pith. sign in

arxiv: 2501.18250 · v3 · pith:KYYGAUITnew · submitted 2025-01-30 · 💻 cs.IT · eess.SP· math.IT

Neural CSI Compression Fine-Tuning: Taming the Communication Cost of Model Updates

Mehdi Sattari , Deniz G\"und\"uz , Tommy Svensson This is my paper

Pith reviewed 2026-05-23 04:46 UTC · model grok-4.3

classification 💻 cs.IT eess.SPmath.IT
keywords neural CSI compressionfine-tuningmodel updatesFDD massive MIMOrate-distortion optimizationstructured priorentropy codingdistribution shift
0
0 comments X

The pith

Full-model fine-tuning improves neural CSI compression rate-distortion even after paying for decoder updates via joint entropy coding and sparse priors.

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

The paper establishes that neural CSI compressors for FDD massive MIMO can be fully fine-tuned on small target-environment samples to handle distribution shifts, with the extra cost of sending updated decoder parameters controlled by folding that bit rate into the training objective and entropy-coding the updates together with the CSI feedback. A structured prior is used to make the parameter changes sparse and selective so the update overhead stays low. Simulation results on multiple CSI datasets show the net rate-distortion performance improves despite the added bits. A reader would care because CSI feedback is a major overhead in these systems and generalization failures are common when environments change.

Core claim

By explicitly including the bit rate of model updates in the fine-tuning loss, jointly entropy-coding those updates with the compressed CSI, and applying a structured prior that encourages sparse and selective parameter changes, full-model fine-tuning on a modest number of recent target-domain CSI samples yields substantial gains in overall rate-distortion performance of the neural compressor across several datasets, even after subtracting the communication cost of the updates.

What carries the argument

Structured prior on decoder parameters that promotes sparse and selective updates, combined with joint rate optimization and entropy coding of CSI feedback plus model updates.

If this is right

  • The net feedback efficiency improves when the evaluation horizon, quantization of updates, and size of the target dataset are varied within the ranges tested.
  • Performance gains hold across multiple independent CSI datasets.
  • Joint entropy coding of CSI and model updates is feasible without separate overhead channels.

Where Pith is reading between the lines

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

  • The same joint-rate-plus-sparse-prior technique could be tested on other neural compression tasks where the decoder must be updated over a rate-limited link.
  • If the sparsity pattern learned by the prior turns out to be stable across environments, the method might support incremental updates with even lower cost than full re-transmission.
  • One could measure whether the same fine-tuning budget yields larger gains when the structured prior is replaced by standard L1 regularization.

Load-bearing premise

The structured prior is able to produce sparse enough parameter updates that the model-update bit rate stays low while most of the fine-tuning performance gain is preserved.

What would settle it

An experiment on the same CSI datasets where the overall rate-distortion curve after fine-tuning lies above the no-fine-tuning baseline once model-update bits are included.

Figures

Figures reproduced from arXiv: 2501.18250 by Deniz G\"und\"uz, Mehdi Sattari, Tommy Svensson.

Figure 1
Figure 1. Figure 1: The neural CSI compressor architecture. On the decoder side, the entropy-coded quantized bit stream is first decoded, and the low-dimensional channel matrix is re￾covered by the decoding network. The entropy encoder and de￾coder share the same prior distribution for the latent variable. In our architecture, we adopt a fully factorized distribution based on a neural network cumulative called DeepFactorized … view at source ↗
Figure 4
Figure 4. Figure 4: The top view of the ‘O2 Dynamic’ scenario [41]. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: RD plots for the backbone neural CSI compressor [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: CSI samples used for fine-tuning and evaluation. The [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The RD plot for different fine-tuning schemes, applied to (a) QuaDRiGa and (b) DeepMIMO datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The effect of spike-and-slap prior and LRDM loss function on the RD performance of full-model fine-tuning. 0 200 400 600 800 1000 Epochs 10 3 10 4 Non-zero updates SSP UP (a) Number of non-zero updates. 0 200 400 600 800 1000 Epochs 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Rate (bits) SSP - Total rate UP - Total rate SSP - Model update rate UP - Model update rate (b) Model update and total bit-rates [PITH_FULL… view at source ↗
Figure 9
Figure 9. Figure 9: Number of non-zero model updates and update bit-rates using spike-and-slap and uniform priors. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Full-model RD performance across varying [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
read the original abstract

Efficient channel state information (CSI) compression is essential in frequency division duplexing (FDD) massive multiple-input multiple-output (MIMO) systems due to the substantial feedback overhead. Recently, deep learning-based compression techniques have demonstrated superior performance for CSI feedback. However, their performance often degrades under distribution shifts across wireless environments, largely due to limited generalization capability. To address this challenge, we consider a full-model fine-tuning scheme, in which both the encoder and decoder are jointly updated using a small number of recent CSI samples from the target environment. A key challenge in this setting is the transmission of updated decoder parameters to the receiver, which introduces additional communication overhead. To mitigate this bottleneck, we explicitly incorporate the bit rate of model updates into the fine-tuning objective and entropy-code the model updates jointly with the compressed CSI. Furthermore, we employ a structured prior that promotes sparse and selective parameter updates, thereby significantly reducing the model-update communication cost. Simulation results across multiple CSI datasets demonstrate that full-model fine-tuning substantially improves the rate-distortion performance of neural CSI compression, despite the additional cost of model updates. We further analyze the impact of the evaluation horizon, the quantization resolution of model updates, and the size of the target-domain dataset on the overall feedback efficiency.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a full-model fine-tuning scheme for neural CSI compression in FDD massive MIMO systems to mitigate performance degradation under distribution shifts. Both encoder and decoder are updated on a small set of target-environment CSI samples; the bit rate of the resulting model updates is explicitly folded into the fine-tuning objective, the updates are entropy-coded jointly with the CSI, and a structured prior is used to encourage sparse and selective parameter changes. Simulation results across multiple CSI datasets are reported to show that the net rate-distortion performance improves despite the added update overhead, with further analysis of evaluation horizon, quantization resolution, and target-dataset size.

Significance. If the empirical claims hold, the work addresses a practically important barrier to deploying learned CSI compressors in time-varying wireless environments. Explicitly penalizing model-update rate and employing a structured prior to control communication cost constitute a direct and relevant engineering contribution. The multi-dataset simulation campaign is a positive feature.

major comments (2)
  1. [Method section describing the structured prior and fine-tuning objective] The central claim that full-model fine-tuning yields net rate-distortion gains rests on the structured prior successfully producing sufficiently sparse/selective updates so that the added communication cost is amortized. The manuscript states that the prior 'promotes sparse and selective parameter updates' and incorporates update bit rate into the objective, yet supplies no ablation that isolates the prior, no sparsity statistics (fraction of non-zero deltas, effective coded parameter count), and no comparison against an unstructured Gaussian prior baseline. Without these, it is impossible to determine whether the reported net improvement is a general consequence of the prior or an artifact of the chosen datasets and horizons.
  2. [Abstract and simulation-results section] The abstract and simulation results assert that full-model fine-tuning 'substantially improves' rate-distortion performance, but the provided description contains no quantitative deltas, no explicit baseline comparisons, and no implementation details on the joint entropy coding of CSI and model updates. These omissions make it difficult to gauge the magnitude and robustness of the claimed gains.
minor comments (2)
  1. [Abstract] The abstract would be strengthened by including at least one concrete performance number (e.g., NMSE or rate reduction) to support the qualitative claim of substantial improvement.
  2. [Notation and objective-function definitions] Notation for bit-rate and distortion terms should be checked for consistency between the objective function and the reported simulation metrics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight areas where the manuscript can be strengthened with additional analysis and details. We address each major comment below and will incorporate the suggested revisions.

read point-by-point responses

  1. Referee: [Method section describing the structured prior and fine-tuning objective] The central claim that full-model fine-tuning yields net rate-distortion gains rests on the structured prior successfully producing sufficiently sparse/selective updates so that the added communication cost is amortized. The manuscript states that the prior 'promotes sparse and selective parameter updates' and incorporates update bit rate into the objective, yet supplies no ablation that isolates the prior, no sparsity statistics (fraction of non-zero deltas, effective coded parameter count), and no comparison against an unstructured Gaussian prior baseline. Without these, it is impossible to determine whether the reported net improvement is a general consequence of the prior or an artifact of the chosen datasets and horizons.

    Authors: We agree that the manuscript would benefit from explicit ablations and statistics to isolate the structured prior's contribution. In the revised version, we will add: (i) an ablation comparing the structured prior to an unstructured Gaussian prior baseline, (ii) sparsity statistics including the fraction of non-zero parameter deltas and effective coded parameter counts after entropy coding, and (iii) discussion of how these elements enable amortization of update costs across the evaluated datasets and horizons. This will strengthen the evidence that the net gains stem from the proposed prior. revision: yes

  2. Referee: [Abstract and simulation-results section] The abstract and simulation results assert that full-model fine-tuning 'substantially improves' rate-distortion performance, but the provided description contains no quantitative deltas, no explicit baseline comparisons, and no implementation details on the joint entropy coding of CSI and model updates. These omissions make it difficult to gauge the magnitude and robustness of the claimed gains.

    Authors: We acknowledge the need for more explicit quantification and details. The revised manuscript will update the abstract to include specific quantitative deltas (e.g., rate-distortion improvements in dB or percentage terms), expand the simulation-results section with explicit baseline comparisons (including no-fine-tuning and partial-update cases), and add implementation details on the joint entropy coding of CSI and model updates, such as the entropy model architecture, coding procedure, and quantization scheme. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method with direct rate incorporation and simulation validation

full rationale

The paper proposes a fine-tuning scheme for neural CSI compression that explicitly adds the measured bit rate of model updates to the training objective and employs a structured prior as a modeling choice. Performance gains are shown via simulations on multiple CSI datasets. No step reduces a claimed prediction or result to a fitted parameter or self-citation by construction; the central claims rest on external empirical benchmarks rather than tautological definitions or load-bearing self-references.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach rests on standard supervised learning assumptions plus the domain claim that recent CSI samples suffice for adaptation; no new physical entities are postulated.

free parameters (2)
  • weight on model-update rate term in fine-tuning objective
    The objective explicitly includes the bit rate of model updates, implying a tunable trade-off hyperparameter whose value is not stated in the abstract.
  • sparsity strength in structured prior
    The structured prior that promotes sparse updates is a modeling choice whose exact parameterization is not given.
axioms (2)
  • domain assumption A small number of recent CSI samples from the target environment are sufficient to adapt the model without overfitting or requiring large target-domain data.
    The fine-tuning scheme is described as using a small number of recent CSI samples.
  • domain assumption The entropy coder can jointly compress the sparse model updates and the CSI payload without significant rate overhead beyond the modeled cost.
    Joint entropy coding is presented as the mechanism that mitigates the update cost.

pith-pipeline@v0.9.0 · 5763 in / 1620 out tokens · 28899 ms · 2026-05-23T04:46:28.528535+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages · 1 internal anchor

  1. [1]

    Noncooperative cellular wireless with unlimited num- bers of base station antennas,

    T. L. Marzetta, “Noncooperative cellular wireless with unlimited num- bers of base station antennas,”IEEE Trans. Wireless Commun., vol. 9, no. 11, pp. 3590–3600, 2010

    work page 2010

  2. [2]

    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, 2014

    work page 2014

  3. [3]

    Distributed compressive CSIT estimation and feedback for FDD multi-user massive MIMO systems,

    X. Rao and V . K. N. Lau, “Distributed compressive CSIT estimation and feedback for FDD multi-user massive MIMO systems,”IEEE Trans. Signal Process., vol. 62, no. 12, pp. 3261–3271, 2014

    work page 2014

  4. [4]

    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, 2016

    work page 2016

  5. [5]

    User- driven adaptive CSI feedback with ordered vector quantization,

    V . Rizzello, M. Nerini, M. Joham, B. Clerckx, and W. Utschick, “User- driven adaptive CSI feedback with ordered vector quantization,”IEEE Wireless Commun. Lett., vol. 12, no. 11, pp. 1956–1960, 2023

    work page 1956

  6. [6]

    Predictive vector quantization for multicell cooperation with delayed limited feedback,

    R. Bhagavatula and R. W. Heath, “Predictive vector quantization for multicell cooperation with delayed limited feedback,”IEEE Trans. Wireless Commun., vol. 12, no. 6, pp. 2588–2597, 2013

    work page 2013

  7. [7]

    Deep learning for massive MIMO CSI feedback,

    C. Wen, W. Shih, and S. Jin, “Deep learning for massive MIMO CSI feedback,”IEEE Wireless Commun. Lett., vol. 7, no. 5, pp. 748–751, Mar. 2018

    work page 2018

  8. [8]

    Convolutional neural network- based multiple-rate compressive sensing for massive MIMO CSI feed- back: 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 feed- back: Design, simulation, and analysis,”IEEE Trans. Wireless Commun., vol. 19, no. 4, pp. 2827–2840, 2020

    work page 2020

  9. [9]

    Multi-resolution CSI feedback with deep learning in massive MIMO system,

    Z. Lu, J. Wang, and J. Song, “Multi-resolution CSI feedback with deep learning in massive MIMO system,” inProc. IEEE Int. Conf. Commun. (ICC), 2020, pp. 1–6

    work page 2020

  10. [10]

    Dilated convolution based CSI feedback compression for massive MIMO sys- tems,

    S. Tang, J. Xia, L. Fan, X. Lei, W. Xu, and A. Nallanathan, “Dilated convolution based CSI feedback compression for massive MIMO sys- tems,”IEEE Trans. V eh. Technol., vol. 71, no. 10, pp. 11 216–11 221, 2022

    work page 2022

  11. [11]

    Transnet: Full attention network for CSI feedback in FDD massive MIMO system,

    Y . Cui, A. Guo, and C. Song, “Transnet: Full attention network for CSI feedback in FDD massive MIMO system,”IEEE Wireless Commun. Lett., vol. 11, no. 5, pp. 903–907, 2022

    work page 2022

  12. [12]

    A learn- able optimization and regularization approach to massive MIMO CSI feedback,

    Z. Hu, G. Liu, Q. Xie, J. Xue, D. Meng, and D. G ¨und¨uz, “A learn- able optimization and regularization approach to massive MIMO CSI feedback,”IEEE Trans. Wireless Commun., vol. 23, no. 1, pp. 104–116, 2024

    work page 2024

  13. [13]

    MIMO channel as a neural function: Implicit neural representations for extreme CSI compression in massive MIMO systems,

    H. Wu, M. Zhang, Y . Shao, K. Mikolajczyk, and D. G ¨und¨uz, “MIMO channel as a neural function: Implicit neural representations for extreme CSI compression in massive MIMO systems,” 2024. [Online]. Available: https://arxiv.org/abs/2403.13615

    work page arXiv 2024

  14. [14]

    Distributed deep convo- lutional compression for massive MIMO CSI feedback,

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

    work page 2021

  15. [15]

    CSI-PPPNet: A one-sided one-for-all deep learning framework for massive MIMO CSI feedback,

    W. Chen, W. Wan, S. Wang, P. Sun, G. Y . Li, and B. Ai, “CSI-PPPNet: A one-sided one-for-all deep learning framework for massive MIMO CSI feedback,” 2023. [Online]. Available: https://arxiv.org/abs/2211.15851

    work page arXiv 2023

  16. [16]

    An introduction to neural data compression,

    Y . Yang, S. Mandt, and L. Theis, “An introduction to neural data compression,”F ound. Trends. Comput. Graph. Vis., vol. 15, no. 2, p. 113–200, apr 2023. [Online]. Available: https://doi.org/10.1561/ 0600000107

    work page 2023

  17. [17]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” in Adv. Neural Inf. Process. Syst., Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K. Weinberger, Eds., vol. 27. Curran Associates, Inc., 2014. [Online]. Available: https://proceedings.neurips.cc/paper files/pap...

    work page 2014

  18. [18]

    Auto-Encoding Variational Bayes,

    D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,” in 2nd Int. Conf. Learn. Represent. (ICLR) 2014, Banff, AB, Canada, Apr . 14-16, 2014, Conf. Track Proc., 2014

    work page 2014

  19. [19]

    The neural autoregressive distribution estimator,

    H. Larochelle and I. Murray, “The neural autoregressive distribution estimator,” inProc. 14th Int. Conf. Artif. Intell. Stat. (AISTATS), ser. Proc. Mach. Learn. Res., G. Gordon, D. Dunson, and M. Dud ´ık, Eds., vol. 15. Fort Lauderdale, FL, USA: PMLR, 11–13 Apr 2011, pp. 29–37. [Online]. Available: https://proceedings.mlr.press/v15/larochelle11a.html

    work page 2011

  20. [20]

    Pixel recurrent neural networks,

    A. van den Oord, N. Kalchbrenner, and K. Kavukcuoglu, “Pixel recurrent neural networks,” inProc. 33rd Int. Conf. Mach. Learn. (ICML), ser. Proc. Mach. Learn. Res., M. F. Balcan and K. Q. Weinberger, Eds., vol. 48. New York, NY , USA: PMLR, 20–22 Jun 2016, pp. 1747–1756. [Online]. Available: https://proceedings.mlr.press/v48/oord16.html

    work page 2016

  21. [21]

    Deep convolutional com- pression for massive MIMO CSI feedback,

    Q. Yang, M. B. Mashhadi, and D. G ¨und¨uz, “Deep convolutional com- pression for massive MIMO CSI feedback,” inProc. IEEE 29th Int. Workshop Mach. Learn. Signal Process. (MLSP), 2019, pp. 1–6

    work page 2019

  22. [22]

    Qui ˜nonero-Candela, M

    J. Qui ˜nonero-Candela, M. Sugiyama, A. Schwaighofer, and N. D. Lawrence,Dataset Shift in Machine Learning. MIT Press, 2022

    work page 2022

  23. [23]

    A comprehensive survey on transfer learning,

    F. Zhuang, Z. Qi, K. Duan, D. Xi, Y . Zhu, H. Zhu, H. Xiong, and Q. He, “A comprehensive survey on transfer learning,”Proc. IEEE, vol. 109, no. 1, pp. 43–76, 2020

    work page 2020

  24. [24]

    Generalizing to unseen domains via adversarial data augmentation,

    R. V olpi, H. Namkoong, O. Sener, J. C. Duchi, V . Murino, and S. Savarese, “Generalizing to unseen domains via adversarial data augmentation,” inAdv. Neural Inf. Process. Syst. (NeurIPS), S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa- Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc.,

    work page

  25. [25]

    Available: https://proceedings.neurips.cc/paper files/ paper/2018/file/1d94108e907bb8311d8802b48fd54b4a-Paper.pdf

    [Online]. Available: https://proceedings.neurips.cc/paper files/ paper/2018/file/1d94108e907bb8311d8802b48fd54b4a-Paper.pdf

    work page 2018

  26. [26]

    Unsupervised domain adaptation by backpropagation,

    Y . Ganin and V . Lempitsky, “Unsupervised domain adaptation by backpropagation,” inInt. Conf. Mach. Learn. (ICML). PMLR, 2015, pp. 1180–1189

    work page 2015

  27. [27]

    Full- duplex millimeter wave MIMO channel estimation: A neural network approach,

    M. Sattari, H. Guo, D. G ¨und¨uz, A. Panahi, and T. Svensson, “Full- duplex millimeter wave MIMO channel estimation: A neural network approach,”IEEE Trans. Mach. Learn. Commun. Netw., vol. 2, pp. 1093– 1108, 2024

    work page 2024

  28. [28]

    Adaptive neural signal detection for massive MIMO,

    M. Khani, M. Alizadeh, J. Hoydis, and P. Fleming, “Adaptive neural signal detection for massive MIMO,”IEEE Trans. Wirel. Commun., vol. 19, no. 8, pp. 5635–5648, May. 2020

    work page 2020

  29. [29]

    Airnet: Neural network transmission over the air,

    M. Jankowski, D. G ¨und¨uz, and K. Mikolajczyk, “Airnet: Neural network transmission over the air,” in2022 IEEE Int. Symp. Inf. Theory (ISIT), 2022, pp. 2451–2456

    work page 2022

  30. [30]

    Downlink CSI feedback algorithm with deep transfer learning for FDD massive MIMO systems,

    J. Zeng, J. Sun, G. Gui, B. Adebisi, T. Ohtsuki, H. Gacanin, and H. Sari, “Downlink CSI feedback algorithm with deep transfer learning for FDD massive MIMO systems,”IEEE Trans. Cogn. Commun. Netw., vol. 7, no. 4, pp. 1253–1265, 2021. 13

    work page 2021

  31. [31]

    Deep learning for efficient CSI feedback in massive MIMO: Adapting to new environments and small datasets,

    Z. Liu, L. Wang, L. Xu, and Z. Ding, “Deep learning for efficient CSI feedback in massive MIMO: Adapting to new environments and small datasets,”IEEE Trans. Wireless Commun., pp. 1–1, 2024

    work page 2024

  32. [32]

    Multi-task learning-based CSI feedback design in multiple scenarios,

    X. Li, J. Guo, C.-K. Wen, S. Jin, S. Han, and X. Wang, “Multi-task learning-based CSI feedback design in multiple scenarios,”IEEE Trans. Commun., vol. 71, no. 12, pp. 7039–7055, 2023

    work page 2023

  33. [33]

    Model transmission- based online updating approach for massive mimo csi feedback,

    B. Zhang, H. Li, X. Liang, X. Gu, and L. Zhang, “Model transmission- based online updating approach for massive mimo csi feedback,”IEEE Communications Letters, vol. 27, no. 6, pp. 1609–1613, 2023

    work page 2023

  34. [34]

    Continuous online learning- based csi feedback in massive mimo systems,

    X. Zhang, J. Wang, Z. Lu, and H. Zhang, “Continuous online learning- based csi feedback in massive mimo systems,”IEEE Communications Letters, vol. 28, no. 3, pp. 557–561, 2024

    work page 2024

  35. [35]

    Communication-efficient personalized federated edge learning for massive mimo csi feedback,

    Y . Cui, J. Guo, C.-K. Wen, and S. Jin, “Communication-efficient personalized federated edge learning for massive mimo csi feedback,” IEEE Transactions on Wireless Communications, vol. 23, no. 7, pp. 7362–7375, 2024

    work page 2024

  36. [36]

    User-centric online gossip training for autoencoder-based csi feedback,

    J. Guo, Y . Zuo, C.-K. Wen, and S. Jin, “User-centric online gossip training for autoencoder-based csi feedback,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 3, pp. 559–572, 2022

    work page 2022

  37. [37]

    On the advantages of stochastic encoders,

    L. Theis and E. Agustsson, “On the advantages of stochastic encoders,”CoRR, vol. abs/2102.09270, 2021. [Online]. Available: https://arxiv.org/abs/2102.09270

    work page arXiv 2021

  38. [38]

    Towards empirical sandwich bounds on the rate-distortion function,

    Y . Yang and S. Mandt, “Towards empirical sandwich bounds on the rate-distortion function,”ArXiv, vol. abs/2111.12166, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:244527239

    work page arXiv 2021

  39. [39]

    Variational image compression with a scale hyperprior,

    J. Ball ´e, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” inInt. Conf. Learn. Represent. (ICLR), 2018. [Online]. Available: https: //openreview.net/forum?id=rkcQFMZRb

    work page 2018

  40. [40]

    Wireless InSite,

    Remcom, “Wireless InSite,” http://www.remcom.com/wireless-insite

    work page

  41. [41]

    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, 2014

    work page 2014

  42. [42]

    DeepMIMO: A generic deep learning dataset for mil- limeter wave and massive MIMO applications,

    A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for mil- limeter wave and massive MIMO applications,” inProc. Inf. Theory Appl. Workshop (ITA), San Diego, CA, Feb 2019, pp. 1–8

    work page 2019

  43. [43]

    Overfitting for fun and profit: Instance-adaptive data compression,

    T. van Rozendaal, I. A. Huijben, and T. Cohen, “Overfitting for fun and profit: Instance-adaptive data compression,” inInt. Conf. Learn. Represent. (ICLR), 2021. [Online]. Available: https: //openreview.net/forum?id=oFp8Mx V5FL

    work page 2021

  44. [44]

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

    Y . Bengio, N. L ´eonard, and A. C. Courville, “Estimating or propagating gradients through stochastic neurons for conditional computation,” CoRR, vol. abs/1308.3432, 2013, accessed: 2025-01-27. [Online]. Available: http://arxiv.org/abs/1308.3432

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  45. [45]

    The spike-and-slab lasso,

    V . Roˇckov´a and E. I. G. and, “The spike-and-slab lasso,”Journal of the American Statistical Association, vol. 113, no. 521, pp. 431–444, 2018. [Online]. Available: https://doi.org/10.1080/01621459.2016.1260469

    work page doi:10.1080/01621459.2016.1260469 2018