Pith. sign in

REVIEW 3 major objections 3 minor 30 references

Phase-Only Positioning: Overcoming Integer Ambiguity Challenge through Deep Learning

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Cell-free networks can locate a user to about one centimeter using only uplink carrier phase, if two trained neural networks resolve the integer ambiguity problem at a fraction of the inference cost of maximum-likelihood search.

desk verdict The stress-test note is right: Eq. (3) mis-specifies the phase-noise covariance, and the MLE baseline in the current draft is consequently handicapped; the complexity-reduction claim needs a re-run with the correct covariance. read the letter →

arxiv 2506.07890 v1 pith:BQDWJBV2 submitted 2025-06-09 eess.SP

Fatih Ayten , Mehmet C. Ilter , Ossi Kaltiokallio , Jukka Talvitie , Akshay Jain , Elena Simona Lohan , Henk Wymeersch , Mikko Valkama This is my paper
classification eess.SP
keywords carrierphasepositioningintegerambiguitycell-freeMIMOdeeplearningphase-onlylocalizationneuralnetworkresolution6Ginferencecomplexity
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

This paper claims that a cell-free network can locate a user from nothing but uplink carrier phase measurements, without time-of-arrival or clock-bias correction, by letting neural networks absorb the integer ambiguity problem. Two approaches are proposed: an MLP that maps differential phase measurements directly to a position estimate, and a two-stage design in which an MLP first classifies the integer ambiguities and a CNN then fuses those estimates with the phase data. In simulations with twenty phase-synchronized antenna points and purely line-of-sight channels, the CNN reaches about 1 cm root-mean-square error at 0 dBm transmit power, and both networks need two to three orders of magnitude fewer floating-point operations than a grid-search maximum-likelihood baseline matched in accuracy. A sympathetic reader would take the contribution to be evidence that phase-only positioning is computationally practical in distributed 6G networks, provided the simulation model holds.

What carries the argument

The load-bearing object is the differential phase measurement $\delta_m = \Delta_m + k_m\lambda + w_m$, where $\Delta_m$ is the true distance difference to a reference antenna, $k_m$ is the integer ambiguity, and $w_m$ is zero-mean Gaussian noise with Fisher-derived covariance $\Sigma_{\mathrm{diff}}$. The integer ambiguity $k_m$ is bounded by the antenna separation in wavelengths, which turns ambiguity resolution into a classification problem with a finite label set. The MLP ambiguity estimator outputs a softmax probability per ambiguity branch, trained with sparse categorical cross-entropy; the CNN positioning model concatenates the phase vector with the argmax ambiguity estimates and regresses the position with MSE. The complexity argument is carried by a FLOP count: the MLE cost is $N_{\mathrm{grid}}(2I^2-2I-1)$, while the pruned networks cost roughly $10^6$ FLOPs, and pruning rates are chosen so accuracy stays essentially unchanged.

What would settle it

Run the same trained MLP and CNN on data from a real phase-synchronized distributed-antenna testbed with ground-truth UE positions at 800 MHz and 1.8 GHz; if the RMSE at 0 dBm exceeds a few tens of centimeters or the ambiguity accuracy drops well below the reported 98%, the line-of-sight simulation model is the point of failure.

Watch

Extended reading notes

Core claim

The central discovery is that integer phase ambiguity, usually treated as a combinatorial estimation barrier, can be resolved or bypassed by supervised deep networks operating on differential carrier phase. The authors show that an MLP trained with mean-squared error can regress user position directly from the wrapped phase differences, and that an MLP-classifier plus CNN pipeline that explicitly predicts the bounded integer ambiguities and then feeds them, along with phase, into a position regressor achieves the best accuracy. They report up to 99.89% element-wise and 98.40% overall ambiguity accuracy at 800 MHz, and roughly 1 cm RMSE positioning at 0 dBm with the CNN. Against a maximum-likelihood baseline whose grid density is constrained to match the network's FLOP count, both networks are more accurate; when the grid is instead refined until MLE matches their accuracy, the networks reduce inference cost by factors of roughly 310 to 1787.

Load-bearing premise

The results depend on the simulated line-of-sight channel with perfectly phase-synchronized antenna points, free-space path loss, and Gaussian phase noise with a Fisher-bound covariance that the paper states without derivation; if real multipath, clock bias, or calibration errors enter, the centimeter-level accuracy is not guaranteed.

Editorial extensions

If this is right

  • Phase-only positioning becomes a candidate for 6G cell-free networks where ToA measurements are corrupted by UE clock bias, since only phase observations and network-side synchronization are required.
  • Integer ambiguity resolution can be treated as a bounded classification task, with accuracy above 98% overall at 800 MHz and above 91% at 1.8 GHz in the simulated setting.
  • At matched inference complexity, both proposed networks beat the MLE grid-search baseline; at matched accuracy, they cut inference FLOPs by two to three orders of magnitude.
  • The CNN variant's centimeter-level accuracy at 0 dBm suggests that high transmit power is not needed for the best performance, only enough SNR for ambiguity estimation.
  • The method's 95th percentile errors (about 2.1 cm at 800 MHz and 3.1 cm at 1.8 GHz for the CNN) indicate the useful operating envelope if the channel model transfers to practice.

Reading between the lines

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

  • Because training and test data come from the same line-of-sight model, the centimeter-level numbers are a demonstration under ideal conditions; a real deployment would likely need training on measured or ray-traced channels including multipath before the complexity advantage survives.
  • The ambiguity estimator's probability distributions could be used beyond argmax, for example to reject low-confidence samples or feed an uncertainty-aware position filter, an option the paper does not explore.
  • The same bounded-ambiguity classification trick transfers to other phase-based distributed sensing problems, such as orientation estimation or near-field localization with large arrays, where the label set is also finite per wavelength.
  • The FLOP comparison counts inference only and omits training cost and memory footprint, so the claimed two-to-three-orders-of-magnitude reduction applies to run-time processing, not to the full life cycle of deploying a network.
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 / 3 minor

Summary. The paper studies uplink carrier-phase positioning in a cell-free / distributed antenna system, where the UE position is estimated exclusively from carrier phase measurements collected at phase-synchronized access points. To overcome the classical integer ambiguity problem, the authors propose two deep learning approaches: a direct MLP that maps differential phase measurements to position, and an integer-ambiguity-aided approach that first estimates the integer ambiguities with an MLP and then feeds them, together with the phase measurements, into a CNN that outputs the position. The numerical evaluation, based on simulated LoS links with free-space path loss, reports RMSEs near 1 cm at 0 dBm transmit power and claims an inference complexity reduction of two to three orders of magnitude relative to a grid-search MLE baseline. The paper also includes a FLOP-based complexity analysis and a pruning study for the NNs.

Significance. If the results hold, the paper would be a valuable demonstration that phase-only positioning with distributed APs is feasible at centimeter-level accuracy and that neural networks can resolve integer ambiguities implicitly or explicitly, at inference costs far below a dense grid-search MLE. The paper has several strengths: it targets a genuinely under-studied setting (phase-only observations, no ToA/CSI), it uses a held-out test set of 150e3 samples, it compares two NN architectures, it performs an ablation with true versus estimated ambiguities, and it provides explicit FLOP formulas for the proposed networks. The credibility of the headline claims, however, hinges on the correctness of the MLE baseline and its complexity accounting, both of which have serious issues detailed below.

major comments (3)
  1. [Section II, Eq. (3)] The covariance matrix Σdiff in (3) is inconsistent with the signal model. From y_i = sqrt(E ρ_i) exp(-j(2π d_i/λ - θ)) s + v_i, the phase error variance of observation r_i is approximately σ_i^2 ≈ N0/(2Eρ_i). Consequently, the differential measurement w_m = (λ/2π)(n_0 - n_m) has variance λ^2 N0/(8π^2 E)(1/ρ_0 + 1/ρ_m) and covariance between w_m and w_l equal to λ^2 N0/(8π^2 E ρ_0). Eq. (3) instead has 1/ρ_m^2 and 1/ρ_0^2, a factor of 1/ρ_i mismatch unless ρ_i is redefined as an amplitude gain, which would contradict the stated signal model. Because the derivation is explicitly omitted, this error is not visible in the text. The incorrect Σdiff is used in the MLE objective (5), mis-weighting weak APs, and if the same erroneous statistics are used to generate the phase observations for training and testing, the NN results themselves inherit the error. This is a load-bearing issue for the numerical comparison in Section V.
  2. [Section IV-A and Section V-B] The MLE complexity count in (7) is incomplete. The text explicitly neglects the cost of computing h(xue) and assumes Σdiff^{-1} is precomputed, counting only the quadratic form. In Section V-B, the benchmark MLE additionally performs a 100-step gradient descent fine-tuning after the grid search, yet this computational cost is never included in CMLE. As a result, the 'matched inference complexity' comparison in Fig. 4 gives the MLE extra computation that is not counted, and the complexity-reduction factors reported in Section V-D divide by an underestimated MLE cost. For the stated reduction factors (e.g., 310 at 800 MHz and 1787 at 1.8 GHz for the MLP), including the gradient steps would reduce the apparent gain. The authors should either include all MLE operations, including the gradient refinement and any h(xue) evaluations, in CMLE, or remove the gradient refinement from the baseline for the complexity-matched comparison.
  3. [Section II and Section V-A] The paper does not state explicitly how the phase noise variance σ_i^2 used to generate the simulated phase observations is set. If it is derived from the Fisher bound leading to (3), then the training data and all reported results depend on the erroneous covariance. If, instead, the correct phase noise variance N0/(2Eρ_i) is used, then the MLE baseline is the only component affected by the wrong Σdiff, creating an unfair comparison. The authors should specify the exact data-generation procedure for the phase observations and confirm whether the covariance in (3) is used there.
minor comments (3)
  1. [Section IV-D and Fig. 3] The CNN output shape is stated as (1, I−3, C) although the text says zero padding with stride 1 is used, which would preserve the input width I−1. The FLOP expression CCNN,c = 11C(I−3) follows the former, contradicting the padding description. Please clarify and make the shape and FLOP count consistent.
  2. [Section III-B] The definition of Q as the total number of labels across all differential ambiguities is clear, but in (8) it is not obvious which term corresponds to which layer, especially the final softmax layer. A short derivation or a pointer to Fig. 2 would help.
  3. [Section V] There are a few minor language issues: 'simulate' appears as 'stimulate' in Section V, and '100 m2' should be '100 m^2' or described as a 10 m × 10 m area. These are cosmetic but should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the NN positioning claims are validated on held-out test data from the same model, and the complexity-reduction claim is a defined benchmark comparison rather than a fitted prediction or a self-citation chain.

full rationale

The paper's central claims are the centimeter-level positioning accuracy of two neural-network approaches and the two-to-three-orders-of-magnitude inference-complexity reduction relative to an MLE baseline. Neither claim reduces by construction to its inputs. The NN models are trained with true labels on simulated phase observations and evaluated on 150,000 held-out test samples, so the reported RMSE values are genuine out-of-sample results, not fitted re-statements of the training data. The complexity comparison is defined by explicit FLOP formulas: the MLE cost is Ngrid(2I^2 - 2I - 1), and the NN costs are explicit pruned-layer expressions; the reduction factors are ratios of these quantities after choosing Ngrid to match either NN FLOPs or comparable accuracy. This is a benchmark construction, not a circular derivation. The paper does contain self-citations, notably [7] and [16] by overlapping authors, but they are used only for contextual statements about carrier-phase positioning and existing CPP studies; they are not load-bearing for the proposed method or its evaluation. There is no imported uniqueness theorem, no ansatz smuggled in through a citation, and no renamed known result presented as organization. The omitted derivation of the phase-noise covariance in Eq. (3) is a support/correctness concern that could affect the fairness of the MLE baseline, but it is not circularity: the NN results do not depend on Eq. (3), and the complexity-reduction claim is not equivalent to any fitted parameter. Overall, the derivation chain is self-contained with respect to its stated model, so no significant circularity is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The free parameters are architecture sizes, pruning rates, and training settings chosen by hand. The axioms are standard domain assumptions about LoS propagation, synchronization, and noise. No new physical entities are introduced.

free parameters (5)
  • Dense layer width A=B=D = 128
    Hand-chosen architecture sizes for the MLP models and the CNN dense layers in Section V.
  • Number of CNN filters C = 32
    Hand-chosen in Section V for the convolutional positioning model.
  • NN pruning rates = 50% (MLP), 75% (CNN), 0% (ambiguity estimator at 1.8 GHz)
    Tuned so that model performance remains largely unaffected; these rates directly determine the reported FLOP counts in Table I.
  • MLE gradient refinement steps = 100
    Fixed in Section V-B for the MLE benchmark and not counted in the MLE FLOP expression.
  • Training hyperparameters = learning rate 1e-4, L2 coefficient 1e-5, batch size 1000, 1000 epochs
    Standard choices, but they affect the fitted network accuracy and the reported performance.
assumptions (5)
  • domain assumption Line-of-sight-only channel with free-space path loss
    System model in Section II; multipath is absent from the simulations and deferred to future work in Section VI.
  • domain assumption Mutually phase-synchronized antenna points with known positions
    Section II assumes the APs are mutually phase-synchronized and their positions are known; the common phase offset is removed by differential measurements.
  • domain assumption Narrowband assumption: maximum delay difference between reference AP and other APs is much smaller than 1/W
    Stated in Section II directly after Eq. (2); this justifies treating the phase observations as a narrowband measurement.
  • domain assumption Phase noise is Gaussian with covariance equal to the Fisher lower bound in Eq. (3)
    Section II uses Eq. (3) as the noise covariance for the MLE, but the derivation is omitted due to space constraints.
  • domain assumption The MLE likelihood in Eq. (5) uses the wrapped differential distance as the mean and ignores summation over integer ambiguity branches
    Section II, Eq. (5); this is an implicit high-SNR modeling choice that assumes the correct ambiguity branch dominates the likelihood.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phase-Only Positioning: Overcoming Integer Ambiguity Challenge through Deep Learning." pith.science (2026). https://pith.science/paper/BQDWJBV2

@misc{pith2026250607890,
  author       = {Pith},
  title        = {Pith review of: Phase-Only Positioning: Overcoming Integer Ambiguity Challenge through Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BQDWJBV2}},
  note         = {Machine review of arXiv:2506.07890}
}
read the original abstract

This paper investigates uplink carrier phase positioning (CPP) in cell-free (CF) or distributed antenna system context, assuming a challenging case where only phase measurements are utilized as observations. In general, CPP can achieve sub-meter to centimeter-level accuracy but is challenged by the integer ambiguity problem. In this work, we propose two deep learning approaches for phase-only positioning, overcoming the integer ambiguity challenge. The first one directly uses phase measurements, while the second one first estimates integer ambiguities and then integrates them with phase measurements for improved accuracy. Our numerical results demonstrate that an inference complexity reduction of two to three orders of magnitude is achieved, compared to maximum likelihood baseline solution, depending on the approach and parameter configuration. This emphasizes the potential of the developed deep learning solutions for efficient and precise positioning in future CF 6G systems.

Figures

Figures reproduced from arXiv: 2506.07890 by the authors.

Figure 1
Figure 1. Uplink positioning with distributed antenna points (APs) where only [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Proposed MLP-based ambiguity estimation model: Rectangles [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Proposed CNN-based positioning model incorporating an MLP-based [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Positioning accuracy results and comparison between the two proposed [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: ECDF of the proposed NNs at a transmit power of 0 dBm. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 28 canonical work pages

  1. [1]

    The integrated sensing and communication revolution for 6G: Vision, techniques, and applications,

    N. Gonz ´alez-Prelcic et al. , “The integrated sensing and communication revolution for 6G: Vision, techniques, and applications,” IEEE Proc., vol. 112, no. 7, pp. 676–723, 2024

    work page 2024

  2. [2]

    Carrier phase-based precise heading and pitch estimation using a low-cost GNSS receiver,

    W. Ding et al. , “Carrier phase-based precise heading and pitch estimation using a low-cost GNSS receiver,” Remote Sensing , vol. 13, no. 18, p. 3642, 2021

    work page 2021

  3. [3]

    5G NR positioning enhancements in 3GPP release-18,

    H.-S. Cha et al., “5G NR positioning enhancements in 3GPP release-18,” IEEE Commun. Stand. Mag. , vol. 9, no. 1, pp. 22–27, 2025

    work page 2025

  4. [4]

    Study on Expanded and Improved NR Positioning (Release 18),

    3GPP, “Study on Expanded and Improved NR Positioning (Release 18),” Tech. Rep. TR 38.859, June 2024

    work page 2024

  5. [5]

    Indoor positioning in 5G-Advanced: Challenges and solution toward centimeter-level accuracy with carrier phase enhancements,

    J. Nikonowicz et al. , “Indoor positioning in 5G-Advanced: Challenges and solution toward centimeter-level accuracy with carrier phase enhancements,” IEEE Wireless Commun. , vol. 31, no. 4, 2024

    work page 2024

  6. [6]

    Carrier phase-based synchronization and high-accuracy positioning in 5G new radio cellular networks,

    S. Fan et al. , “Carrier phase-based synchronization and high-accuracy positioning in 5G new radio cellular networks,” IEEE Trans. Commun. , vol. 70, no. 1, pp. 564–577, 2022

    work page 2022

  7. [7]

    Orientation and location tracking of XR devices: 5G carrier phase-based methods,

    J. Talvitie, M. S¨aily, and M. Valkama, “Orientation and location tracking of XR devices: 5G carrier phase-based methods,” IEEE J. Sel. Topics Signal Process., vol. 17, no. 5, pp. 919–934, 2023

    work page 2023

  8. [8]

    Integer ambiguity resolution by mixture Kalman filter for improved GNSS precision,

    K. Berntorp, A. Weiss, and S. D. Cairano, “Integer ambiguity resolution by mixture Kalman filter for improved GNSS precision,” IEEE Trans. Aerosp. Electron. Syst. , vol. 56, no. 4, pp. 3170–3181, 2020

    work page 2020

Show all 30 references
  1. [9]

    Integer parameter estimation in linear models with applications to GPS,

    A. Hassibi and S. Boyd, “Integer parameter estimation in linear models with applications to GPS,” IEEE Trans. Signal Process. , vol. 46, no. 11, pp. 2938–2952, 1998

    1998

  2. [10]

    Neural network-based ambiguity resolution for precise attitude estimation with GNSS sensors,

    R. de Celis et al. , “Neural network-based ambiguity resolution for precise attitude estimation with GNSS sensors,” IEEE Trans. Aerosp. Electron. Syst., vol. 60, no. 5, pp. 6702–6716, 2024

    2024

  3. [11]

    Best integer equivariant estimation based on unsupervised machine learning for GNSS precise positioning and navigation in complex environments,

    Z. Zhang, X. Li, and H. Yuan, “Best integer equivariant estimation based on unsupervised machine learning for GNSS precise positioning and navigation in complex environments,” IEEE Trans. Aerosp. Electron. Syst., vol. 60, no. 3, pp. 2672–2682, 2024

    2024

  4. [12]

    Foundations of user- centric cell-free massive MIMO,

    ¨O. T. Demir, E. Bj ¨ornson, and L. Sanguinetti, “Foundations of user- centric cell-free massive MIMO,” F oundations and Trends® in Signal Processing, vol. 14, no. 3-4, pp. 162–472, 2021

    2021

  5. [13]

    A tutorial on extremely large-scale MIMO for 6G: Fundamentals, signal processing, and applications,

    Z. Wang et al. , “A tutorial on extremely large-scale MIMO for 6G: Fundamentals, signal processing, and applications,” IEEE Commun. Surveys Tuts., vol. 26, no. 3, pp. 1560–1605, 2024

    2024

  6. [14]

    Joint localization, synchronization and mapping via phase-coherent distributed arrays,

    A. Fascista et al. , “Joint localization, synchronization and mapping via phase-coherent distributed arrays,” IEEE J. Sel. Topics Signal Process. , pp. 1–16, (early access) 2025

    2025

  7. [15]

    Direct wideband coherent localization by distributed antenna arrays,

    N. Vukmirovic et al. , “Direct wideband coherent localization by distributed antenna arrays,” Sensors, vol. 19, no. 20, 2019

    2019

  8. [16]

    Fundamental performance bounds for carrier phase positioning in cellular networks,

    H. Wymeersch, R. Amiri, and G. Seco-Granados, “Fundamental performance bounds for carrier phase positioning in cellular networks,” in Proc. IEEE GLOBECOM , Kuala Lumpur, Malaysia, 2023, pp. 7478– 7483

    2023

  9. [17]

    Accurate direct positioning in distributed mimo using delay-doppler channel measurements,

    B. Deutschmann et al., “Accurate direct positioning in distributed mimo using delay-doppler channel measurements,” in Proc. IEEE SPAWC, Lucca, Italy, 2024, pp. 606–610

    2024

  10. [18]

    Near-field tracking with large antenna arrays: Fundamental limits and practical algorithms,

    A. Guerra et al. , “Near-field tracking with large antenna arrays: Fundamental limits and practical algorithms,” IEEE Trans. Signal Process., vol. 69, pp. 5723–5738, 2021

    2021

  11. [19]

    What is the state of neural network pruning?

    D. Blalock et al. , “What is the state of neural network pruning?” Proc. of Machine Learning and Syst. , vol. 2, pp. 129–146, 2020

    2020

  12. [20]

    Deep residual learning for image recognition,

    K. He et al. , “Deep residual learning for image recognition,” in Proc. IEEE CVPR , June Las Vegas, NV , USA, 2016

    2016

  13. [21]

    Importance Estimation for Neural Network Pruning,

    P. Molchanov et al. , “Importance Estimation for Neural Network Pruning,” in Proc. IEEE/CVF CVPR , Los Alamitos, CA, USA, Jun. 2019, pp. 11 256–11 264

    2019

  14. [22]

    Time vs. frequency domain DPD for massive MIMO: Methods and performance analysis,

    Y . Wuet al. , “Time vs. frequency domain DPD for massive MIMO: Methods and performance analysis,” IEEE Trans. Wireless Commun. , pp. 1–1, 2025

    2025

  15. [23]

    Number of FLOPs of training DNNs for learning precoding,

    P. Cong and C. Yang, “Number of FLOPs of training DNNs for learning precoding,” in Proc. IEEE VTC-Spring , Florence, Italy, 2023, pp. 1–6

    2023

  16. [24]

    A comparative analysis of the complexity/accuracy tradeoff in power amplifier behavioral models,

    A. S. Tehrani et al., “A comparative analysis of the complexity/accuracy tradeoff in power amplifier behavioral models,” IEEE Trans. Microwave Theory Techn., vol. 58, no. 6, pp. 1510–1520, 2010

    2010

  17. [25]

    Deep learning based power control for cell- free massive MIMO with MRT,

    L. Sala ¨un and H. Yang, “Deep learning based power control for cell- free massive MIMO with MRT,” in Proc. IEEE GLOBECOM , Madrid, Spain, 2021

    2021

  18. [26]

    Predicting parameters in deep learning,

    M. Denil et al. , “Predicting parameters in deep learning,” Advances in Neural Information Processing Systems , vol. 26, 2013

    2013

  19. [27]

    Rethinking the value of network pruning,

    Z. Liu et al. , “Rethinking the value of network pruning,” in International Conference on Learning Representations , 2019. [Online]. Available: https://openreview.net/forum?id=rJlnB3C5Ym

    2019

  20. [28]

    Learning both weights and connections for efficient neural network,

    S. Han et al. , “Learning both weights and connections for efficient neural network,” Advances in Neural Information Processing Systems , vol. 28, 2015

    2015

  21. [29]

    Physical Channels and Modulation (Release 18),

    3GPP, “Physical Channels and Modulation (Release 18),” Tech. Spec. TS 38.211, v18.5.0, Dec. 2024

    2024

  22. [30]

    To prune, or not to prune: exploring the efficacy of pruning for model compression,

    M. Zhu and S. Gupta, “To prune, or not to prune: exploring the efficacy of pruning for model compression,” 2017. [Online]. Available: https://arxiv.org/abs/1710.01878

    2017 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.