User-Centric Stream Sensing for Grant-Free Access: Deep Learning with Covariance Differencing

Hyun Jong Yang; Sojeong Park; Yeongjun Kim

REVIEW 4 major objections 5 minor 16 references

Covariance differencing isolates newly activated streams in overloaded grant-free access.

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

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

T0 review · deepseek-v4-flash

2026-08-03 10:39 UTC pith:SFTUBESJ

load-bearing objection A plausible incremental method for detecting new streams via covariance differencing plus a DL classifier; the math holds under stated assumptions, but the key assumption of static pre-existing channels across windows is untested and could be a dealbreaker in realistic mobility. the 4 major comments →

arxiv 2601.09168 v1 pith:SFTUBESJ submitted 2026-01-14 eess.SP

User-Centric Stream Sensing for Grant-Free Access: Deep Learning with Covariance Differencing

Sojeong Park , Yeongjun Kim , Hyun Jong Yang This is my paper

classification eess.SP

keywords grant-free accessmassive MIMOstream sensingcovariance differencingdeep learningoverloaded regimesingular value decompositionlisten-before-talk

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that a user device with only a few antennas can still detect how many new streams just began transmitting, even when the total number of active streams exceeds the number of receive antennas. The method subtracts the previous sensing window's covariance matrix from the current one, so pre-existing streams cancel and only the newly activated streams remain visible. A deep-learning classifier then reads the singular values of both covariances and their difference to estimate the count of new streams. The approach is tested on flat Rayleigh fading and a standardized frequency-selective channel model, where it outperforms non-deep-learning baselines and remains accurate in overloaded conditions.

Core claim

On its own terms, the paper's central claim is equation (19): under static pre-existing channels and stationary noise power across two consecutive windows, the ideal covariance difference matrix is spanned solely by the d newly activated streams, free from interference by pre-existing users. This transforms an ill-posed absolute enumeration problem into a tractable relative detection problem: instead of estimating the total number of streams K, one estimates only the increment d. The paper also derives a theoretical bound on covariance deviation induced by channel variation, scaling as O(sqrt(1 - rho_th)), to justify choosing a sensing window over which the channel is sufficiently correlated

What carries the argument

Covariance differencing is the central mechanism: subtracting the sample covariance of the previous window from that of the current window cancels the spatial covariance of pre-existing streams, leaving a difference matrix whose dominant subspace is spanned by the newly activated streams. The feature vector is the concatenation of the singular values of the two individual covariance matrices and the difference matrix. A two-stream fully connected neural network maps this feature vector to an estimate of d, the number of newly activated streams.

Load-bearing premise

The channels of the pre-existing streams are assumed to stay exactly the same across the two sensing windows; if they change between windows, the covariance differencing cannot cancel them and the estimate of newly activated streams becomes biased.

What would settle it

Introduce a controlled channel variation between the two sensing windows (for example, a small Doppler shift or antenna motion affecting only the pre-existing users) and measure whether sensing accuracy degrades toward chance level as the differencing fails to cancel the pre-existing streams.

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

If this is right

If the central claim holds, user equipment can execute listen-before-talk-style collision avoidance without requiring more antennas than the number of active streams.
Incremental stream-count tracking becomes feasible in dense grant-free scenarios, where total stream counts are unobservable but changes are observable.
The derived O(sqrt(1 - rho_th)) bound provides a concrete rule for choosing the sensing window size based on channel correlation.
Fusing raw and differenced spectral features yields better accuracy than using either alone, indicating complementary information in the two representations.

Where Pith is reading between the lines

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

A natural extension is to relax the static-channel assumption by predicting or interpolating the previous covariance forward in time, which would make the method robust to mild Doppler or phase noise.
The same covariance-differencing principle could be applied to cognitive-radio scenarios where a device must detect newly arriving interferers in a crowded spectrum.
The correlation-based window-size bound suggests a testable design tradeoff: larger windows improve statistical averaging but must respect the channel coherence constraint to keep differencing effective.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A plausible incremental method for detecting new streams via covariance differencing plus a DL classifier; the math holds under stated assumptions, but the key assumption of static pre-existing channels across windows is untested and could be a dealbreaker in realistic mobility.

read the letter

Straight answer: the paper is honest about its core assumption, and the math is correct under that assumption. The idea is to stop estimating the total number of streams and instead detect only the new ones by subtracting the covariance of a previous window from the current one; the pre-existing streams' channels are assumed static, so they cancel exactly. On top of that, a small MLP maps singular values of the raw and difference SCMs to the number of new streams. That combination for user-centric grant-free sensing is not something I've seen cited elsewhere, so the application is genuinely new, even though the components are known.

What the paper does well: the problem is real, the exposition is clear, the bound in Section III-A is fine as far as it goes, and the simulations show consistent gains over MDL and thresholding in overloaded cases. The TDL-A results give a bit more credibility than pure i.i.d. Rayleigh.

The soft spots are real but not fatal if the static assumption holds. The stress-test is right: the covariance-deviation bound in (16) covers channel variation within a single window only. It does not cover variation of a pre-existing channel between the two windows. Equation (19) cancels the old users only if the h_k are identical in window t and t+1. If they change, the leftover term can have rank up to N_U and power scaling with K_t, which would swamp the d new streams. The simulations explicitly keep the channel and noise power static over the frame, so the one assumption the whole method rests on is never tested. The paper should at least simulate with channel aging, phase noise, or a mobility model, and possibly provide a bound on inter-window variation. Other issues: no code or data, no error bars, N_o=140 seems arbitrary, and MDL is only applied to the difference matrix—a fairer baseline would also run MDL on the raw covariance of the current window. Not citing classic covariance-differencing work (e.g., from array processing with unknown noise fields) is a gap, but not a correctness problem.

This is a modest, clearly-written engineering contribution. If you work on grant-free access or DL-based sensing, it's worth a read; if the authors can show the method degrades gracefully when pre-existing channels move, it becomes a solid contribution. For me, it's a 'maybe' for the reading group, and I wouldn't cite it in the next year, but I'd send it to review because the problem is relevant and the idea is testable.

Referee Report

4 major / 5 minor

Summary. The paper proposes a user-centric stream sensing method for grant-free access in overloaded MIMO scenarios, where the number of active streams K_t can exceed the number of receive antennas N_U. Instead of estimating the total active stream count, the method estimates the number d of newly activated streams between two consecutive observation windows. The key idea is covariance differencing: subtracting the sample covariance matrix (SCM) of the previous window from that of the current window to cancel the contribution of pre-existing streams, leaving a subspace spanned chiefly by the new streams. The authors provide a theoretical bound on the covariance deviation induced by channel variation within a single window, use this bound to justify a correlation-threshold-based choice of subcarrier grouping, and integrate a deep-learning classifier that fuses singular-value features from the two SCMs and their difference to estimate d. Simulations under i.i.d. Rayleigh fading and 3GPP TR 38.901 TDL-A channels show that the proposed method outperforms non-DL baselines (thresholding and MDL) and ablation baselines (difference-only and raw-only). The central claim is that covariance differencing plus DL enables reliable detection of newly activated streams even in overloaded regimes.

Significance. If the stated assumptions hold, the proposed framework is a useful incremental contribution to grant-free access: it reframes an ill-posed total-stream-count estimation problem into a better-conditioned incremental-count problem, and it demonstrates that a DL classifier can exploit fused spectral features to outperform conventional source-enumeration methods in overloaded conditions. The theoretical derivation in Section III-A is mathematically sound under the stated assumptions, and the simulation study covers both synthetic and standardized channel models, with the proposed method consistently achieving the highest accuracy. However, the significance is moderated by two gaps: (i) the core cancellation in Eq. (19) relies on pre-existing channels being static across the two windows, an assumption that is enforced in simulation but not modeled or stress-tested; and (ii) the theoretical bound in Section III-A concerns the true statistical covariance, not the finite-sample SCM that the DL classifier is intended to correct. If these gaps are addressed, the paper could be a solid practical contribution to UE-side sensing for grant-free access.

major comments (4)

[§III-B, Eq. (19); §IV-A] The central cancellation in Eq. (19) assumes the channels of the K_t pre-existing streams are identical in windows t and t+1. The theoretical bound in Section III-A (Eqs. (10)–(16)) only characterizes channel variation within a single window (sample ℓ vs. the reference at the first sample), not variation between the two windows. No theoretical analysis or simulation is provided for the case where h_k changes between windows; the residual term Σ_{k=1}^{K_t}(h_{t+1,k}h_{t+1,k}^H − h_{t,k}h_{t,k}^H) then has rank up to min(N_U,K_t) and power scaling with K_t, which can overwhelm the d new streams in overloaded conditions. The simulations explicitly enforce 'both the channel and noise power remain static over a frame' (Sec. IV-A), so the claimed robustness is demonstrated only under the very assumption on which the cancellation rests. This is load-bearing: a sensing UE in practice cannot gua
[§III-A, title and Eqs. (7)–(16)] The section is titled 'SCM Deviation Analysis' and claims to bound the 'estimation error' of the SCM, but the derivation actually bounds the deviation of the true statistical covariance R_t from the reference R*_t (Eqs. (7)–(16)). Finite-sample estimation error—the difference between the SCM \hat R_t and R_t—is not modeled. Since the DL classifier is introduced precisely to compensate for finite-sample residual interference, the theoretical analysis does not cover the regime the method is designed for. This is a mismatch between the analytical claim and the actual quantity being bounded. Please clarify that the bound applies to the true covariance, not the SCM, or extend the analysis to include finite-sample terms.
[§IV-A, N_o=140] The window size parameter N_o=140 (number of OFDM symbols) is fixed without justification or sensitivity analysis. The theoretical bound in Section III-A is used to choose N_s=7 subcarriers via the correlation threshold ρ_th, but N_o is chosen ad hoc. Since the total number of samples L = N_o N_s directly affects the quality of the SCM and hence the difficulty of the classification task, the lack of any sensitivity study for N_o undermines the claim that the theoretical bound determines the sensing window size. Please either derive a principled criterion for N_o or provide experiments varying N_o to show that the performance is robust to this choice.
[§IV-C, Figs. 5–6] The accuracy comparisons are reported without error bars or confidence intervals, and several reported gains are small (e.g., 5.5% over Raw Only and 10.8% over Difference Only at SNR=20 dB in the i.i.d. case). Without information on the number of Monte Carlo runs or variance, it is difficult to assess whether the observed differences are statistically significant, especially for the baseline comparisons in the TDL-A channel where the curves are close. Please include error bars, confidence bounds, or at least the number of independent trials.

minor comments (5)

[§II, first paragraph] Typo: 'up toK GF GF UEs' should be 'up to K_GF GF UEs'.
[§III-A, Eq. (6)] Notation inconsistency: H_{t,ℓ} is defined in Eq. (5), but the reference channel is denoted H_t, which previously (Eq. (1)) was the composite channel matrix. The notation is understandable but could be made cleaner, e.g., using a subscript for the reference sample.
[§III-A, Eq. (16)] The bound is stated as O(√(1−ρ_th)) under the implicit assumption that E||H_t||²_F is bounded. It would be helpful to state explicitly that the O(·) hides the channel norm and that the bound grows with K_t.
[§IV-A, Fig. 4] The figure title and caption state 'Δf = 7; ρ_th = 0.99', but the text says N_s=7 is chosen. Please clarify that the maximum subcarrier difference in the window is N_s−1=6, so the condition |ρ_h|≥0.99 is satisfied for the relevant range.
[References] The MDL baseline is attributed to a general MDL book [10], but the classic source-enumeration MDL reference (Wax and Kailath, 1985) would be more appropriate and helpful for readers.

Circularity Check

0 steps flagged

No significant circularity: the derivation is self-contained and the central cancellation is explicit algebra under stated assumptions.

full rationale

The paper's central claim, Eq. (19), is obtained by directly subtracting the reference covariance of window t, Eq. (17), from that of window t+1, Eq. (18), under the explicitly stated assumptions that pre-existing channels are static across the two windows and that noise power is stationary. This is a straightforward algebraic consequence of those assumptions, not a prediction that is equivalent to its own inputs by construction. The Section III-A covariance-deviation bound follows from the definition of the channel correlation coefficient (11), the triangle inequality, submultiplicativity of matrix norms, and Cauchy-Schwarz; it is a mathematical derivation rather than a fitted or assumed result. The DL classifier is trained on labeled data and evaluated on held-out realizations with external baselines (MDL, thresholding) and the Sionna/3GPP TDL-A channel model; this is disclosed empirical fitting, not a claimed first-principles derivation, so it does not constitute circularity. There are no load-bearing self-citations: the authors do not rely on their own prior uniqueness theorems or ansatz-carrying citations. The main weakness noted by a skeptical reader — that inter-window channel variation is assumed away and never tested in simulation — is a robustness/assumption-validation gap, not a circularity. Under the review rules, that concern belongs to correctness risk, not to the circularity score. Because the derivation chain is self-contained and benchmarked externally, the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The central claim rests on domain assumptions about channel stationarity and noise stationarity, plus several hand-chosen simulation parameters. No new physical entities are introduced.

free parameters (3)

ρ_th = 0.99
Correlation threshold for window size selection; hand-chosen, not derived.
N_s = 7
Number of subcarriers per window; set so that frequency-domain correlation >0.99 (Fig. 4).
N_o = 140
Number of OFDM symbols per window; fixed without stated justification in Section IV-A.

axioms (6)

standard math The triangle inequality and submultiplicativity of matrix norms hold.
Used to obtain Eq. (10).
domain assumption Transmitted signals x are zero-mean i.i.d. with E[x x^H] = I and independent of noise.
Assumed in Section III-B before Eq. (17), needed for the covariance differencing cancellation.
domain assumption Channels of the K_t pre-existing streams are static across the two observation windows.
Load-bearing assumption for ΔR*_t cancellation; stated in Section III-B.
domain assumption Noise power is stationary across consecutive windows.
Stated near Eq. (19); without it, ΔR*_t has a noise-scale term.
domain assumption Sensing UE is positioned to reliably capture aggregate uplink energy.
System model Section II; needed for the observation model to hold.
domain assumption Channel samples within a window have positive real correlation ≥ ρ_th.
Eq. (12); the choice ρ_th=0.99 defines coherence.

pith-pipeline@v1.3.0-alltime-deepseek · 7418 in / 10620 out tokens · 89598 ms · 2026-08-03T10:39:04.173936+00:00 · methodology

0 comments

read the original abstract

Grant-free (GF) access is essential for massive connectivity but faces collision risks due to uncoordinated transmissions. While user-side sensing can mitigate these collisions by enabling autonomous transmission decisions, conventional methods become ineffective in overloaded scenarios where active streams exceed receive antennas. To address this problem, we propose a differential stream sensing framework that reframes the problem from estimating the total stream count to isolating newly activated streams via covariance differencing. We analyze the covariance deviation induced by channel variations to establish a theoretical bound based on channel correlation for determining the sensing window size. To mitigate residual interference from finite sampling, a deep learning (DL) classifier is integrated. Simulations across both independent and identically distributed flat Rayleigh fading and standardized channel environments demonstrate that the proposed method consistently outperforms non-DL baselines and remains robust in overloaded scenarios.

Figures

Figures reproduced from arXiv: 2601.09168 by Hyun Jong Yang, Sojeong Park, Yeongjun Kim.

**Figure 2.** Figure 2: The proposed stream sensing method based on covariance differencing and a deep learning model. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Figure 3: Time-frequency resource grid structure for sampling [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: Frequency-domain correlation of the 3GPP TR 38.901 [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: Sensing accuracy versus the SNR with Kt = 4. 4 5 6 7 8 9 10 Kt 0.2 0.4 0.6 0.8 1 Accuracy Proposed Raw Only Difference Only Thresholding MDL (a) I.i.d. flat Rayleigh fading 4 5 6 7 8 9 10 Kt 0.2 0.4 0.6 0.8 1 Accuracy Proposed Raw Only Difference Only Thresholding MDL (b) 3GPP TR 38.901 TDL-A [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 6.** Figure 6: Sensing accuracy versus the number of pre-existing [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 1 linked inside Pith

[1]

The road towards 6G: A comprehensive survey,

W. Jiang, B. Han, M. A. Habibi, and H. D. Schotten, “The road towards 6G: A comprehensive survey,”IEEE Open J. Commun. Soc., vol. 2, pp. 334–366, 2021

2021
[2]

6G internet of things: A comprehensive survey,

D. C. Nguyenet al., “6G internet of things: A comprehensive survey,” IEEE Internet Things J., vol. 9, no. 1, pp. 359–383, 2022

2022
[3]

URLLC in beyond 5G and 6G networks: An interference management perspective,

M. U. A. Siddiquiet al., “URLLC in beyond 5G and 6G networks: An interference management perspective,”IEEE Access, vol. 11, pp. 54 639–54 663, 2023

2023
[4]

Millimeter-wave massive MIMO communication for future wireless systems: A survey,

S. A. Busari, K. M. S. Huq, S. Mumtaz, L. Dai, and J. Rodriguez, “Millimeter-wave massive MIMO communication for future wireless systems: A survey,”IEEE Commun. Surv. Tutor., vol. 20, no. 2, pp. 836–869, 2018

2018
[5]

NR; Medium Access Control (MAC) protocol specification,

3GPP, “NR; Medium Access Control (MAC) protocol specification,” TS 38.321, Oct. 2025, ETSI TS 138 321 V19.0.0

2025
[6]

Robust multiuser physical layer security for grant-free mMTC in beyond 5G/6G networks,

U. P. Enwereonye, A. S. Shahraki, H. Alavizadeh, and A. S. M. Kayes, “Robust multiuser physical layer security for grant-free mMTC in beyond 5G/6G networks,”IEEE Internet Things J., vol. 12, no. 23, pp. 50 110–50 125, 2025

2025
[7]

Grant-free random access in machine-type communication: Approaches and challenges,

J. Choi, J. Ding, N.-P. Le, and Z. Ding, “Grant-free random access in machine-type communication: Approaches and challenges,”IEEE Wirel. Commun., vol. 29, no. 1, pp. 151–158, 2022

2022
[8]

Uplink grant-free access solutions for URLLC services in 5G new radio,

N. H. Mahmoodet al., “Uplink grant-free access solutions for URLLC services in 5G new radio,” inProc. 16th Int. Symp. Wireless Commun. Syst. (ISWCS), 2019, pp. 607–612

2019
[9]

Novel listen-before-talk access scheme with adaptive backoff procedure for uplink centric broadband communication,

H. Zhou, Y . Deng, and A. Nallanathan, “Novel listen-before-talk access scheme with adaptive backoff procedure for uplink centric broadband communication,”IEEE Internet Things J., vol. 10, no. 22, pp. 19 981– 19 992, 2023

2023
[10]

P. D. Gr ¨unwald,The minimum description length principle. MIT Press, 2007

2007
[11]

An introduction to deep learning for the physical layer,

T. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,”IEEE Trans. Cogn. Commun. Netw., vol. 3, no. 4, pp. 563–575, 2017

2017
[12]

Physical layer communication via deep learning,

H. Kim, S. Oh, and P. Viswanath, “Physical layer communication via deep learning,”IEEE J. Sel. Areas Inf. Theory, vol. 1, no. 1, pp. 5–18, 2020

2020
[13]

Deep learning at the physical layer: System challenges and applications to 5G and beyond,

F. Restuccia and T. Melodia, “Deep learning at the physical layer: System challenges and applications to 5G and beyond,”IEEE Commun. Mag., vol. 58, no. 10, pp. 58–64, 2020

2020
[14]

Goodfellow, Y

I. Goodfellow, Y . Bengio, and A. Courville,Deep learning. MIT Press Cambridge, 2016

2016
[15]

Study on channel model for frequencies from 0.5 to 100 GHz,

3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,” TR 38.901, Oct. 2025, 3GPP TR 38.901 V19.1.0

2025
[16]

Sionna: An open-source library for next-generation physical layer research,

J. Hoydiset al., “Sionna: An open-source library for next-generation physical layer research,”arXiv preprint arXiv:2203.11854, 2022

Pith/arXiv arXiv 2022

[1] [1]

The road towards 6G: A comprehensive survey,

W. Jiang, B. Han, M. A. Habibi, and H. D. Schotten, “The road towards 6G: A comprehensive survey,”IEEE Open J. Commun. Soc., vol. 2, pp. 334–366, 2021

2021

[2] [2]

6G internet of things: A comprehensive survey,

D. C. Nguyenet al., “6G internet of things: A comprehensive survey,” IEEE Internet Things J., vol. 9, no. 1, pp. 359–383, 2022

2022

[3] [3]

URLLC in beyond 5G and 6G networks: An interference management perspective,

M. U. A. Siddiquiet al., “URLLC in beyond 5G and 6G networks: An interference management perspective,”IEEE Access, vol. 11, pp. 54 639–54 663, 2023

2023

[4] [4]

Millimeter-wave massive MIMO communication for future wireless systems: A survey,

S. A. Busari, K. M. S. Huq, S. Mumtaz, L. Dai, and J. Rodriguez, “Millimeter-wave massive MIMO communication for future wireless systems: A survey,”IEEE Commun. Surv. Tutor., vol. 20, no. 2, pp. 836–869, 2018

2018

[5] [5]

NR; Medium Access Control (MAC) protocol specification,

3GPP, “NR; Medium Access Control (MAC) protocol specification,” TS 38.321, Oct. 2025, ETSI TS 138 321 V19.0.0

2025

[6] [6]

Robust multiuser physical layer security for grant-free mMTC in beyond 5G/6G networks,

U. P. Enwereonye, A. S. Shahraki, H. Alavizadeh, and A. S. M. Kayes, “Robust multiuser physical layer security for grant-free mMTC in beyond 5G/6G networks,”IEEE Internet Things J., vol. 12, no. 23, pp. 50 110–50 125, 2025

2025

[7] [7]

Grant-free random access in machine-type communication: Approaches and challenges,

J. Choi, J. Ding, N.-P. Le, and Z. Ding, “Grant-free random access in machine-type communication: Approaches and challenges,”IEEE Wirel. Commun., vol. 29, no. 1, pp. 151–158, 2022

2022

[8] [8]

Uplink grant-free access solutions for URLLC services in 5G new radio,

N. H. Mahmoodet al., “Uplink grant-free access solutions for URLLC services in 5G new radio,” inProc. 16th Int. Symp. Wireless Commun. Syst. (ISWCS), 2019, pp. 607–612

2019

[9] [9]

Novel listen-before-talk access scheme with adaptive backoff procedure for uplink centric broadband communication,

H. Zhou, Y . Deng, and A. Nallanathan, “Novel listen-before-talk access scheme with adaptive backoff procedure for uplink centric broadband communication,”IEEE Internet Things J., vol. 10, no. 22, pp. 19 981– 19 992, 2023

2023

[10] [10]

P. D. Gr ¨unwald,The minimum description length principle. MIT Press, 2007

2007

[11] [11]

An introduction to deep learning for the physical layer,

T. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,”IEEE Trans. Cogn. Commun. Netw., vol. 3, no. 4, pp. 563–575, 2017

2017

[12] [12]

Physical layer communication via deep learning,

H. Kim, S. Oh, and P. Viswanath, “Physical layer communication via deep learning,”IEEE J. Sel. Areas Inf. Theory, vol. 1, no. 1, pp. 5–18, 2020

2020

[13] [13]

Deep learning at the physical layer: System challenges and applications to 5G and beyond,

F. Restuccia and T. Melodia, “Deep learning at the physical layer: System challenges and applications to 5G and beyond,”IEEE Commun. Mag., vol. 58, no. 10, pp. 58–64, 2020

2020

[14] [14]

Goodfellow, Y

I. Goodfellow, Y . Bengio, and A. Courville,Deep learning. MIT Press Cambridge, 2016

2016

[15] [15]

Study on channel model for frequencies from 0.5 to 100 GHz,

3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,” TR 38.901, Oct. 2025, 3GPP TR 38.901 V19.1.0

2025

[16] [16]

Sionna: An open-source library for next-generation physical layer research,

J. Hoydiset al., “Sionna: An open-source library for next-generation physical layer research,”arXiv preprint arXiv:2203.11854, 2022

Pith/arXiv arXiv 2022