Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Learning to Gridize: Segment Physical World by Wireless Communication Channel

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that partitioning coverage zones by channel spectra, not location, lets a network predict beam-level received power under new beam patterns without retraining — cutting error by 30% on active grids and 65% overall.

desk verdict A solid application paper with a genuine cross-beam real-world test, but the physical CAPS-recovery story is not backed by the synthetic results. read the letter →

arxiv 2507.15386 v1 pith:LSEJERV7 submitted 2025-07-21 cs.LG eess.SP

Juntao Wang , Feng Yin , Tian Ding , Tsung-Hui Chang , Zhi-Quan Luo , Qi Yan This is my paper · ORCID
classification cs.LGeess.SP
keywords channelspacegridizationanglepowerspectrum(CAPS)vectorquantizedautoencoderlocalizedstatisticalmodelRSRPpredictionbeampatterntransferwirelessnetworkoptimizationPIDAtrainingscheme
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 wireless operators can partition their coverage area into grids of users who genuinely share the same propagation channel, using nothing but the beam-level reference signal received power (RSRP) reports the network already collects. The proposal is to estimate each measurement point's channel angle power spectrum (CAPS) — the distribution of path gain over departure angles — and to cluster in that hidden channel space rather than in geographical or signal-strength space, where different channels can look identical. The paper's autoencoder couples this estimation to a sparse codebook of grid centers, with a fixed physics-based decoder, so the only unknowns are the channel spectra and the grid assignments. On a real-world walking-test dataset, the method improves RSRP prediction under beam patterns never seen in training by 30% for active grids and 65% overall relative to the strongest baselines using the same data. This is the first claimed unification of channel estimation and gridization, and it matters because it removes the two things that make current practice expensive or unreliable: location data, and the assumption that equal signal strength means equal channel.

What carries the argument

The load-bearing object is the identity $y = A x$ linking beam-level RSRP to the channel angle power spectrum, with $A$ the $M \times N$ beam pattern matrix fixed by the antenna array and $x$ the sparse nonnegative CAPS over $N = N_V \times N_H$ discrete angles. Three components carry the argument: the encoder $E_\Theta$ maps dB-scale RSRP to nonnegative CAPS; the quantizer $Q_\Xi$ assigns each CAPS to the nearest of $K$ codebook columns, sparsified by keeping only the $L$ largest entries and applying ReLU; and the decoder is fixed to the localized statistical channel model, reconstructing RSRP as $10\log_{10}(A x)$. Around this architecture, the PIDA training scheme — encoder pretraining, K-means codebook initialization, detached encoder/codebook gradient flows, and asynchronous assignment refresh — is what prevents codebook collapse and keeps the quantization objective from fighting the reconstruction objective, which is why the joint formulation can be trained end-to-end at all.

What would settle it

Run a ray-tracing simulation with known scatterer geometry so the true CAPS is available: train CSG-AE on RSRP synthesized under one beam pattern, then evaluate its predicted RSRP for a second beam pattern against the simulator's actual measurements under that pattern. If the error under the unseen pattern is no better than a trivial baseline that re-weights the observed average RSRP by the beam gains, or if two training runs with different random seeds produce CAPS with disjoint angular supports while both reconstruct the training RSRP equally well, then the underlying CAPS is not identifiable from the linear model and the transfer claim fails.

Watch

Extended reading notes

Core claim

Stated on the paper's own terms, the discovery is that gridization belongs in channel space. With the linear model $y_i = A x_i$, where $A \in \mathbb{R}^{M \times N}$ is the known beam pattern matrix and $x_i$ is the nonnegative, approximately sparse CAPS of sample $i$, the environment is summarized by $K$ sparse centers $x^{(k)}$ such that each sample is the sum of its assigned center and a perturbation whose dominant-path components have zero mean. The paper claims that a vector-quantized autoencoder (VQ-VAE, i.e., a neural discrete representation learner) with a fixed LSCM decoder solves this joint estimation-and-clustering problem without any location information, provided the training scheme stabilizes the discrete codebook. The transfer claim is one-shot: once trained under beam pattern $A$, the grid-average CAPS $\bar{x}^*_{(k)}$ predicts average RSRP under any new pattern $A_{\mathrm{new}}$ as $A_{\mathrm{new}} \bar{x}^*_{(k)}$, with no retraining. The real-world experiment, training on round one of a ten-round walking test and testing on the other nine rounds under distinct beam patterns, is what supports the reported 30% active / 65% overall MAE reduction over the beam-space baseline.

Load-bearing premise

The whole transfer argument rests on the assumption that the true channel spectrum can be recovered from far fewer beam measurements than angular unknowns; if the CAPS learned under one beam pattern is merely one of many spectra consistent with those measurements, predictions under a new beam pattern can be wrong even when the model fits its training data perfectly.

Editorial extensions

If this is right

  • Operators can retire drive tests for gridization: on the reported real dataset, CSG-AE running on ordinary RSRP reports cuts Active MAE from about 6.5 dB to about 4.6 dB and Overall MAE from 35.2 dB to 12.3 dB versus the best baselines that use the same input data.
  • Beam-pattern selection becomes cheap: because each grid is stored as one average CAPS, the optimization problem in (28) reduces to choosing $A$ to maximize $A \sum_k \bar{x}^*_{(k)}$, so candidate configurations can be compared without new measurements or simulations.
  • Grid-level optimization no longer depends on which users are present: channel-based grids summarize the stable localized environment, so the grid-level utility formulation in (2) stays valid as individual users come and go.
  • The PIDA training scheme is a transferable ingredient: encoder pretraining, K-means codebook initialization, and detached asynchronous updates keep codebook utilization above 95%, where naive end-to-end training collapses to roughly 26% active codewords.

Reading between the lines

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

  • If CAPS transfer across beam patterns generalizes beyond the walking test, the same encoder-quantizer template could support offline evaluation of precoding and beamforming candidates on measurement reports already sitting in the network, effectively turning gridization into a low-cost digital twin of the propagation environment.
  • A direct test the paper does not run: train on data from two beam patterns at once and check whether the encoder maps co-located samples to the same CAPS; agreement would be direct evidence that the null-space ambiguity of $A$ is benign in practice.
  • The architecture is a generic amortized solver for underdetermined inverse problems with a known forward operator and a clustering objective, so the template — encoder, sparse codebook quantizer, fixed physics decoder, detached asynchronous training — could carry to other sensing domains such as tomography or electromagnetic mapping.
  • The reported pattern of higher Homogeneity and lower Completeness relative to geographical grids hints that channel-space grids are a genuinely different partition of the physical world; correlating those grids with operational indicators such as handover rates or interference would test which partition is more useful in practice.
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

4 major / 5 minor

Summary. The paper introduces Channel Space Gridization (CSG), a framework that clusters RSRP measurement samples into grids whose members share similar channel characteristics, without using location data. The proposed CSG-AE is an autoencoder whose encoder maps beam-level RSRP to a high-dimensional CAPS vector, whose quantizer learns K sparse grid centers, and whose decoder is the fixed LSCM linear map from CAPS to RSRP. A training scheme called PIDA (pretraining, K-Means initialization, detached and asynchronous updates) is proposed to avoid codebook collapse and training instability. The paper evaluates CSG-AE on synthetic data and on a 10-round real-world walking test, reporting improved RSRP prediction on held-out beam patterns, with 30% lower Active MAE and 65% lower Overall MAE than BSG baselines, together with improved clustering metrics and active ratio.

Significance. If the central claim holds, this is a practically useful contribution: it would enable grid-level network optimization from ubiquitous RSRP measurements alone, without costly drive tests or user location data, and would allow RSRP prediction under new beam patterns without retraining. The real-world evaluation is a genuine external test, since the model is trained on round 1 and evaluated on rounds 2-10 with beam patterns never seen during training, and the RSRP prediction is not a fitted quantity. The PIDA training scheme is supported by ablation studies and codebook-utilization curves. The paper also clearly identifies a real limitation of BSG, namely that RSRP similarity does not imply channel similarity. However, the core transfer claim relies on recovering the true physical CAPS from an underdetermined linear system, and the current evidence does not yet establish that the learned CAPS are physically identifiable rather than merely consistent with the training beam pattern.

major comments (4)
  1. [Section IV-D, Eqs. (29)-(31)] The one-shot transfer claim presupposes that the CAPS estimates are identifiable from y = A x even though A has M << N rows. The training objective L1 in Eq. (18) penalizes only beam-space mismatch, so any null-space vector n with A n = 0 can be added to a solution without changing L1. When the learned CAPS are multiplied by a new beam pattern A_new in Eq. (30), those invisible null-space components become active and produce uncontrolled prediction error. The paper provides no RIP/incoherence analysis, no sparse-recovery guarantee, and Assumption 1 constrains only the grid centers, not the per-sample CAPS produced by the encoder; the encoder's ReLU output can be dense. Since the ability to predict under other beam patterns is advertised as a central contribution, this identifiability gap is load-bearing and needs either a theoretical argument or a dedicated empirical test with ground-truth CAPS under held-out beam patterns.
  2. [Section V-A, Fig. 8] The synthetic experiments do not test the one-shot cross-beam scenario. The ground-truth CAPS samples are generated and evaluated under the same beam pattern A that is used to train the model, and the CAPS estimation metrics (Sample Mean NMSE and Center Wasserstein Distance) are computed on that same A. Consequently, these results cannot distinguish a CAPS estimate that matches the true physical channel from one that differs by a null-space component of A. A synthetic version of the Section IV-D procedure, training under A and evaluating RSRP prediction under a held-out A_new against ground-truth RSRP, is needed to validate the transfer mechanism. In addition, the Sample Mean NMSE values in Fig. 8b appear to be around 1.0, meaning the normalized error energy is comparable to the signal energy even under the training beam pattern; this is weak absolute evidence for the paper's claim of 'exceptional CAPS estimation accuracy' and should be discussed explicitly.
  3. [Section V-B-3b] Using K-Means on location measurements as surrogate ground-truth labels for channel similarity is in tension with the paper's own premise that location is not a reliable proxy for channel characteristics. The paper acknowledges the converse direction, but the clustering metrics (ARI, NMI, Homogeneity, V-measure) are still reported as evidence of channel consistency. The authors should either justify why this surrogate is informative despite the premise, or provide an evaluation of clustering quality that does not depend on location labels, for example by measuring within-grid predictive consistency of RSRP under the held-out beam patterns or by comparing per-grid channel estimates obtained from independent data.
  4. [Section V-B-4a, Eq. (34)] The headline 30% Active MAE and 65% Overall MAE improvements are sensitive to the evaluation protocol. The active-grid sets differ across methods: BSG has a low active ratio, so its Overall MAE is inflated by many grids with zero real average RSRP, while CSG's higher active ratio removes many of those grids from the comparison. Since grid assignments are method-dependent, the two methods are not evaluated on the same set of grids. To make the comparison fair and interpretable, the authors should report prediction error on a common fixed partition or on the union/intersection of active grids, and should report per-grid statistics over all K grids without the active/inactive split.
minor comments (5)
  1. [Section IV-C-3b] There is a typo in 'inherent in the naive training shceme'; it should be 'scheme'.
  2. [Fig. 8 legend] The legend lists 'NNOMP-KMeans-X' while the text and Section V-A-2 refer to 'NOMP-KMeans-X'; the naming should be consistent.
  3. [Section IV-D, Eq. (31)] The expression 'max_{A in C(A)} A sum_{k in K} xbar*_{(k)}' is missing parentheses and is notationally ambiguous; it should be written as a sum over grids of A applied to each grid center.
  4. [Section V-B-4b] The claim that higher Homogeneity 'indicates that our grids contain more samples sharing the same surrogate label and thus exhibit greater channel similarity' conflates agreement with the location-based surrogate with actual channel similarity; this is a presentation issue that should be clarified in light of Major Comment 3.
  5. [Section II-A, Eq. (3)] The notation indexing x_{v,h} and the vector ordering in Eq. (5) are not explicitly reconciled; a sentence stating the vectorization order would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central transfer prediction is trained on round 1 and tested on unseen beam patterns; cited LSCM is external support.

full rationale

I walked the paper's claimed derivation chain and found no step that reduces by construction to its own inputs. The encoder is trained only on first-round RSRP samples under beam pattern A(1), and the RSRP predictions for test rounds 2-10 are computed under unseen beam patterns A(r) using Eq. (30) and the evaluation in Eq. (34); thus the reported 30% Active MAE and 65% Overall MAE improvements are held-out predictions rather than fitted quantities. The reconstruction loss L1 in Eq. (18) is a standard data-fidelity objective, and the quantization loss L2 in Eqs. (19) and (26) is a standard clustering objective that by design makes codebook centers track projected sample averages; this is a modeling assumption, not an assumed empirical result, so it is not circular. The linear physical model y = A(Ψ)x in Eq. (4) is cited from [5] with stated antenna parameters and is not fitted to the paper's target claims; although there is author overlap with [5], the cited model is a published, parameter-free physical expectation and therefore independent support rather than load-bearing self-citation. The paper does not invoke a uniqueness theorem from its authors' prior work, and no ansatz is smuggled in via citation. The underdetermined nature of A (M << N) raises a legitimate identifiability and generalization concern for the one-shot transfer claim, but that is a correctness or validity risk, not a circular reduction: the paper's equations do not make the predicted RSRP equal to the training objective by construction. Consequently, no circular step is identified.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the LSCM linear model, sparsity and zero-mean assumptions, and on the implicit identifiability of CAPS from low-dimensional RSRP measurements. K, L, and w1/w2 are hand-chosen inputs or unreported hyperparameters.

free parameters (3)
  • K (number of grids) = 100
    Chosen by hand in all experiments; the gridization result and evaluation metrics depend on it.
  • L (sparsity of CAPS grid centers) = 5 (synthetic, likely real)
    Assumed sparsity level in Assumption 1; used in hard thresholding of codebook; not estimated from data.
  • loss weights w1, w2 = unspecified
    Weights in LCSG-AE (Eq. 24) that balance reconstruction and quantization; values are not reported in experiments, affecting reproducibility.
assumptions (6)
  • domain assumption LSCM linear model: y = A x (Eq. 4)
    Borrowed from [5]; assumes expectation over random phase errors yields a deterministic linear map from CAPS to RSRP.
  • domain assumption Assumption 1: each CAPS grid center is L-sparse (Eq. 9)
    Physical scatterers are limited; used to constrain codebook via hard thresholding.
  • domain assumption Assumption 2: zero-mean perturbations on dominant paths (Eq. 15)
    Dominant paths stable over time; used to define L2 loss and the projected average center.
  • ad hoc to paper CAPS identifiability from M RSRP measurements (implicit in Eq. 7)
    The linear map A has M rows with M << N; the paper does not prove that minimizing reconstruction loss yields the true CAPS, yet one-shot RSRP prediction under new beam patterns relies on this.
  • domain assumption Grid centers are stable and independent of beam pattern (Sec. IV-D)
    Used to predict RSRP under A_new as A_new times average CAPS; not validated with ground-truth CAPS on real data.
  • ad hoc to paper K-Means on location is a valid surrogate for channel similarity in real-world clustering evaluation (Sec. V-B-3b)
    Used as ground truth labels for NMI/ARI; proximity of location does not guarantee identical channel characteristics, so clustering metrics may not reflect true channel consistency.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Gridize: Segment Physical World by Wireless Communication Channel." pith.science (2026). https://pith.science/paper/LSEJERV7

@misc{pith2026250715386,
  author       = {Pith},
  title        = {Pith review of: Learning to Gridize: Segment Physical World by Wireless Communication Channel},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LSEJERV7}},
  note         = {Machine review of arXiv:2507.15386}
}
read the original abstract

Gridization, the process of partitioning space into grids where users share similar channel characteristics, serves as a fundamental prerequisite for efficient large-scale network optimization. However, existing methods like Geographical or Beam Space Gridization (GSG or BSG) are limited by reliance on unavailable location data or the flawed assumption that similar signal strengths imply similar channel properties. We propose Channel Space Gridization (CSG), a pioneering framework that unifies channel estimation and gridization for the first time. Formulated as a joint optimization problem, CSG uses only beam-level reference signal received power (RSRP) to estimate Channel Angle Power Spectra (CAPS) and partition samples into grids with homogeneous channel characteristics. To perform CSG, we develop the CSG Autoencoder (CSG-AE), featuring a trainable RSRP-to-CAPS encoder, a learnable sparse codebook quantizer, and a physics-informed decoder based on the Localized Statistical Channel Model. On recognizing the limitations of naive training scheme, we propose a novel Pretraining-Initialization-Detached-Asynchronous (PIDA) training scheme for CSG-AE, ensuring stable and effective training by systematically addressing the common pitfalls of the naive training paradigm. Evaluations reveal that CSG-AE excels in CAPS estimation accuracy and clustering quality on synthetic data. On real-world datasets, it reduces Active Mean Absolute Error (MAE) by 30\% and Overall MAE by 65\% on RSRP prediction accuracy compared to salient baselines using the same data, while improving channel consistency, cluster sizes balance, and active ratio, advancing the development of gridization for large-scale network optimization.

Figures

Figures reproduced from arXiv: 2507.15386 by the authors.

Figure 1
Figure 1. (a) Sample-level vs. grid-level network optimization. (b) Different gridization methods in wireless communication networks. • Robustness—Environmental Modeling and Reusability: Gridization aggregates inherently random measurements into meaningful clusters, forming a reusable “digital asset” that characterizes the communication environment, where each cluster of data corresponds to a subregion of the network coverage… view at source ↗
Figure 2
Figure 2. (a) Discrete angular space. (b) Visualization of a beam pattern matrix [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An illustration of the zero-mean perturbation assumption. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The architecture of the CSG-AE. The encoder is a neural network [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of the centers x(k) (triangles) and the projected embeddings PN(ki ) (xi) (circles) using t-SNE [22]. Colors represent the cluster labels. The first row illustrates the states under different initialization strategies at the beginning of training and imme…
Figure 6
Figure 6. Figure 6: Codebook utilization dynamics across training strategies. The naive [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visualization of synthetic dataset under scale factor [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 9
Figure 9. Figure 9: Illustration of the walking test dataset collection process (round [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Performance comparison of different methods on real-world dataset. All methods run [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 28 canonical work pages

  1. [5]

    A physics-based and data-driven approach for localized statistical channel modeling,

    S. Zhang, X. Ning, X. Zheng, Q. Shi, T.-H. Chang, and Z.-Q. Luo, “A physics-based and data-driven approach for localized statistical channel modeling,” IEEE Transactions on Wireless Communications , vol. 23, no. 6, pp. 5409–5424, 2024

    work page 2024

  2. [1]

    SRCON: A data-driven network performance simulator for real-world wireless networks,

    Z.-Q. Luo, X. Zheng, D. L ´opez-P´erez, Q. Yan, X. Chen, N. Wang, Q. Shi, T.-H. Chang, and A. Garcia-Rodriguez, “SRCON: A data-driven network performance simulator for real-world wireless networks,” IEEE Communications Magazine, vol. 61, no. 6, pp. 96–102, 2023

    work page 2023

  3. [2]

    A survey of recent advances in optimization methods for wireless communications,

    Y .-F. Liu, T.-H. Chang, M. Hong, Z. Wu, A. Man-Cho So, E. A. Jorswieck, and W. Yu, “A survey of recent advances in optimization methods for wireless communications,” IEEE Journal on Selected Areas in Communications, vol. 42, no. 11, pp. 2992–3031, 2024

    work page 2024

  4. [3]

    Parallel digital twin-driven deep rein- forcement learning for user association and load balancing in dynamic wireless networks,

    Z. Tao, W. Xu, and X. You, “Parallel digital twin-driven deep rein- forcement learning for user association and load balancing in dynamic wireless networks,” arXiv preprint arXiv:2410.07611 , 2024

    arXiv 2024

  5. [4]

    Multi-grid-based localized statistical channel modeling: A radio map approach,

    X. Ning, S. Zhang, X. Zheng, and T.-H. Chang, “Multi-grid-based localized statistical channel modeling: A radio map approach,” in 2022 International Symposium on Wireless Communication Systems (ISWCS) , 2022, pp. 1–6

    work page 2022

  6. [6]

    FedLoc: Federated learning framework for data-driven cooperative localization and location data processing,

    F. Yin, Z. Lin, Q. Kong, Y . Xu, D. Li, S. Theodoridis, and S. R. Cui, “FedLoc: Federated learning framework for data-driven cooperative localization and location data processing,” IEEE Open Journal of Signal Processing, vol. 1, pp. 187–215, 2020

    work page 2020

  7. [7]

    NR; Requirements for Support of Radio Resource Manage- ment,

    3GPP, “NR; Requirements for Support of Radio Resource Manage- ment,” 3rd Generation Partnership Project (3GPP), Technical Specifi- cation (TS) 38.133, 2024, version 18.6.0

    work page 2024

  8. [8]

    Spatial partitioning method, apparatus, device, medium, and program product,

    F. Yin, Y . Sun, R. Jiang, W. Yan, J. Wang, and Z.-Q. Luo, “Spatial partitioning method, apparatus, device, medium, and program product,” Chinese Patent CN114 430 530A, Jun. 21, 2022

    work page 2022

Show all 36 references
  1. [9]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” in Advances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017

    2017

  2. [10]

    Zero-shot text-to-image generation,

    A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. V oss, A. Radford, M. Chen, and I. Sutskever, “Zero-shot text-to-image generation,” in International Conference on Machine Learning . PMLR, 2021, pp. 8821–8831

    2021

  3. [11]

    Jukebox: A generative model for music,

    P. Dhariwal, H. Jun, C. Payne, J. W. Kim, A. Radford, and I. Sutskever, “Jukebox: A generative model for music,” arXiv preprint arXiv:2005.00341, 2020

    2005 arXiv

  4. [12]

    Tokenflow: Unified image tokenizer for multi- modal understanding and generation,

    L. Qu, H. Zhang, Y . Liu, X. Wang, Y . Jiang, Y . Gao, H. Ye, D. K. Du, Z. Yuan, and X. Wu, “Tokenflow: Unified image tokenizer for multi- modal understanding and generation,” in Proceedings of the Computer Vision and Pattern Recognition Conference , 2025, pp. 2545–2555

    2025

  5. [13]

    VQ-V AE empowered wireless commu- nication for joint source-channel coding and beyond,

    M. Nemati, J. Park, and J. Choi, “VQ-V AE empowered wireless commu- nication for joint source-channel coding and beyond,” in GLOBECOM 2023-2023 IEEE Global Communications Conference. IEEE, 2023, pp. 3155–3160

    2023

  6. [14]

    Cox, An Introduction to 5G: The New Radio, 5G Network and Beyond

    C. Cox, An Introduction to 5G: The New Radio, 5G Network and Beyond. John Wiley & Sons, 2020

    2020

  7. [15]

    Tse and P

    D. Tse and P. Viswanath, Fundamentals of Wireless Communication . Cambridge, U.K.: Cambridge University Press, 2005

    2005

  8. [16]

    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,” 3rd Generation Partnership Project (3GPP), Technical Report (TR) 38.901, 2024, version 18.6.0

    2024

  9. [17]

    D. P. Bertsekas, Nonlinear Programming, 3rd ed. Belmont, MA: Athena Scientific, 2016

    2016

  10. [18]

    Inexact block coordinate descent algorithms for nonsmooth nonconvex optimization,

    Y . Yang, M. Pesavento, Z.-Q. Luo, and B. Ottersten, “Inexact block coordinate descent algorithms for nonsmooth nonconvex optimization,” IEEE Transactions on Signal Processing , vol. 68, pp. 947–961, 2019

    2019

  11. [19]

    Proximal algorithms,

    N. Parikh, S. Boyd et al. , “Proximal algorithms,” Foundations and Trends in Optimization, vol. 1, no. 3, pp. 127–239, 2014

    2014

  12. [20]

    R. Shu. (2017) Amortized optimization. [Online]. Available: https: //ruishu.io/2017/11/07/amortized-optimization/

    2017

  13. [21]

    Tutorial on amortized optimization,

    B. Amos, “Tutorial on amortized optimization,” Foundations and Trends in Machine Learning , vol. 16, no. 5, pp. 592–732, 2023

    2023

  14. [22]

    Visualizing data using t-sne,

    L. van der Maaten and G. Hinton, “Visualizing data using t-sne,” Journal of Machine Learning Research , vol. 9, no. 86, pp. 2579–2605, 2008. [Online]. Available: http://jmlr.org/papers/v9/vandermaaten08a.html

    2008

  15. [23]

    Fast decoding in sequence models using discrete latent variables,

    L. Kaiser, S. Bengio, A. Roy, A. Vaswani, N. Parmar, J. Uszkoreit, and N. Shazeer, “Fast decoding in sequence models using discrete latent variables,” in International Conference on Machine Learning . PMLR, 2018, pp. 2390–2399

    2018

  16. [24]

    Theory and experiments on vector quantized autoencoders,

    A. Roy, A. Vaswani, A. Neelakantan, and N. Parmar, “Theory and experiments on vector quantized autoencoders,” arXiv preprint arXiv:1805.11063, 2018. 16

    2018 arXiv

  17. [25]

    Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks,

    M. Huh, B. Cheung, P. Agrawal, and P. Isola, “Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks,” in International Conference on Machine Learning. PMLR, 2023, pp. 14 096–14 113

    2023

  18. [26]

    Least squares quantization in PCM,

    S. Lloyd, “Least squares quantization in PCM,” IEEE Transactions on Information Theory, vol. 28, no. 2, pp. 129–137, 1982

    1982

  19. [27]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  20. [28]

    torch.tensor.detach

    PyTorch. torch.tensor.detach. [Online]. Available: https://docs.pytorch. org/docs/stable/generated/torch.Tensor.detach.html

  21. [29]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101 , 2017

    2017 arXiv

  22. [30]

    Discrete latent variables,

    C. M. Bishop and H. Bishop, “Discrete latent variables,” in Deep Learning: Foundations and Concepts . Springer, 2024, pp. 459–493

    2024

  23. [31]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Łukasz Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017

    2017

  24. [32]

    Inductive representation learning on large graphs,

    W. L. Hamilton, R. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” in Advances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017

    2017

  25. [33]

    Attentional graph neural networks for robust massive network localization,

    W. Yan, J. Wang, F. Yin, Y . Tian, and A. M. Zoubir, “Attentional graph neural networks for robust massive network localization,” arXiv preprint arXiv:2311.16856, 2023

    2023 arXiv

  26. [34]

    Comparing partitions,

    L. Hubert and P. Arabie, “Comparing partitions,” Journal of Classifica- tion, vol. 2, pp. 193–218, 1985

    1985

  27. [35]

    Cluster ensembles—a knowledge reuse frame- work for combining multiple partitions,

    A. Strehl and J. Ghosh, “Cluster ensembles—a knowledge reuse frame- work for combining multiple partitions,” Journal of Machine Learning Research, vol. 3, no. Dec, pp. 583–617, 2002

    2002

  28. [36]

    V-measure: A conditional entropy- based external cluster evaluation measure,

    A. Rosenberg and J. Hirschberg, “V-measure: A conditional entropy- based external cluster evaluation measure,” in Proceedings of the 2007 joint conference on empirical methods in natural language processing and computational natural language learning (EMNLP-CoNLL) , 2007, pp. 410–420

    2007

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.