IEEE Journal on Selected Areas in Communications , vol. 41, no. 6, pp. 1592–1619, 2023

    2023

  4. [2]

    Industrial iot in 5g-and-beyond networks: Vision, architecture, and design trends,

    A. Mahmood, L. Beltramelli, S. F. Abedin, S. Zeb, N. I. Mowla, S. A. Hassan, E. Sisinni, and M. Gidlund, “Industrial iot in 5g-and-beyond networks: Vision, architecture, and design trends,” IEEE Transactions on Industrial Informatics , vol. 18, no. 6, pp. 4122–4137, 2021

    work page 2021

  5. [3]

    Industry 5.0 is Coming: A Survey on Intelligent NextG Wireless Networks as Technological Enablers

    S. Zeb, A. Mahmood, S. A. Khowaja, K. Dev, S. A. Hassan, N. M. F. Qureshi, M. Gidlund, and P. Bellavista, “Industry 5.0 is coming: A survey on intelligent nextg wireless networks as technological enablers,” arXiv preprint arXiv:2205.09084 , 2022

    work page Pith review arXiv 2022

  6. [4]

    The roadmap to 6g security and privacy,

    P. Porambage, G. G ¨ur, D. P. M. Osorio, M. Liyanage, A. Gurtov, and M. Ylianttila, “The roadmap to 6g security and privacy,” IEEE Open Journal of the Communications Society , vol. 2, pp. 1094–1122, 2021

    2021

  7. [5]

    A streamlit- based artificial intelligence trust platform for next-generation wireless networks,

    M. Kuzlu, F. O. Catak, S. Sarp, U. Cali, and O. Gueler, “A streamlit- based artificial intelligence trust platform for next-generation wireless networks,” in 2022 IEEE Future Networks World Forum (FNWF), 2022, pp. 94–97

    work page 2022

  8. [6]

    Massive mimo,

    H. Sun, C. Ng, Y . Huo, R. Q. Hu, N. Wang, C.-M. Chen, K. Vasudevan, J. Yang, W. Montlouis, D. Ayanda, K. V . Mishra, K. Tekbiyik, N. Hus- sain, H. K. Sahoo, Y . Miao, B. Di, H. Zhang, and O. T. Demir, “Massive mimo,” in 2023 IEEE Future Networks World Forum (FNWF), 2023, pp. 1–70

    work page 2023

Show all 32 references
  1. [7]

    A literature survey on ai-aided beamforming and beam management for 5g and 6g systems,

    D. d. S. Brilhante, J. C. Manjarres, R. Moreira, L. de Oliveira Veiga, J. F. de Rezende, F. M ¨uller, A. Klautau, L. Leonel Mendes, and F. A. P. de Figueiredo, “A literature survey on ai-aided beamforming and beam management for 5g and 6g systems,” Sensors, vol. 23, no. 9, p. ...

    2023

  2. [8]

    Federated learning for distributed spectrum sensing in nextg communication networks,

    Y . Shi, Y . E. Sagduyu, and T. Erpek, “Federated learning for distributed spectrum sensing in nextg communication networks,” in Artificial Intelli- gence and Machine Learning for Multi-Domain Operations Applications IV, vol. 12113. SPIE, 2022, pp. 472–478

    2022

  3. [9]

    A cryptographic federated learning-based channel estimation for next-generation networks,

    F. O. Catak and M. Kuzlu, “A cryptographic federated learning-based channel estimation for next-generation networks,” in 2023 IEEE Virtual Conference on Communications (VCC) . IEEE, 2023, pp. 92–97

    2023

  4. [10]

    Artificial intelligence-assisted network slicing: Network assurance and service provisioning in 6g,

    J. Wang, J. Liu, J. Li, and N. Kato, “Artificial intelligence-assisted network slicing: Network assurance and service provisioning in 6g,” IEEE Vehicular Technology Magazine, vol. 18, no. 1, pp. 49–58, 2023

    2023

  5. [11]

    Automatic modulation recognition using parallel feature extraction ar- chitecture,

    H. Tang, Y . Zhao, M. Kuzlu, C. Luo, F. O. Catak, and W. Wang, “Automatic modulation recognition using parallel feature extraction ar- chitecture,” in International Conference on Wireless Artificial Intelligent Computing Systems and Applications . Springer, 2024, pp. 216–228

    2024

  6. [12]

    Security hardening of intelligent reflecting surfaces against adversarial machine learning attacks,

    F. O. Catak, M. Kuzlu, H. Tang, E. Catak, and Y . Zhao, “Security hardening of intelligent reflecting surfaces against adversarial machine learning attacks,” IEEE Access, vol. 10, pp. 100 267–100 275, 2022

    2022

  7. [13]

    A research on the new generation artificial intelligence technology generative pretraining transformer 3,

    N. Aydın and O. A. Erdem, “A research on the new generation artificial intelligence technology generative pretraining transformer 3,” in 2022 3rd International Informatics and Software Engineering Conference (IISEC). IEEE, 2022, pp. 1–6

    2022

  8. [14]

    Generative ai in medicine and healthcare: promises, opportunities and challenges,

    P. Zhang and M. N. Kamel Boulos, “Generative ai in medicine and healthcare: promises, opportunities and challenges,” Future Internet , vol. 15, no. 9, p. 286, 2023

    2023

  9. [15]

    The rise of generative artificial intelligence in healthcare,

    M. Kuzlu, Z. Xiao, S. Sarp, F. O. Catak, N. Gurler, and O. Guler, “The rise of generative artificial intelligence in healthcare,” in 2023 12th Mediterranean Conference on Embedded Computing (MECO) . IEEE, 2023, pp. 1–4

    2023

  10. [16]

    A comprehensive review of generative ai in finance,

    D. K. C. Lee, C. Guan, Y . Yu, and Q. Ding, “A comprehensive review of generative ai in finance,” FinTech, vol. 3, no. 3, pp. 460–478, 2024

    2024

  11. [17]

    Generative ai in education and research: Opportunities, concerns, and solutions,

    E. A. Alasadi and C. R. Baiz, “Generative ai in education and research: Opportunities, concerns, and solutions,” Journal of Chemical Education, vol. 100, no. 8, pp. 2965–2971, 2023

    2023

  12. [18]

    Integration of generative artificial intelligence in higher education: Pedagogy factors and best practices,

    M. Chere and N. Wayi-Mgwebi, “Integration of generative artificial intelligence in higher education: Pedagogy factors and best practices,” in Redefining Education and Development: Innovative Approaches in the Era of the Sustainable Development Goals . Springer, 2024, pp. 93–112

    2024

  13. [19]

    A survey of generative ai for intelligent transportation systems,

    H. Yan and Y . Li, “A survey of generative ai for intelligent transportation systems,” arXiv preprint arXiv:2312.08248 , 2023

    2023 arXiv

  14. [20]

    Generative ai for advanced uav networking,

    G. Sun, W. Xie, D. Niyato, H. Du, J. Kang, J. Wu, S. Sun, and P. Zhang, “Generative ai for advanced uav networking,” arXiv preprint arXiv:2404.10556, 2024

    2024 arXiv

  15. [21]

    Human- robot interaction in e-commerce: The role of personality traits and chatbot mechanisms-a neuromarketing research,

    Y .-W. Chang, S.-Y . Chien, Y .-C. Chan, and C.-C. Tsao, “Human- robot interaction in e-commerce: The role of personality traits and chatbot mechanisms-a neuromarketing research,” in Companion of the 2024 ACM/IEEE International Conference on Human-Robot Interaction, 2024, pp. 312–316

    2024

  16. [22]

    Generative artificial intelligence in manufacturing: opportunities for actualizing industry 5.0 sustainability goals,

    M. Ghobakhloo, M. Fathi, M. Iranmanesh, M. Vilkas, A. Grybauskas, and A. Amran, “Generative artificial intelligence in manufacturing: opportunities for actualizing industry 5.0 sustainability goals,” Journal of Manufacturing Technology Management , vol. 35, no. 9, pp. 94–121, 2024

    2024

  17. [23]

    Generative ai agent for next-generation mimo design: Fundamentals, challenges, and vision,

    Z. Wang, J. Zhang, H. Du, R. Zhang, D. Niyato, B. Ai, and K. B. Letaief, “Generative ai agent for next-generation mimo design: Fundamentals, challenges, and vision,” arXiv preprint arXiv:2404.08878 , 2024

    2024

  18. [24]

    Generative ai for physical layer communications: A survey,

    N. Van Huynh, J. Wang, H. Du, D. T. Hoang, D. Niyato, D. N. Nguyen, D. I. Kim, and K. B. Letaief, “Generative ai for physical layer communications: A survey,” IEEE Transactions on Cognitive Communications and Networking , 2024

    2024

  19. [26]

    Unravelling the impact of gen- erative artificial intelligence (gai) in industrial applications: A review of scientific and grey literature,

    A. K. Kar, P. Varsha, and S. Rajan, “Unravelling the impact of gen- erative artificial intelligence (gai) in industrial applications: A review of scientific and grey literature,” Global Journal of Flexible Systems Management, vol. 24, no. 4, pp. 659–689, 2023

    2023

  20. [27]

    Deep neural network based channel estimation for massive mimo-ofdm systems with imperfect channel state information,

    L. Ge, Y . Guo, Y . Zhang, G. Chen, J. Wang, B. Dai, M. Li, and T. Jiang, “Deep neural network based channel estimation for massive mimo-ofdm systems with imperfect channel state information,” IEEE Systems Journal, vol. 16, no. 3, pp. 4675–4685, 2022

    2022

  21. [28]

    An overview of enhanced massive mimo with array signal processing techniques,

    M. Wang, F. Gao, S. Jin, and H. Lin, “An overview of enhanced massive mimo with array signal processing techniques,”IEEE Journal of Selected Topics in Signal Processing , vol. 13, no. 5, pp. 886–901, 2019

    2019

  22. [29]

    Deterministic and statistical-based channel models in the mimo link evaluation,

    S. M. Ali, S. Qu, F. Kohandani, and P. Lusina, “Deterministic and statistical-based channel models in the mimo link evaluation,” IEEE Antennas and Wireless Propagation Letters , vol. 8, pp. 927–930, 2009

    2009

  23. [31]

    Temporal-channel modeling in multi-head self-attention for synthetic speech detection,

    D.-T. Truong, R. Tao, T. Nguyen, H.-T. Luong, K. A. Lee, and E. S. Chng, “Temporal-channel modeling in multi-head self-attention for synthetic speech detection,” arXiv preprint arXiv:2406.17376 , 2024

    2024 arXiv

  24. [32]

    One wide feedforward is all you need,

    T. P. Pires, A. V . Lopes, Y . Assogba, and H. Setiawan, “One wide feedforward is all you need,” arXiv preprint arXiv:2309.01826 , 2023

    2023 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.