Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Synesthesia of Machines (SoM)-Empowered Wireless Image Transmission over Complex Dynamic Channel

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Wireless image codec that adapts to fast-fading channels beats JSCC baselines

desk verdict Useful JSCC adaptation for time-selective fading, but a load-bearing sign error in Eq. (20) and missing error bars/code make the strong claims shaky. read the letter →

arxiv 2509.11243 v1 pith:X7DC63XV submitted 2025-09-14 eess.SP

Haozhen Li , Ruide Zhang , Rongqing Zhang , Xiang Cheng This is my paper · ORCID
classification eess.SP
keywords wirelessimagetransmissionjointsource-channelcodingtime-selectivefadingchannelagingSwinTransformersemanticcommunicationphysical-layersideinformationadaptation
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

The paper proposes DCAT, a joint source-channel coding scheme that uses physical-layer side information—SNR, CSI estimates, Doppler shift, and channel aging delay—to adapt image features during transmission. It claims that this tight coupling makes DCAT consistently outperform existing JSCC baselines under time-selective fading and channel aging, with at least 10.5% PSNR improvement and 12.6% LPIPS reduction at identical communication overhead. The architecture builds on a Swin Transformer backbone and adds two lightweight modules: DC-attn, which scales feature channels based on channel state, and DC-permu, which reorders feature-channel symbols to place important information in more favorable channel moments. A three-stage training strategy progressively transfers knowledge from noiseless to AWGN to dynamic channels. If correct, this offers a path to deploying learned image transmission in realistic, high-mobility wireless conditions.

What carries the argument

Two complementary adaptive modules inserted into a Swin Transformer codec. DC-attn (Dynamic Channel-Attention) converts side information into a per-stage attention vector that scales feature-channels, with an Attention Ratio Adjustment module that modulates confidence in CSI using aging delay and Doppler. DC-permu (Dynamic Channel-Permutation) generates a per-time-instant score from jointly embedded side information, sorts those scores, and permutes the real and imaginary components of the C feature-channels accordingly; the decoder applies the inverse permutation. Both modules are supervised by physically meaningful regularization terms: CSI NMSE for DC-attn and a tanh-mapped symbol impairm

What would settle it

Run DCAT over an over-the-air testbed (or with deliberately corrupted/outdated CSI feedback) where the true channel state differs from the side information used by the encoder and decoder. If the PSNR gain over a non-adaptive SwinT JSCC baseline falls below the reported ~10.5% under aging conditions—or if LPIPS degradation grows beyond the reported 12.6%—the central claim of robust real-world adaptation is falsified.

Watch

Extended reading notes

Core claim

The central claim is that image codecs should be aware of the physical-layer transmission process, not just the noise level. DCAT encodes the image into a token sequence, maps it to complex symbols, and transmits it over a time-selective fading channel with periodic pilot-based CSI that may be stale. Using the causal side information set P = {SNR, CSI estimate, Doppler, aging delay}, the encoder and decoder jointly learn to adapt: DC-attn generates an attention vector that globally scales feature-channels according to channel quality, while DC-permu scores each transmission moment and permutes feature-channels so that the most critical information is transmitted when the channel is most reli

Load-bearing premise

The reported gains rely on the transmitter and receiver having causal access to accurate physical-layer side information—SNR, CSI estimates, Doppler, and aging delay—and on the simulated channel generator faithfully representing real propagation; if the side information is stale, mismatched, or the real channel deviates from the simulated statistics, the adaptive modules may misalign and the gains could diminish or vanish.

Editorial extensions

If this is right

  • DCAT maintains or improves its advantage as compression ratio grows, because longer transmission durations expose more structure in time-selective fading and periodic aging that the modules can exploit.
  • Under severe channel aging (high velocity), the permutation module contributes substantially more than attention, since stale CSI makes global scaling less reliable while per-symbol reordering still protects important features.
  • The physical-layer side information embedding (Doppler and aging delay, not just SNR/CSI) is critical: ablating it costs an average 1.206 dB PSNR and 26.6% LPIPS degradation under aging.
  • A fine-tuned DCAT with a frozen backbone surpasses fully retrained baselines on an unseen dataset and an unseen channel configuration, indicating transferable, interpretable adaptation.
  • The architecture adds only modest overhead: 13.66M parameters and 210.18 GFLOPs, comparable to the plain SwinT backbone while outperforming heavier baselines.

Reading between the lines

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

  • If the side-information assumption holds, the same design principle could extend to MIMO or OFDM systems, where per-subcarrier or per-stream channel state would give even richer signals for attention and permutation to exploit.
  • A natural next test is whether DC-permu's learned ordering approaches a known optimality criterion, such as matching feature importance (e.g., gradient-based saliency) to instantaneous channel quality; such a connection would strengthen the interpretability claim.
  • The paper's gains are measured on simulated QuaDRiGa channels; a field trial with real pilot-based CSI and mobile receivers would test whether the aging-delay model in Eq. (8) captures enough real-world dynamics to preserve the reported 10.5% PSNR advantage.
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

3 major / 5 minor

Summary. The paper proposes DCAT, a joint source-channel coding (JSCC) scheme for wireless image transmission over time-selective fading channels with channel aging. It builds on a Swin Transformer backbone and adds two adaptive modules: DC-attn, which re-weights feature channels using physical-layer side information (SNR, CSI, Doppler, aging delay) and includes an Attention Ratio Adjustment (ARA) module, and DC-permu, which permutes feature channels before modulation according to per-symbol channel-quality scores. Training uses a three-stage transfer-learning strategy with auxiliary losses based on CSI NMSE and symbol impairment. Experiments on UDIS-D and PLACES365 with QuaDRiGa 3GPP channel configurations compare DCAT against SwinT, ADJSCC, WITT, and SwinT-L, reporting consistent PSNR/LPIPS gains, ablations across velocities, and a scalability/fine-tuning study.

Significance. If the reported results survive correction of the specification issues below, DCAT is a valuable step beyond JSCC evaluations that treat the channel as AWGN or Rayleigh noise: it explicitly couples image coding with physical-layer effects (Doppler, pilot-based aging, stale CSI) and demonstrates the benefit of side-information-driven adaptation. The paper's strengths include a clearly described system model, a realistic QuaDRiGa-based evaluation, informative ablations (Tables V, VI and the 'w/o embed' row), and a transfer-learning/scalability study. The use of ground-truth CSI and impairment labels in auxiliary losses is acceptable because these labels are used only for training and are not needed at inference. However, the physics-guided interpretation is currently undermined by an internal inconsistency in the Tx-side aging delay feature and by an ill-defined auxiliary loss for the permutation module; these must be corrected before the central claims can be accepted.

major comments (3)
  1. [Sec. III-B, Eq. (20)] The Tx-side aging delay is mis-specified. With \hat H_tx(t)=\hat H(t_beg) (Eq. (19)) and T[k]>t_beg, the age of the Tx CSI at symbol time T[k] is \tau_ag(t_beg)+(T[k]-t_beg), which increases with k. Eq. (20) instead uses \tau_ag(t_beg)+(t_beg-T[k]), which decreases and becomes negative for most k; in the perfect-CP case (\tau_ag=0) it is negative for every k>1, which is unphysical. Because the ARA module combines this feature with f_D to produce δ, and Eq. (22) trains δ against CSI NMSE (which grows with actual age), the module's input contradicts the paper's own aging model in Eqs. (7)-(8). The claimed physics-guided adaptation can therefore not be attributed to the design as written. Please correct the sign and verify whether the experimental code used the corrected value.
  2. [Sec. III-C, Eqs. (26)-(27)] L_permu is not well defined as written. The text states that Score[k] is generated for each discrete time instant k, so Score is a length-K vector. Eq. (26) defines Imp(z,\hat z) as a single scalar obtained by summing over all K real and imaginary channels. Eq. (27) then writes |Score - Imp(z,\hat z)|, which is a vector-minus-scalar expression and does not define a scalar loss. Since L_permu is the only direct training signal for the DC-permu scoring network, this needs to be a per-index impairment, e.g., Imp_k = tanh(||z_re[k]-\hat z_re[k]||_2^2 + ||z_im[k]-\hat z_im[k]||_2^2) and L_permu = (1/K) \sum_k |Score[k]-Imp_k|.
  3. [Sec. V-A / Abstract] The headline empirical claim rests on single-run comparisons without error bars or significance tests. For example, the CR=1/12 gain over the suboptimal baseline is 0.089 dB (Sec. V-A), which is small relative to typical run-to-run variability in deep JSCC training. The abstract's 'across all conditions' also exceeds the tested scope (SISO narrowband, QuaDRiGa UMa/RMa, nine SNR values, five velocities). Please report mean±std over multiple seeds and qualify the claim to the tested conditions. The percentage wording is also misleading: a 0.431 dB PSNR gain corresponds to roughly 10.4% MSE reduction, not '10.5% PSNR improvement.'
minor comments (5)
  1. [Sec. III-B, Eq. (22)] The auxiliary loss uses ground-truth CSI H_gt. Please state explicitly that H_gt is available only during training and is not needed at inference, so the auxiliary loss is not circular with respect to the test-time claims.
  2. [Sec. II-B, Eq. (3)] E_n is used in the SNR definition but never defined. Also, the scaling constant λ in f_D = λ f_c v / c appears arbitrary; please either define it or use the standard normalized Doppler expression.
  3. [Fig. 5] The ARA module output δ is not specified to be bounded. If δ is used in (1−δ) or (1+δ) weighting, specify its normalization/clipping and how values above 1 (c.f. NMSE>1 in Table II) are handled.
  4. [Sec. III-C, Eq. (25)] DC-permu uses \hat H_tx[k]=\hat H(T[1]) for both encoder and decoder while \tau_permu[k]=\tau_ag(T[k]) is the age of the Rx equalization CSI. This mixes two different reference times; please clarify whether this is an intentional approximation and what information is lost.
  5. [General] Reproducibility would be substantially improved by releasing code and training seeds. In addition, there are minor typos: 'Comperssion Ratio' in Figs. 7 and 8, 'Velocitiy' in Figs. 10 and 11, and 'Additionly' in Sec. V-D.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: DCAT's gains are empirical performance comparisons, and the side-information regularizers are supervised auxiliary losses rather than outputs derived from the model.

full rationale

The paper does not derive its performance from its inputs; it measures image reconstruction quality against external baselines (SwinT, ADJSCC, WITT, SwinT-L) under QuaDRiGa/3GPP channels. The regularization terms in Eqs. (22) and (27) train the CSI-confidence head and the permutation scorer to match known labels (CSI NMSE and end-to-end symbol impairment) during training; these are supervised auxiliary tasks, not circular predictions. The SoM concept is cited from the authors' prior work [29], [30] as motivation, but it does no technical work: removing the SoM label leaves the DC-attn/DC-permu architecture and all experiments unchanged. No equation reduces by construction to its inputs. The reviewer's concern about the sign of tau_attn in Eq. (20) is a potential modeling correctness issue, not a circularity, and therefore does not affect the circularity score.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical performance comparison, not a derivation from first principles. The main assumptions are the fidelity of the simulated dynamic channel, availability of physical-layer side information, and the validity of the straight-through gradient approximation. The hand-chosen loss weights are free parameters that could affect the quantitative results. No new physical entities are introduced.

free parameters (2)
  • loss regularization weights omega_1, omega_2, eta = 0.5, 0.5, 5.0
    Hand-chosen in Eq. (28); no sensitivity analysis is provided, so reported gains could depend on these values.
  • perceptual loss mixing schedule coefficients = mu1 = (5*SNR+42)/72, mu2 = 85/24 * (-5*SNR+108)
    Hand-designed SNR-dependent weights in Eq. (31); not justified by experiment.
assumptions (4)
  • domain assumption The channel is narrowband with no ISI, Y(t)=H(t)X(t)+N(t), and ZF equalization with estimated CSI is sufficient.
    Sec. II-A, Eqs. (1)-(6). If frequency selectivity or imperfect equalization is present, the model and results change.
  • domain assumption QuaDRiGa with 3GPP TR 38.901 parameterizations faithfully represents real dynamic wireless channels for training and evaluation.
    Sec. IV-A. The whole value of the results depends on this simulation fidelity.
  • standard math The straight-through estimator in Eq. (15) provides valid gradients for training despite ignoring the channel's dependence on the transmitted symbol z.
    Sec. II-B3. This is a widely used approximation but can bias learning under strong nonlinearities.
  • domain assumption Side information P is causally available at Tx and Rx as specified in Eq. (9), including CSI estimates and aging delay.
    Sec. II-B. If feedback is delayed or unavailable, the DC-attn and DC-permu modules may be misaligned with actual channel conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Synesthesia of Machines (SoM)-Empowered Wireless Image Transmission over Complex Dynamic Channel." pith.science (2026). https://pith.science/paper/X7DC63XV

@misc{pith2026250911243,
  author       = {Pith},
  title        = {Pith review of: Synesthesia of Machines (SoM)-Empowered Wireless Image Transmission over Complex Dynamic Channel},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X7DC63XV}},
  note         = {Machine review of arXiv:2509.11243}
}
read the original abstract

Wireless image transmission underpins diverse networked intelligent services and becomes an increasingly critical issue. Existing works have shown that deep learning-based joint source-channel coding (JSCC) is an effective framework to balance image transmission fidelity and data overhead. However, these studies oversimplify the communication system as a mere pipeline with noise, failing to account for the complex dynamics of wireless channels and concrete physical-layer transmission process. To address these limitations, we propose a Synesthesia of Machines (SoM)-empowered Dynamic Channel Adaptive Transmission (DCAT) scheme, designed for practical implementation in real communication scenarios. Building upon the Swin Transformer backbone, our DCAT scheme demonstrates robust adaptability to time-selective fading and channel aging effects by effectively utilizing the physical-layer transmission characteristics of wireless channels. Comprehensive experimental results confirm that DCAT consistently achieves superior performance compared with JSCC baseline approaches across all conditions. Furthermore, our neural network architecture demonstrates high scalability due to its interpretable design, offering substantial potential for cost-efficient deployment in practical applications.

Figures

Figures reproduced from arXiv: 2509.11243 by the authors.

Figure 2
Figure 2. The model of the point-to-point image transmission system with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overall architecture of DCAT, our proposed wireless image transmission scheme over the complex dynamic channel. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. The processing flow within a single stage of the encoder. SwinT blocks [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (8 more)
Figure 5
Figure 5. Figure 5: Diagram of the Dynamic Channel-Attention modules. Each embedding [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Diagram of the Dynamic Channel-Permutation modules. Each embedding submodule is an MLP employing Leaky ReLU for hidden layer activation. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Comparison of PSNR performance between the proposed DCAT and baseline schemes. (a) versus SNR under [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Comparison of LPIPS performance between the proposed DCAT and baseline schemes. (a) versus SNR under [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Visualization comparison under Aging Scenario. All schemes under each SNR undergo the same channel aging condition with a fixed CR of 1/6. The performance gain increases from 0.0029 (CR=1/12) to 0.0079 (CR=1/3) with longer image transmission duration. B. Qualitative Re…
Figure 10
Figure 10. Figure 10: The scalability experiment on the PLACES365 image dataset, [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: The scalability experiment on the Case B channel configuration, averaged the metrics across all tested SNRs. certain threshold, the combined effects of time-selective fading and channel aging significantly degrade transmission quality, leading to a rapid escalation in…
Figure 12
Figure 12. Figure 12: The performance of different transfer learning strategies. All tests [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 4 linked inside Pith

  1. [1]

    Semantic-aware scene recognition,

    A. L ´opez-Cifuentes, M. Escudero-Vinolo, J. Besc ´os, and ´A. Garc ´ıa- Mart´ın, “Semantic-aware scene recognition,”Pattern Recognit., vol. 102, p. 107256, Jun. 2020

    2020

  2. [2]

    VQA: Visual question answering,

    S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, and D. Parikh, “VQA: Visual question answering,” inProc. ICCV, Santiago, Chile, Dec. 2015, pp. 2425–2433

    2015

  3. [3]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” inProc. CVPR, Las Vegas, NV , USA, Jun. 2016, pp. 779–788

    2016

  4. [4]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,”The Bell Syst. Tech. J., vol. 27, no. 3, pp. 379–423, Oct. 1948

    1948

  5. [5]

    Design of low-density parity check codes for 5G new radio,

    T. Richardson and S. Kudekar, “Design of low-density parity check codes for 5G new radio,”IEEE Commun. Mag., vol. 56, no. 3, pp. 28–34, Mar. 2018

    2018

  6. [6]

    Channel polarization: A method for constructing capacity- achieving codes for symmetric binary-input memoryless channels,

    E. Arikan, “Channel polarization: A method for constructing capacity- achieving codes for symmetric binary-input memoryless channels,”IEEE Trans. Inf. Theory, vol. 55, no. 7, pp. 3051–3073, Jul. 2009

    2009

  7. [7]

    Joint source and channel coding,

    M. Fresia, F. Per ´ez-Cruz, H. V . Poor, and S. Verd ´u, “Joint source and channel coding,”IEEE Signal Process Mag., vol. 27, no. 6, pp. 104–113, Nov. 2010

    2010

  8. [8]

    Deep joint source- channel coding for wireless image transmission,

    E. Bourtsoulatze, D. B. Kurka, and D. G ¨und¨uz, “Deep joint source- channel coding for wireless image transmission,”IEEE Trans. Cognit. Commun. Networking, vol. 5, no. 3, pp. 567–579, Sept. 2019

    2019

Show all 39 references
  1. [9]

    Deep joint source-channel coding for person re-identification with related image sources,

    Z. Chu, W. Chen, N. Wang, and B. Ai, “Deep joint source-channel coding for person re-identification with related image sources,” inProc. IEEE Int. Conf. Commun. Technol. (ICCT), Wuxi, China, Oct. 2023, pp. 501–505

    2023

  2. [10]

    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 J. Sel. Areas Commun., vol. 40, no. 8, pp. 2300–2316, Aug. 2022

    2022

  3. [11]

    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 Trans. Wireless Commun., vol. 23, no. 10, pp. 15 002– 15 017, Oct. 2024

    2024

  4. [12]

    WITT: A wireless image transmission transformer for semantic communications,

    K. Yang, S. Wang, J. Dai, K. Tan, K. Niu, and P. Zhang, “WITT: A wireless image transmission transformer for semantic communications,” inProc. ICASSP, Rhodes Island, Greece, Jun. 2023

    2023

  5. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  6. [14]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProc. ICCV, Montreal, QC, Canada, Oct. 2021, pp. 10 012– 10 022

    2021

  7. [15]

    Generative joint source-channel coding for semantic image transmission,

    E. Erdemir, T.-Y . Tung, P. L. Dragotti, and D. G¨und¨uz, “Generative joint source-channel coding for semantic image transmission,”IEEE J. Sel. Areas Commun., vol. 41, no. 8, pp. 2645–2657, Aug. 2023

    2023

  8. [16]

    Diffusion-aided joint source channel coding for high realism wireless image transmission,

    M. Yang, B. Liu, B. Wang, and H.-S. Kim, “Diffusion-aided joint source channel coding for high realism wireless image transmission,”arXiv preprint arXiv:2404.17736, 2024

    2024 arXiv

  9. [17]

    Latent feature- guided conditional diffusion for high-fidelity generative image semantic communication,

    Z. Chen, X. Wei, H. Tong, Z. Yang, and C. Yin, “Latent feature- guided conditional diffusion for high-fidelity generative image semantic communication,”arXiv preprint arXiv:2504.21577, 2025

    2025 arXiv

  10. [18]

    Recent contributions to the mathematical theory of com- munication,

    W. Weaver, “Recent contributions to the mathematical theory of com- munication,”ETC Rev. Gen. Semant., vol. 10, no. 4, pp. 261–281, 1953

    1953

  11. [19]

    Semantic communications for future internet: Fundamentals, applications, and challenges,

    W. Yang, H. Du, Z. Q. Liew, W. Y . B. Lim, Z. Xiong, D. Niyato, X. Chi, X. Shen, and C. Miao, “Semantic communications for future internet: Fundamentals, applications, and challenges,”IEEE Commun. Surv. Tutorials, vol. 25, no. 1, pp. 213–250, 1st Quart. 2023

    2023

  12. [20]

    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 Trans. Circuits Syst. Video Technol., vol. 32, no. 4, pp. 2315–2328, Apr. 2022

    2022

  13. [21]

    SNR-adaptive deep joint source- channel coding for wireless image transmission,

    M. Ding, J. Li, M. Ma, and X. Fan, “SNR-adaptive deep joint source- channel coding for wireless image transmission,” inProc. ICASSP, Toronto, ON, Canada, Jun. 2021, pp. 1555–1559

    2021

  14. [22]

    SwinJSCC: Taming swin transformer for deep joint source-channel coding,

    K. Yang, S. Wang, J. Dai, X. Qin, K. Niu, and P. Zhang, “SwinJSCC: Taming swin transformer for deep joint source-channel coding,”IEEE Trans. Cognit. Commun. Networking, vol. 11, no. 1, pp. 90–104, Feb. 2025

    2025

  15. [23]

    Learned image transmission over MIMO fading channels,

    S. Yao, S. Wang, J. Dai, and K. Niu, “Learned image transmission over MIMO fading channels,” inProc. PIMRC, Toronto, ON, Canada, Sept. 2023, pp. 1–6

    2023

  16. [24]

    SCAN: Semantic communica- tion with adaptive channel feedback,

    G. Zhang, Q. Hu, Y . Cai, and G. Yu, “SCAN: Semantic communica- tion with adaptive channel feedback,”IEEE Trans. Cognit. Commun. Networking, vol. 10, no. 5, pp. 1759–1773, Oct. 2024

    2024

  17. [25]

    Deep learning channel estimation for OFDM 5G systems with different channel models,

    A. S. M. Mohammed, A. I. A. Taman, A. M. Hassan, and A. Zekry, “Deep learning channel estimation for OFDM 5G systems with different channel models,”Wireless Pers. Commun., vol. 128, no. 4, pp. 2891– 2912, Oct. 2023

    2023

  18. [26]

    WiFo: Wireless foundation model for channel prediction,

    B. Liu, S. Gao, X. Liu, X. Cheng, and L. Yang, “WiFo: Wireless foundation model for channel prediction,”Sci. China Inf. Sci., vol. 68, no. 6, pp. 1–13, May 2025

    2025

  19. [27]

    CSI-LLM: A novel downlink channel prediction method aligned with LLM pre-training,

    S. Fan, Z. Liu, X. Gu, and H. Li, “CSI-LLM: A novel downlink channel prediction method aligned with LLM pre-training,” inProc. WCNC, Milan, Italy, Mar. 2025, pp. 1–6

    2025

  20. [28]

    Effects of channel aging in massive MIMO systems,

    K. T. Truong and R. W. Heath, “Effects of channel aging in massive MIMO systems,”J. Commun. Networks, vol. 15, no. 4, pp. 338–351, Aug. 2013

    2013

  21. [29]

    Intelligent multi-modal sensing-communication integration: Synesthesia of Machines,

    X. Cheng, H. Zhang, J. Zhang, S. Gao, S. Li, Z. Huang, L. Bai, Z. Yang, X. Zheng, and L. Yang, “Intelligent multi-modal sensing-communication integration: Synesthesia of Machines,”IEEE Commun. Surv. Tutorials, vol. 26, no. 1, pp. 258–301, 1st Quart. 2024

    2024

  22. [30]

    Foundation model empowered Synesthesia of Machines (SoM): AI-native intelligent multi- modal sensing-communication integration,

    X. Cheng, B. Liu, X. Liu, E. Liu, and Z. Huang, “Foundation model empowered Synesthesia of Machines (SoM): AI-native intelligent multi- modal sensing-communication integration,”IEEE Trans. Network Sci. Eng., pp. 1–21, early access 2025

    2025

  23. [31]

    Estimating or propagating gradients through stochastic neurons for conditional computation,

    Y . Bengio, N. L ´eonard, and A. Courville, “Estimating or propagating gradients through stochastic neurons for conditional computation,”arXiv preprint arXiv:1308.3432, 2013

    2013 arXiv

  24. [32]

    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,” inAdv. Neural Inf. Process. Syst. (NeurIPS), vol. 30, Long Beach, CA, USA, Oct. 2017, pp. 6000–6010

    2017

  25. [33]

    Enhanced AI-based CSI prediction solutions for massive MIMO in 5G and 6G systems,

    D. Burghal, Y . Li, P. Madadi, Y . Hu, J. Jeon, J. Cho, A. F. Molisch, and J. Zhang, “Enhanced AI-based CSI prediction solutions for massive MIMO in 5G and 6G systems,”IEEE Access, vol. 11, pp. 117 810– 117 825, Oct. 2023

    2023

  26. [34]

    Unsupervised deep image stitching: Reconstructing stitched features to images,

    L. Nie, C. Lin, K. Liao, S. Liu, and Y . Zhao, “Unsupervised deep image stitching: Reconstructing stitched features to images,”IEEE Trans. Image Process., vol. 30, no. 7, pp. 6184–6197, Jul. 2021

    2021

  27. [35]

    Places: A 10 million image database for scene recognition,

    B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba, “Places: A 10 million image database for scene recognition,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 40, no. 6, pp. 1452–1464, Jun. 2018

    2018

  28. [36]

    QuaDRiGa: A 3- D multi-cell channel model with time evolution for enabling virtual field trials,

    S. Jaeckel, L. Raschkowski, K. B ¨orner, and L. Thiele, “QuaDRiGa: A 3- D multi-cell channel model with time evolution for enabling virtual field trials,”IEEE Trans. Antennas Propag., vol. 62, no. 6, pp. 3242–3256, Jun. 2014

    2014

  29. [37]

    Study on channel model for frequencies from 0.5 to 100 GHz (Release 15),

    3GPP Radio Access Network Working Group, “Study on channel model for frequencies from 0.5 to 100 GHz (Release 15),” 3GPP, Tech. Rep. TR 38.901, 2018

    2018

  30. [38]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” inAdv. Neural Inf. Process. Syst. (NeurIPS), vol. 25, Lake Tahoe, NV , USA, Dec. 2012

    2012

  31. [39]

    The 14 unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The 14 unreasonable effectiveness of deep features as a perceptual metric,” in Proc. CVPR, Salt Lake City, UT, USA, Jun. 2018, pp. 586–595

    2018

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.