Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

Multi-Modal Variable-Rate CSI Reconstruction for FDD Massive MIMO Systems

T0 review · 3 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read Fusing quantized CSI feedback with RGB images or uplink CSI in a variable-rate autoencoder can restore downlink channel accuracy toward ideal beamforming in 5G NR-compliant simulations.

desk verdict The variable-rate quantizer is a genuine innovation; the RGB fusion claims are plausible only in-distribution on a private synthetic dataset, so the near-optimal beamforming headline is not yet grounded for deployment. read the letter →

arxiv 2501.11926 v2 pith:USQ2TD3N submitted 2025-01-21 cs.IT eess.SPmath.IT

classification cs.ITeess.SPmath.IT
keywords massiveMIMOCSIfeedbackvariable-ratequantizationmulti-modalfusionRGBimagesuplinkFDDsystemsbeamforming
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 tries to establish that the rate-distortion bottleneck of CSI feedback in FDD massive MIMO can be pushed back by fusing the fed-back bits with auxiliary sensor data collected at the base station, specifically RGB images or uplink CSI. The proposed framework is a variable-rate autoencoder that compresses the channel into binary streams of any length between 48 and 144 bits using an ordered element-wise quantizer with only a few hundred parameters, and a transfer-learning stage attaches sensor encoders while leaving the wireless-only mode intact. If the claim holds, base stations can approach ideal beamforming gains with far less feedback overhead, which in 5G NR settings means more downlink throughput under tight reporting budgets. The supporting evidence is simulation in 5G-NR-compliant CDL channels: the wireless-only model beats Type II PMI at 76 feedback bits, and RGB fusion lifts reconstruction by up to 4.7 dB at the lowest feedback rate while reaching near-optimal LOS beamforming gains.

What carries the argument

The load-bearing mechanism is the downsample-then-interpolate variable-rate quantizer. For each of $N$ continuous channel features, the quantizer learns $2^{B_{\max}}-1$ ordered decision boundaries and outputs a vector of signs; at a lower rate it keeps only every $2^{B_{\max}-B_i}$-th boundary and fills erased signs with zeros, so the decoder can interpolate them from the ordering. The quantized feature spaces at different rates are disjoint, $\mathcal{S}(B_i)\cap\mathcal{S}(B_j)=\emptyset$, which tells the decoder how much precision is missing and lets the multi-modal fusion network weight sensor data adaptively. A surrogate tanh gradient makes the non-differentiable sign operation trainable end-to-end from the reconstruction loss alone, and the two-stage training recipe preserves a universal wireless-only codebook while adding site-dependent sensor refinement.

What would settle it

Run the two-stage pipeline on a real urban dataset with a co-located camera and base station where mobile terminals appear outside the camera field of view and in heavy NLOS or weather, then compare precoded beamforming gain against the CSI-only baseline; if the image-assisted gain is not statistically positive, the central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that quantization itself can be made rate-adaptive and fusion-friendly: the channel encoder learns an ordered set of decision boundaries per feature, and to emit a shorter CSI stream it uniformly subsamples those boundaries, with the decoder interpolating the erased comparisons from the ordering. The quantized feature spaces at different rates are disjoint, so the decoder inherently knows how much precision is missing, and a multi-modal Transformer can then refine the discrete CSI feature using RGB or uplink-sensor features. Training is split into a wireless-only stage that learns a universal variable-rate codebook and a transfer-learning stage, with the backbone frozen, that learns site-dependent sensor refinement. In 5G NR-compliant CDL simulation, the result is beamforming gains that approach the ideal channel-matched beamformer, with image-assisted fusion performing best under low feedback rates and low SNR.

Load-bearing premise

The load-bearing premise is that synthetic paired data with mobile terminals inside the camera field of view and at most two ray reflections transfers to real propagation; if it fails, multi-modal gains collapse to wireless-only performance.

Editorial extensions

If this is right

  • At untried feedback lengths such as 56, 64, 80, 88, 104, 112, 128, and 136 bits, the proposed quantizer maintains smooth reconstruction performance where nested vector quantization collapses, meaning the system can switch rates in real time without retraining.
  • With RGB fusion, reconstruction loss improves by 4.7 dB at 48 feedback bits and by 2 dB at 144 bits, so the benefit of sensor data grows precisely when feedback is scarce.
  • At 96 bits and an SNR of -15 dB, RGB fusion adds 3.6 dB over CSI-only reconstruction, showing that image context can compensate for channel estimation noise.
  • Uplink CSI fusion raises NLOS reconstruction accuracy at all simulated rates, although noisy labels leave a visible worst-case gap to ideal beamforming.
  • An autoencoder trained on several discrete rates can be set to 76 bits and still beat Type II PMI beamforming in the simulated LOS and NLOS comparisons.

Reading between the lines

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

  • A direct corollary the authors do not emphasize: because low-rate CSI plus RGB approaches the quality of high-rate CSI alone, an operator could cut feedback overhead by moving bits to the image channel, effectively trading camera pixels for uplink resources.
  • The disjoint-rate quantizer suggests an untested calibration use: the number of erased quantization levels in the received stream is a natural, side-information-free confidence signal that a fusion network could use to gate sensor data; an ablation removing the rate indicator would test this.
  • If the synthetic-to-real transfer assumption fails, the visible failure mode would likely be graceful: the fusion branch should learn to down-weight uninformative images and degrade to wireless-only performance rather than worse.
  • The same two-stage recipe should extend to other co-located sensors, such as LiDAR or radar point clouds, provided their features can be paired with channels from the same scene; the authors' LOS-restricted pairing would then need a richer dataset with occluded and non-co-located sensors, which they list as future work.
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 / 7 minor

Summary. The paper proposes a multi-modal CSI reconstruction framework for FDD massive MIMO systems. The core contribution is a variable-rate autoencoder quantizer that uses a trainable element-wise quantization rule, enabling CSI bit streams of arbitrary length with only a few hundred quantizer parameters. The framework is trained in two stages: first a wireless-only autoencoder, then a transfer-learning stage that fuses the quantized CSI features with auxiliary sensor data (RGB images or uplink CSI) at the BS. The authors generate a synthetic paired dataset using Blender and ray tracing, and evaluate channel reconstruction loss and precoded beamforming gains under 5G NR parameters. They claim that the proposed quantizer outperforms nested VQ at untrained rates and Type II PMI, and that RGB image fusion achieves near-optimal beamforming gains.

Significance. If the results hold, the work provides a lightweight, variable-rate CSI feedback mechanism and demonstrates that environment sensing can meaningfully reduce feedback distortion. The paper ships no code or data, but the simulation design includes held-out test samples, comparisons to Type II PMI and a recent nested VQ method, and evaluation at untrained rates, which are positive aspects. The central multi-modal claim, however, is only supported on the same synthetic distribution used for training, and two technical claims in the paper (the loss-equivalence derivation and the disjointness of quantization levels) are incorrect. These issues are load-bearing for the paper's headline conclusion, so the manuscript needs substantive revision.

major comments (3)
  1. [Section II-B, Eqs. (6)-(7)] The stated equivalence between minimizing L(H,\hat H)=||H/||H||_F - \hat H/||\hat H||_F||_F^2 and maximizing the average column-wise cosine similarity is incorrect. Expanding the norm gives 2 - 2 Re{tr(\hat H^H H)}/(||H||_F ||\hat H||_F), which is a global normalized inner product, not an average of per-column cosines. A simple counterexample is N_s=2 with h_1=[1,0]^T, h_2=[0,100]^T, and \hat h_1=[0,1]^T, \hat h_2=[0,100]^T; then the average column cosine is 0.5 while L is near zero. Since the beamforming matrix in Eq. (8) normalizes each column independently, the training loss is not equivalent to the evaluation metric. The derivation should be corrected or the loss choice should be justified differently.
  2. [Section IV-A and Section V-B] The RGB fusion results, including the 'near-optimal' beamforming CDF in Fig. 13, are evaluated on the same synthetic distribution used for training. The dataset generation restricts MTs to the camera FOV, fixes the camera/BS at 15 m, limits reflection order to 2, and Section IV-B explicitly acknowledges the resulting LOS positional bias; that mitigation only covers the wireless-only mode, not the RGB fusion mode. No out-of-distribution test is provided (e.g., different camera pose, different city segment, NLOS with images, or real data such as ViWi [33]), so the central claim of near-optimal beamforming gains in 5G NR-compliant scenarios is only an in-distribution result. This is the load-bearing weakness of the multi-modal contribution.
  3. [Section III-B, Remark 1, Eq. (21)] The claimed disjointness S(B_i) \cap S(B_j) = \emptyset for B_i \neq B_j is false. For B_max=3, take a feature value z > b_i^{(6)}. For B_i=1, the selected boundary is b_i^{(4)}; the inference rule in Section III-B forces all entries to +1, giving z(s)_i = +7. For B_i=2 and B_i=3, the same value +7 is obtained. Thus the quantization level sets intersect at the upper saturation point (and analogously at lower saturation in some configurations). Since Remark 1 uses disjointness as the basis for the fusion network's ability to assess feedback reliability, this claim must be corrected or the surrounding argument revised.
minor comments (7)
  1. [Section V-C, text near Fig. 15] The text says 'at B=192' when describing uplink CSI fusion gains, but the maximum possible rate with N=48 and B_max=3 is 144 bits; this should be B=144 or the sentence should be reworded.
  2. [Section V-A, Fig. 10] For the untrained rate B=76, the paper does not explain how the 48 features are assigned B_i bits (1, 2, or 3) for rates not in the training set; please specify the allocation rule.
  3. [Section II-A] The word 'deontes' should be 'denotes'.
  4. [Reference [25]] The page range '2382-2376' is not well-formed; please check the correct pages for the published version.
  5. [Fig. 9 and Fig. 12] The x-axes start at 50 while the text states rates range from 48 to 144; please align the axes with the reported training rates.
  6. [Section III-C, Eqs. (22)-(24)] Please verify the dimensionality of the MSA output; the current notation with concatenated attention heads and matrix W appears to have transposition errors that make the dimensions in Eq. (24) inconsistent with the input shape in Eq. (26).
  7. [Section V (general)] The paper reports no error bars, confidence intervals, or multiple-seed runs; adding these would help assess whether small differences in reconstruction loss (e.g., Fig. 12 at high rates) are statistically meaningful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical results from a learned model on a synthetic dataset, not derivations that reduce to their own inputs.

full rationale

I examined the derivation chain for the circularity patterns specified. The variable-rate quantizer, the downsample-then-interpolate rule of Eq. (16)-(17), the weighted-average loss of Eq. (20), and the fusion equations (22)-(26) are all constructive definitions; none of the reported beamforming gains is obtained by re-inserting the evaluated quantity into the training objective. The loss in Eq. (6) is optimized and then beamforming gains are measured on held-out test samples; this is standard supervised evaluation, not a fitted parameter being renamed as a prediction. The paper uses no load-bearing self-citations: references to prior work, such as the Swin Transformer [37] and the ViWi dataset framework [33], are external and are not used to justify the central result. The main limitation, acknowledged in Section IV-B, is that the RGB fusion results are evaluated on the same synthetic, FOV-restricted, LOS-heavy distribution used for training, so real-world generalization is unproven; that is an external-validity and robustness concern, not circularity. No equation in the paper is equivalent by construction to another claimed prediction, so the circularity score is 0.

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

The central result rests on fitted model components (quantizer thresholds, network weights) and on the assumption that the synthetic paired dataset is representative. The trainable thresholds and network weights are optimized on the authors' generated data, so their values are not independent evidence. The axioms listed are the domain assumptions required for the simulation results to transfer to real FDD massive MIMO deployments; none is validated by measurements in the paper.

free parameters (4)
  • Quantizer boundaries B_i (trainable thresholds) = Learned; 336 parameters (N=48, Bmax=3)
    Each feature's quantization boundaries are trainable and define the variable-rate mapping in Eq. (10). The reported performance depends on their learned positions.
  • Autoencoder and fusion network weights = Trained; roughly 1e5 to 6e5 parameters per model
    Encoder, decoder, sensor feature extractor, and refinement transformer are all trained on the synthetic dataset. No pretrained weights are provided.
  • Hyperparameters (gamma, target rates, dimensions) = gamma=2^(1/96); rates {48,72,96,120,144}; dims [24,32,32,32,4]; M=4; patch 2x12
    Chosen by hand and used in the weighted-average loss (Eq. 20); results at untrained rates are sensitive to these choices.
  • Synthetic scene and channel parameters = 28/27 GHz, 4x4 UPA, N_RB=48, max reflection order 2, city scene parts, car model, lighting
    The dataset generator is constructed by the authors; all performance numbers are conditional on this generator.
assumptions (6)
  • domain assumption Ray tracing with reflection order limited to 2 is sufficient to model mmWave channels relevant for beamforming.
    Section IV-A sets this limit; no higher-order reflection analysis or measurement validation is provided.
  • domain assumption The Blender-generated city scenes are representative of real deployment environments.
    Section IV-A creates virtual 3D cities; the network is never tested on real images or measurements.
  • ad hoc to paper The tanh surrogate gradient in Eq. (11) provides a faithful approximation of the sign function for end-to-end training.
    The paper relies on this surrogate to train the quantizer without auxiliary losses; its fidelity is asserted rather than proven.
  • ad hoc to paper Weighted-average loss over target rates {48,72,96,120,144} induces usable quantization at untrained rates.
    Generalization to untrained rates is demonstrated empirically in Fig. 9 but is not guaranteed or derived.
  • domain assumption MTs inside the camera FOV in LOS conditions make RGB images informative for channel reconstruction.
    Section IV-B acknowledges the LOS bias; the multi-modal fusion gain is measured only under this favorable condition.
  • domain assumption Uplink and downlink channels share the same ray geometry at reciprocal frequencies.
    Section V-C uses uplink CSI as a fusion modality; this is a standard geometrical-reciprocity assumption, not validated with measurements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Modal Variable-Rate CSI Reconstruction for FDD Massive MIMO Systems." pith.science (2026). https://pith.science/paper/USQ2TD3N

@misc{pith2026250111926,
  author       = {Pith},
  title        = {Pith review of: Multi-Modal Variable-Rate CSI Reconstruction for FDD Massive MIMO Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/USQ2TD3N}},
  note         = {Machine review of arXiv:2501.11926}
}
read the original abstract

In frequency division duplex (FDD) systems, acquiring channel state information (CSI) at the base station (BS) traditionally relies on limited feedback from mobile terminals (MTs). However, the accuracy of channel reconstruction from feedback CSI is inherently constrained by the rate-distortion trade-off. To overcome this limitation, we propose a multi-modal channel reconstruction framework that leverages auxiliary data, such as RGB images or uplink CSI, collected at the BS. By integrating contextual information from these modalities, the framework mitigates CSI distortions caused by noise, compression, and quantization. At its core, the framework utilizes an autoencoder network capable of generating variable-length CSI, tailored for rate-adaptive multi-modal channel reconstruction. By augmenting the foundational autoencoder network using a transfer learning-based multi-modal fusion strategy, we enable accurate channel reconstruction in both single-modal and multi-modal scenarios. To train and evaluate the network under diverse and realistic wireless conditions, we construct a synthetic dataset that pairs wireless channel data with sensor data through 3D modeling and ray tracing. Simulation results demonstrate that the proposed framework achieves near-optimal beamforming gains in 5G New Radio (5G NR)-compliant scenarios, highlighting the potential of sensor data integration to improve CSI reconstruction accuracy.

Figures

Figures reproduced from arXiv: 2501.11926 by the authors.

Figure 1
Figure 1. Multi-modal channel reconstruction using image data and wireless data. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of multi-modal channel reconstruction processes at the BS and the MT. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Hierarchical self-attention computation using local windows/shifted-windows. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Network layers of the autoencoder. keep the complexity of W-MSA computation low. The network size can be further tuned using the embedding dimensions [𝑁L0, 𝑁L1, 𝑁L2, 𝑁L3, 𝑁p] (see [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Architecture of the channel-relevant feature extraction network. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Sensor data-assisted channel feature vector refinement procedure using multi-modal Transformer. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: RGB image-wireless channel paired dataset generation for LOS environments and wireless channel-only dataset generation [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Sample RGB images for fixed (𝑃1, 𝑃2, 𝐶) and random (𝐿, 𝑆) values. • Scenario modeling and image rendering: We create virtual 3D city environments using the 3D modeling software, Blender. Each scene includes various objects such as buildings, cars, roads, traffic lights…
Figure 9
Figure 9. Figure 9: Channel reconstruction losses using benchmarks with respect to various feedback rates at SNR [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: CDFs on precoded NLOS channel gain using various CSI feedback methods with target feedback length [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Beamforming matrices using various channel reconstruction methods at SNR [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: Channel reconstruction loss with and without using RGB images at various feedback rates. [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: CDFs on precoded LOS channel gain of various CSI feedback methods at SNR [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]
Figure 14
Figure 14. Figure 14: CDFs of quantized feature z(s) and refined feature zr at various feedback rates [PITH_FULL_IMAGE:figures/full_fig_p026_14.png]
Figure 15
Figure 15. Figure 15: Channel reconstruction loss with and without using uplink CSI fusion at various feedback rates. [PITH_FULL_IMAGE:figures/full_fig_p028_15.png]
Figure 16
Figure 16. Figure 16: CDFs on precoded NLOS channel gain of various CSI feedback methods at SNR [PITH_FULL_IMAGE:figures/full_fig_p028_16.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. 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.

Reference graph

Works this paper leans on

41 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [33]

    ViWi: a deep learning dataset framework for vision-aided wireless communications,

    M. Alrabeiah, A. Hredzak, Z. Liu, and A. Alkhateeb, “ViWi: a deep learning dataset framework for vision-aided wireless communications,” IEEE V eh. Tech. Conf., 2020

  2. [1]

    Millimeter-wave massive MIMO: the next wireless revolution?,

    A. L. Swindlehurst, E. Ayanoglu, P. Heydari, and F. Capolino, “Millimeter-wave massive MIMO: the next wireless revolution?,” IEEE Commun. Mag., vol. 52, no. 9, pp. 56-62, Sep. 2014

  3. [2]

    Five disruptive technology directions for 5G,

    F. Boccardi, R. W. Heath Jr., A. Lozano, T. L. Marzetta, and P. Popovski, “Five disruptive technology directions for 5G,” IEEE Commun. Mag., vol. 52, no. 2, pp. 74-80, Feb. 2014

  4. [3]

    Massive MIMO for next generation wireless systems,

    E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive MIMO for next generation wireless systems,” IEEE Commun. Mag., vol. 52, no. 2, pp. 186-195, Feb. 2014

  5. [4]

    Spatially sparse precoding in millimeter wave MIMO systems,

    O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath Jr., “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 1499-1513, Mar. 2014

  6. [5]

    Large-scale antenna systems with hybrid analog and digital beamforming for millimeter wave 5G,

    S. Han, C.-L. I, Z. Xu, and C. Rowell, “Large-scale antenna systems with hybrid analog and digital beamforming for millimeter wave 5G,” IEEE Commun. Mag., vol. 53, no. 1, pp. 186-194, Jan. 2015

  7. [6]

    D. J. Love, R. W. Heath Jr., V . K. N. Lau, D. Gesbert, B. D. Rao, and M. Andrews, “An overview of limited feedback in wireless communication systems, IEEE J. Sel. Areas Commun., vol. 26, no. 8, pp. 1341-1365, Oct. 2008

  8. [7]

    Overview of deep learning-based CSI feedback in massive MIMO systems,

    J. Guo, C.-K. Wen, S. Jin, and G. Y . Li, “Overview of deep learning-based CSI feedback in massive MIMO systems,” IEEE Trans. Commun., vol. 70, no. 12, pp. 8017-8045, Dec. 2022

Show all 41 references
  1. [8]

    Channel feedback based on AoD-adaptive subspace codebook in FDD massive MIMO systems,

    W. Shen, L. Dai, B. Shim, Z. Wang, and R. W. Heath Jr., “Channel feedback based on AoD-adaptive subspace codebook in FDD massive MIMO systems,” IEEE Trans. Commun., vol. 66, no. 11, pp. 5235-5248, Nov. 2018

  2. [9]

    H. Ju, S. Jeong, B. Lee, and B. Shim, ‘Transformer-assisted parametric CSI feedback for mmWave massive MIMO systems,” IEEE Trans. Wireless Commun., vol. 23, no. 12, pp. 18774-18787, Dec. 2024

  3. [10]

    Technical specification group radio access network, physical layer procedures for data (Release 18),

    3GPP, “Technical specification group radio access network, physical layer procedures for data (Release 18),” TS 38.214 V18.5.0, Dec. 2024

  4. [11]

    A review of codebooks for CSI feedback in 5G new radio and beyond,

    Z. Qin and H. Yin, “A review of codebooks for CSI feedback in 5G new radio and beyond,” arXiv preprint arXiv:2302.09222, 2023

  5. [12]

    Compressive sensing for feedback reduction in MIMO broadcast channels,

    M. E. Eltayeb, T. Y . Al-Naffouri, and H. R. Bahrami, “Compressive sensing for feedback reduction in MIMO broadcast channels,” IEEE Trans. Commun., vol. 62, no. 9, pp. 3209-3222, Sep. 2014. 30

  6. [13]

    Rate-adaptive feedback with Bayesian compressive sensing in multiuser MIMO beamforming systems,

    X.-L. Huang, J. Wu, Y . Wen, F. Hu, Y . Wang, and T. Jiang, “Rate-adaptive feedback with Bayesian compressive sensing in multiuser MIMO beamforming systems,” IEEE Trans. Wireless Commun., vol. 15, no. 7, pp. 4839-4851, July 2016

  7. [14]

    Compressive sensing techniques for next-generation wireless communications,

    Z. Gao, L. Di, S. Han, C.-L. I, Z. Wang, and L. Hanzo, “Compressive sensing techniques for next-generation wireless communications,” IEEE Wireless Commun., vol. 25, no. 3, pp. 144-153, June 2018

  8. [15]

    Joint sparse channel recovery with quantized feedback for multi-user massive MIMO systems,

    F. Kulsoom, A. Vizziello, H. N. Chaudhry, and P. Savazzi, “Joint sparse channel recovery with quantized feedback for multi-user massive MIMO systems,” IEEE Access, vol. 8, pp. 11046-11060, Jan. 2020

  9. [16]

    Technical specification group radio access network, study on artificial intelligence (AI)/machine learning (ML) for NR air interface (Release 18),

    3GPP, “Technical specification group radio access network, study on artificial intelligence (AI)/machine learning (ML) for NR air interface (Release 18),” TR 38.843, V18.0.0, Dec. 2023

  10. [17]

    Deep learning and compressive sensing-based CSI feedback in FDD massive MIMO systems,

    P. Liang, J. Fan, W. Shen, Z. Qin, and G. Y . Li, “Deep learning and compressive sensing-based CSI feedback in FDD massive MIMO systems,” IEEE Trans. V eh. Tech., vol. 69, no. 8, pp. 9217-9222, Aug. 2020

  11. [18]

    Deep learning for massive MIMO CSI feedback,

    C.-K. Wen, W.-T. Shih, and S. Jin, “Deep learning for massive MIMO CSI feedback,” IEEE Wireless Commun. Letters, vol. 7, no. 5, pp. 748-751, Oct. 2018

  12. [19]

    Distributed deep convolutional compression for massive MIMO CSI feedback,

    M. B. Mashhadi, Q. Yang, and D. G ¨und¨uz, “Distributed deep convolutional compression for massive MIMO CSI feedback,” IEEE Trans. Wireless Commun., vol. 20, no. 4, pp. 2621-2633, Apr. 2021

  13. [20]

    MIMO channel information feedback using deep recurrent network,

    C. Lu, W. Xu, H. Shen, J. Zhu, and K. Wang, “MIMO channel information feedback using deep recurrent network,” IEEE Commun. Letters, vol. 23, no. 1, pp. 188-191, Jan. 2019

  14. [21]

    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 Trans. Commun., vol. 68, no. 8, pp. 4761-4772, Aug. 2020

  15. [22]

    Quantization adaptor for bit-level deep learning-based massive MIMO CSI feedback,

    X. Zhang, Z. Lu, R. Zeng, and J. Wang, “Quantization adaptor for bit-level deep learning-based massive MIMO CSI feedback,” IEEE Trans. V eh. Tech., vol. 73, no. 4, pp. 5443-5453, Apr. 2024

  16. [23]

    Convolutional neural network-based multiple-rate compressive sensing for massive MIMO CSI feedback: design, simulation, and analysis,

    J. Guo, C.-K. Wen, S. Jin, and G. Y . Li, “Convolutional neural network-based multiple-rate compressive sensing for massive MIMO CSI feedback: design, simulation, and analysis,” IEEE Trans. Wireless Commun., vol. 19, no. 4, pp. 2827-2840, Apr. 2020

  17. [24]

    Changeable rate and novel quantization for CSI feedback based on deep learning,

    X. Liang, H. Chang, H. Li, X. Gu and L. Zhang, “Changeable rate and novel quantization for CSI feedback based on deep learning,” IEEE Trans. Wireless Commun., vol. 21, no. 12, pp. 10100-10114, Dec. 2022

  18. [25]

    Vector quatization for deep-learning-based CSI feedback in massive MIMO systems,

    J. Shin, Y . Kang, and Y .-S. Jeon, “Vector quatization for deep-learning-based CSI feedback in massive MIMO systems,” IEEE Wireless Commun. Letters, vol. 13, no. 9, pp. 2382-2376, Sep. 2024

  19. [26]

    T. M. Cover and J. A. Thomas, Elements of Information Theory, New york, NY , USA: Wiley, 2006

  20. [27]

    Integrated sensing and communications: toward dual-function wireless networks for 6G and beyond,

    F. Liu, Y . Cui, C. Masouros, J. Xu, T. X. Han, Y . C. Eldar, and S. BuzziM “Integrated sensing and communications: toward dual-function wireless networks for 6G and beyond,” IEEE J. Sel. Areas Commun., vol. 40, no. 6, pp. 1728-1767, June 2022

  21. [28]

    Vision-assisted millimeter-wave beam management for next- generation wireless systems: concepts, solutions, and open challenges,

    K. Zheng, H. Yang, Z. Ying, P. Wang, and L. Hanzo, “Vision-assisted millimeter-wave beam management for next- generation wireless systems: concepts, solutions, and open challenges,” IEEE V eh. Tech. Mag., vol. 18, no. 3, pp. 58-68, Sep. 2023

  22. [29]

    Vision-aided positioning and beam focusing for 6G terahertz communications,

    S. Kim, J. Moon, J. Wu, B. Shim, and M. Z. Win, “Vision-aided positioning and beam focusing for 6G terahertz communications,” IEEE J. Sel. Areas Commun., vol. 42, no.9, pp. 2503-2519, Sep. 2024

  23. [30]

    Sensing and computer vision-aided mobility management for 6G millimeter and terahertz communication systems,

    Y . Ahn, J. Kim, S. Kim, S. Kim, and B. Shim, “Sensing and computer vision-aided mobility management for 6G millimeter and terahertz communication systems,” IEEE Trans. Commun., vol. 72, no. 10, pp. 6044-6058, Oct. 2024

  24. [31]

    Toward intelligent millimeter and terahertz communication for 6G: computer vision-aided beamforming,

    Y . Ahn, J. Kin, S. Kim, K. Shim, J. Kim, S. Kim, and B. Shim, “Toward intelligent millimeter and terahertz communication for 6G: computer vision-aided beamforming,” IEEE Wireless Commun., vol. 30, no. 5, pp. 179-186, Oct. 2023

  25. [32]

    Vision-aided frame-capture-based CSI recomposition for WiFi sensing: a multi modal approach,

    H. Shimomura, Y . Koda, T. Kanda, K. Yamamoto, T. Nishio, and A. Taya, “Vision-aided frame-capture-based CSI recomposition for WiFi sensing: a multi modal approach,” IEEE Consumer Commun. Net. Conf., 2023. 31

  26. [34]

    Technical specification group radio access network, physical channels and modulation (Release 18),

    3GPP, “Technical specification group radio access network, physical channels and modulation (Release 18),” TS 38.211 V18.5.0, Dec. 2024

  27. [35]

    Technical specification group radio access network, study on channel model for frequencies from 0.5 to 100GHz (Release 18),

    3GPP, “Technical specification group radio access network, study on channel model for frequencies from 0.5 to 100GHz (Release 18),” TR 38.901 V18.0.0, Mar. 2024

  28. [36]

    Neural discrete representation learning,

    A. V . D. Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” Adv. Neural Inf. Process. Syst., 2017

  29. [37]

    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,” IEEE Int. Conf. Comp. Vision, 2021

  30. [38]

    Multimodal learning with Transformers: a survey,

    P. Xu, X. Zhu, and D. A. Clifton, “Multimodal learning with Transformers: a survey,” arXiv preprint arXiv:2206.06488, 2022

  31. [39]

    User identification: a key enabler for multi-user vision-aided communications,

    G. Charan and A. Alkhateeb, “User identification: a key enabler for multi-user vision-aided communications,” IEEE Open Journal Commun. Society, vol. 5, pp. 472-488, Dec. 2023

  32. [40]

    Attention is all you need,

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

  33. [41]

    Latent Bernoulli autoencoder,

    J. Fajtl, V . Argyriou, D. Monekosso, and P. Remagnino, “Latent Bernoulli autoencoder,” Int. Conf. Machine Learn., 2020

Pith tools

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