Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

CSI2Vec: Towards a Universal CSI Feature Representation for Positioning and Channel Charting

T0 review · 4 major / 5 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 A 16-dimensional CSI vector embedding can replace 1024-dimensional features for positioning and channel charting with only a small accuracy loss.

desk verdict CSI2Vec is a solid, useful compact embedding method with a clear experimental story, but the 'universal' claim and the temporal-proxy training signal need more evidence before I'd trust the headline. read the letter →

arxiv 2506.05237 v1 pith:JHSYH3MV submitted 2025-06-05 cs.IT eess.SPmath.IT

Victoria Palhares , Sueda Taner , Christoph Studer This is my paper · ORCID
classification cs.ITeess.SPmath.IT
keywords CSI2Vecchannelstateinformationself-supervisedrepresentationlearningtripletlosspositioningchartingdataaugmentationMIMO
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes CSI2Vec, a self-supervised method that turns channel state information (CSI) from many different wireless deployments into one shared, low-dimensional vector per user position. The central claim is that a 16-dimensional embedding trained across simulated and measured scenarios preserves spatial relationships well enough that positioning and channel charting nearly match a scenario-specific network using the full 1024-dimensional CSI features. The authors argue that this removes the need for CSI reconstruction or ground-truth positions during representation learning, and that the same embedding network generalizes across environments, antenna counts, and subcarrier allocations. If true, a single compact code could serve as a universal front end for location-based wireless tasks, reducing storage and computation by a factor of 64.

What carries the argument

The load-bearing mechanism is the triplet loss with a temporal-proximity proxy for spatial proximity, combined with cross-scenario data augmentation. For each anchor CSI sample, a close sample is one recorded within a scenario-specific time window $T_c$, and a far sample is one recorded much later in the same scenario or taken from a different scenario; the loss $\mathcal{L}_t = \frac{1}{|\mathcal{T}|}\sum \left(\|v_n - v_c\| - \|v_n - v_f\| + M_t\right)_+$ then shapes the embedding. The input side zero-pads all CSI tensors to common maximum dimensions, randomly drops antennas and subcarrier bands, adds noise, transforms to the delay domain with an inverse discrete Fourier transform, truncates to 16 taps, and extracts hardware-robust large-scale-fading features. This combination lets one small multi-layer perceptron with a 32-unit hidden layer and output dimension $D' = 16$ serve as a universal encoder across scenarios.

What would settle it

Record a trajectory that passes through the same physical location at two very different timestamps (a loop or a repeated path). If CSI2Vec places those two visits far apart in embedding space while a position-supervised model places them close, the time-based proxy is overriding geometry; a quantitative check compares pairwise embedding distances between same-location visits against distances between genuinely different locations.

Watch

Extended reading notes

Core claim

CSI2Vec's core discovery is that a contrastive, timestamp-based objective is enough to learn a universal spatial representation of wireless channels. The network maps delay-domain, large-scale-fading CSI features into a 16-dimensional latent space such that CSI samples recorded at nearby times, and hence nearby positions, are pulled together, while samples far apart in time or drawn from different scenarios are pushed apart. Because training mixes scenarios with different deployment geometries and radio hardware, the resulting embedding abstracts away the radio setup while keeping location information. On a ray-traced outdoor scene, a ray-traced indoor office, and a measured factory, CSI2Vec with augmentation positions users with mean errors only 0.898 m, 0.493 m, and 0.566 m above the fully supervised end-to-end baseline, and produces channel charts of comparable quality, while autoencoder-based embeddings without spatial constraints mostly fail. The paper concludes that reconstruction-based CSI compression is unnecessary, and even harmful, for positioning and channel charting.

Load-bearing premise

The whole self-supervised signal rests on equating time with space: samples recorded close in time are assumed to come from nearby positions, so a trajectory that revisits a location later would teach the embedding to separate the same physical place.

Editorial extensions

If this is right

  • A single 16-dimensional embedding per CSI sample can replace 1024-dimensional features for positioning and channel charting, shrinking the downstream network from activation counts $\{1024,320,160,80,40,20,10,5,2\}$ to $\{1024,32,16,12,8,6,4,2\}$.
  • Data augmentation that randomly drops antennas and subcarrier bands makes the embedding robust to different radio hardware and OFDMA-style subcarrier allocations, so one encoder works across deployments.
  • Autoencoder-based CSI compression does not preserve spatial relationships and fails at positioning and channel charting in most tested scenarios, so reconstructing CSI is not needed for these tasks.
  • Semi-supervised training with ground-truth positions from one scenario yields only a slight gain and only in that scenario, so the fully self-supervised version is sufficient in most cases.
  • Because the embedding is compact and discards hardware-specific details, it can double as a lightweight, privacy-friendly interface between user equipment and network.

Reading between the lines

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

  • Beyond the paper: the timestamp proxy is the obvious failure mode, so a natural stress test is a looping trajectory; if the embedding separates same-location revisits, it is encoding trajectory order rather than geography.
  • Beyond the paper: because the encoder is scenario-agnostic, the same embedding should transfer zero-shot to a new deployment without retraining, provided the input is zero-padded to the same dimensions; the paper's three-scenario test suggests but does not prove this.
  • Beyond the paper: the universal embedding could be attached to small task heads for other spatial radio problems, such as radio environment mapping or handover prediction, reusing the same encoder.
  • Beyond the paper: the augmentation strategy is effectively a form of channel dropout; testing stronger variants, such as masking entire access points, could reveal how far the robustness argument extends.
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 proposes CSI2Vec, a self-supervised framework that learns a compact D'-dimensional CSI embedding using a triplet loss. Training samples are collected across three heterogeneous scenarios (two ray-tracing, one measured). Positive pairs are samples with close timestamps within a scenario; negative pairs are temporally distant samples or samples from other scenarios. After embedding, scenario-specific MLPs perform positioning and channel charting. The authors report that with D'=16, CSI2Vec-AUG achieves MDE values of 4.026 m, 1.164 m, and 1.191 m on the outdoor, indoor, and measured scenarios, compared with 3.128 m, 0.671 m, and 0.625 m for the fully supervised SCS-EE baseline, and substantially outperforms AE-based embeddings for most channel-charting tasks.

Significance. The paper addresses an important problem: whether a single, low-dimensional CSI representation can serve multiple environments and downstream tasks. The comparison setup is well designed in several respects: it includes a strong fully supervised baseline (SCS-EE), two AE baselines with and without augmentation, a real-world DICHASUS dataset, and multiple positioning and charting metrics. The 64x dimensionality reduction with moderate positioning loss, if reproducible, is practically valuable. However, the stated universal claim is not established: the shared embedding is only evaluated on the same three scenarios used for training, D'=16 is selected from test-set behavior, and the self-supervised label (temporal proximity) is not validated as a proxy for spatial proximity. The paper's conclusions are therefore stronger than the evidence supports.

major comments (4)
  1. [Sec. II-E, Eqs. (5)-(8)] The triplet loss in Sec. II-E (Eqs. (5)-(8)) uses timestamp differences as a proxy for spatial distance, but the paper never validates this proxy on the actual trajectories. The three trajectories are described as 'meandering' over bounded areas with dense sampling (e.g., 40 cm and 2.5 cm spacings in Table II), so a path can loop back and produce samples with |t_n - t_f| > T_c that are nevertheless spatially close. In such cases the loss actively pushes apart embeddings of nearby positions, and the embedding could encode trajectory time rather than physical location. Since the self-supervised training deliberately avoids ground-truth positions, this proxy is load-bearing for every POS and CC result; the paper should include an empirical check (e.g., conditional distributions of spatial distance given timestamp difference) or an experiment on a trajectory with revisits.
  2. [Sec. V-A and Fig. 6] The 'universal' claim is not tested with a held-out scenario: the CSI2Vec models are trained on all three scenarios and evaluated on the same three scenarios. In addition, the embedding dimension D'=16 is selected from test-set MDE curves in Fig. 6, which optimistically biases the reported numbers. The paper should either evaluate on a left-out scenario or dataset, or at minimum select hyperparameters on a validation split and report the resulting test numbers. The future-work sentence in Sec. VI that applying CSI2Vec to more scenarios is left for the future also implicitly limits the universality claim.
  3. [Sec. II-E, Eq. (6)] The cross-scenario triplet set TOUT pushes embeddings of samples from different scenarios apart. Because the three scenarios have different coordinate systems and no common spatial metric, 'different scenario' does not imply 'spatially far'; this term may force the encoder to separate by scenario ID rather than learn a common spatial structure, which is the opposite of a universal representation. The paper does not analyze this potential conflict or ablate it. Please provide an ablation without TOUT, or evidence that cross-scenario negatives improve the learned representation rather than merely creating scenario clusters.
  4. [Table V and Sec. V-D/VI] In the measured indoor scenario, the AE-AUG baseline achieves a POS MDE of 1.102 m, which is lower than CSI2Vec-AUG's 1.191 m (Table V). The text acknowledges that AEs succeed for POS in this scenario, but the conclusion in Sec. VI that CSI2Vec 'was proven to be better suited' for spatial tasks is too strong given this result. The discussion should account for this exception and clarify that the advantage of CSI2Vec over AEs is primarily in channel charting and in the simulated scenarios.
minor comments (5)
  1. [Fig. 3 caption] The caption labels the method 'CSI2Vec-AUG-SS', but the method is called 'CSI2Vec-AUG-SEMI' in Sec. IV-A3 and in Tbl. III; please use one consistent name.
  2. [Sec. V-A] The sentence 'For training and testing, we use 240 and 120 batches, respectively' is ambiguous; please clarify whether these are epochs, batches per epoch, or total gradient steps.
  3. [References] Reference [53] misspells Hotelling; reference [27] should include the full author list or institutional DOI details for the DICHASUS dataset.
  4. [Abstract and Sec. II-F] The abstract's statement that CSI2Vec captures spatial relationships 'without relying on ... ground-truth position information' is only true for the self-supervised variants; the SEMI variant uses ground-truth positions for scenario 2, as described in Sec. II-F.
  5. [General] No code, model checkpoints, or random seeds are provided, which limits the reproducibility of the reported numbers.

Circularity Check

1 steps flagged · score 2.0 of 10

No circular derivation: the triplet objective (Sec. II-E) uses timestamps while all evaluations use external ground-truth positions, so spatial fidelity is an empirical outcome; the only self-confirmatory element is the test-set-based selection of D′=16, which mildly inflates confidence in that configuration without forcing the central scenario comparisons.

  1. fitted input called prediction [Sec. V-A (Simulation Settings) and Sec. V-F (Impact of Vector Embedding Dimension), Fig. 6, Tbls. III-V]
    "Unless stated otherwise, we set our vector embedding dimension to D′ = 16. ... All results and performance metrics are computed using the test set. ... We observe that for all proposed CSI2Vec-based methods, the MDE decreases as the vector embedding dimension increases, reaching the best performance at D′ = 16. ... We conclude that (i) D′ = 16 is an excellent choice for the CSI2Vec embedding dimension."

    The paper selects the embedding dimension D′=16 by minimizing MDE on the test set, then reports all headline numbers on that same test set. Sec. V-A fixes "D′ = 16" and states that "All results and performance metrics are computed using the test set"; Sec. V-F then presents the MDE-versus-dimension curve and announces that "the best performance at D′=16" is reached, concluding that "D′=16 is an excellent choice." Consequently the optimality of D′=16 is a selected statistic over the evaluation set, and the accompanying claim of "64× smaller" features with small MDE loss is partly self-confirmatory for the reported configuration.

full rationale

Walking the paper's claimed derivation chain, I find no equation-level reduction that makes a prediction equivalent to its inputs by construction. The self-supervised loss in Sec. II-E (Eqs. (5)-(8)) is defined purely on timestamps: samples with |t_n - t_c| ≤ Tc are pulled together, and samples farther apart in time, or from another scenario, are pushed apart. No term of L_t uses ground-truth positions (except in the optional SEMI variant, Sec. II-F, which the main self-supervised claims do not rely on). The evaluations, in contrast, are externally grounded: POS MDE is measured against ground-truth UE positions in meters relative to the SCS-EE baseline (Secs. III-A and IV-C), and the CC metrics (TW, CT, KS, RD) compare embedding-derived neighborhoods with true position neighborhoods. Because the training objective and the evaluation use different quantities (time versus position), spatial fidelity is an empirical outcome that could have failed — and does fail for the AE baselines. The main correctness risk is unvalidated, not circular: the temporal-proximity proxy (close timestamps imply close positions) is assumed throughout, and no self-intersecting or revisiting trajectory is tested, so the embedding may partly encode path time. The self-citations ([15], [16], [48] with C. Studer) are used as building blocks — feature extraction, triplet-based and Siamese CC, and metric definitions — and import no uniqueness theorem or forbidden-alternative argument; they do not raise the score. The one mild self-confirmation is the selection of D′=16 on the same test set that produces the headline numbers (Secs. V-A and V-F), which makes the optimality of that configuration a selected statistic. This affects confidence in the specific configuration, not the central claim that 16-dimensional CSI2Vec embeddings approach SCS-EE accuracy while AE embeddings fail. The "universal" claim is additionally tested only on the three training scenarios, without a held-out scenario — a scope limitation, not a circular step.

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

The method is empirical and does not introduce new physical entities. The central claim rests on four domain or design assumptions and on six hand-set hyperparameters; the embedding dimension is selected using test-set performance, which weakens the reported numbers.

free parameters (6)
  • Triplet margin M_t = 10
    Set jointly for all scenarios in Sec. IV-A with no sensitivity analysis; the loss in Eq. (8) depends directly on it.
  • Subcarrier removal percentage q = 20%
    Fixed in Sec. V-A for the augmentation in Sec. II-B2; no ablation is provided.
  • Delay tap truncation C'_max = 16 taps
    Truncation in Sec. II-C set in Sec. V-A; it determines the input dimension D=1024.
  • Embedding dimension D' = 16
    Chosen in Sec. V-F because it gave the best MDE in Fig. 6, while Sec. V-A states all metrics are computed on the test set.
  • Time close/far threshold T_c = 10, 1, 1 per scenario
    Set per scenario in Secs. V-B, V-C, and V-D with no sensitivity analysis; it defines the triplet labels in Eqs. (5)-(6).
  • Augmentation SNR ranges = [10,21] dB simulated, [25,41] dB measured
    Chosen in Sec. V-A to account for channel estimation errors; the range differs between simulated and measured data.
assumptions (4)
  • domain assumption Timestamp proximity is a proxy for spatial proximity along the UE trajectory
    Triplet labels in Eqs. (5)-(8) define close and far samples purely by timestamps; if trajectories self-intersect, the labels are wrong.
  • domain assumption Large-scale fading features suffice for positioning and channel charting
    Sec. II-C discards small-scale fading by taking element-wise magnitudes of truncated delay-domain CSI and normalizing, following [9], [15].
  • ad hoc to paper Zero-padding plus random removal of antennas and subcarriers mimics variation across deployments and radio setups
    Sec. II-B assumes augmentation lets one model trained on three scenarios generalize to different DSs and RSs; no held-out scenario tests this.
  • ad hoc to paper Pushing embeddings from different scenarios apart does not undermine a single universal representation
    Eq. (6) makes any cross-scenario pair a negative pair, so scenarios are separated in latent space; downstream tasks are still scenario-specific.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CSI2Vec: Towards a Universal CSI Feature Representation for Positioning and Channel Charting." pith.science (2026). https://pith.science/paper/JHSYH3MV

@misc{pith2026250605237,
  author       = {Pith},
  title        = {Pith review of: CSI2Vec: Towards a Universal CSI Feature Representation for Positioning and Channel Charting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JHSYH3MV}},
  note         = {Machine review of arXiv:2506.05237}
}
read the original abstract

Natural language processing techniques, such as Word2Vec, have demonstrated exceptional capabilities in capturing semantic and syntactic relationships of text through vector embeddings. Inspired by this technique, we propose CSI2Vec, a self-supervised framework for generating universal and robust channel state information (CSI) representations tailored to CSI-based positioning (POS) and channel charting (CC). CSI2Vec learns compact vector embeddings across various wireless scenarios, capturing spatial relationships between user equipment positions without relying on CSI reconstruction or ground-truth position information. We implement CSI2Vec as a neural network that is trained across various deployment setups (i.e., the spatial arrangement of radio equipment and scatterers) and radio setups (RSs) (i.e., the specific hardware used), ensuring robustness to aspects such as differences in the environment, the number of used antennas, or allocated set of subcarriers. CSI2Vec abstracts the RS by generating compact vector embeddings that capture essential spatial information, avoiding the need for full CSI transmission or reconstruction while also reducing complexity and improving processing efficiency of downstream tasks. Simulations with ray-tracing and real-world CSI datasets demonstrate CSI2Vec's effectiveness in maintaining excellent POS and CC performance while reducing computational demands and storage.

Figures

Figures reproduced from arXiv: 2506.05237 by the authors.

Figure 1
Figure 1. CSI2Vec pipeline overview: A UE transmits pilot signals from () [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Ground-truth UE positions for the three presented scenarios (green-to-red gradient colored area), AP positions (blue triangles): (a) simulated outdoor [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Results for the simulated outdoor scenario: (a-e) POS estimates and (f-j) channel charts using CC-SN for both the proposed and baseline methods. The [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results for the simulated indoor scenario: (a-e) POS estimates and (f-j) channel charts using CC-SN for both the proposed and baseline methods. The [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Results for the measurement-based indoor scenario: (a-e) POS estimates and (f-j) channel charts using CC-SN for both the proposed and baseline [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: MDE of both the proposed and baseline methods when varying the vector embedding dimension [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Towards channel foundation models (CFMs): Motivations, methodologies and opportunities

    eess.SP 2025-07 conditional novelty 4.0 of 10

    A survey and position paper proposing channel foundation models, with experiments on two pretrained CSI models showing gains over a vanilla ViT baseline.

Reference graph

Works this paper leans on

62 extracted references · 47 canonical work pages · cited by 1 Pith paper

  1. [1]

    (2022) Wireless InSite

    Remcom, Inc. (2022) Wireless InSite. [Online]. Available: 12 TABLE V POSANDCCPERFORMANCE COMPARISON FOR THE MEASUREMENT-BASED INDOOR SCENARIO. Positioning error [m] Latent space quality metrics Method Downstream task Figure Dim. Mean↓95th percentile↓TW↑CT↑KS↓RD↓ CSI2Vec (across scenarios) POS Fig. 5(a) 16 1.244 2.700 0.915 0.933 0.226 0.832 CC-SN Fig. 5(f...

    work page 2022

  2. [2]

    Efficient estimation of word representations in vector space,

    T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of word representations in vector space,” Sep. 2013. [Online]. Available: https://arxiv.org/abs/1301.3781

    arXiv 2013

  3. [3]

    Dis- tributed representations of words and phrases and their compositionality,

    T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean, “Dis- tributed representations of words and phrases and their compositionality,” inAdvances in Neural Inf. Process. Syst., vol. 26, Dec. 2013

    work page 2013

  4. [4]

    Survey of CSI fingerprinting-based indoor positioning and mobility tracking systems,

    J. M. Rocamora, I. Wang-Hei Ho, W.-M. Mak, and A. P.-T. Lau, “Survey of CSI fingerprinting-based indoor positioning and mobility tracking systems,”IET Signal Process., vol. 14, no. 7, pp. 407–419, May 2020

    work page 2020

  5. [5]

    Non-invasive detection of moving and stationary human with WiFi,

    C. Wu, Z. Yang, Z. Zhou, X. Liu, Y . Liu, and J. Cao, “Non-invasive detection of moving and stationary human with WiFi,”IEEE J. Sel. Areas Commun., vol. 33, no. 11, pp. 2329–2342, Nov. 2015

    work page 2015

  6. [6]

    Towards in-baggage suspicious object detection using commodity WiFi,

    C. Wang, J. Liu, Y . Chen, H. Liu, and Y . Wang, “Towards in-baggage suspicious object detection using commodity WiFi,” in2018 IEEE Conf. on Commun. and Netw. Secur. (CNS), May 2018, pp. 1–9

    work page 2018

  7. [7]

    MAP-CSI: Single-site map-assisted localization using massive MIMO CSI,

    K. Vuckovic, F. Hejazi, and N. Rahnavard, “MAP-CSI: Single-site map-assisted localization using massive MIMO CSI,” inIEEE Global Telecommun. Conf. (GLOBECOM), Oct. 2021, pp. 1–6

    work page 2021

  8. [8]

    CSI-based multi-antenna and multi-point indoor positioning using probability fusion,

    E. Gönülta¸ s, E. Lei, J. Langerman, H. Huang, and C. Studer, “CSI-based multi-antenna and multi-point indoor positioning using probability fusion,” IEEE Trans. Wireless Commun., vol. 21, no. 4, pp. 2162–2176, Apr. 2022

    work page 2022

Show all 62 references
  1. [9]

    Deep convolutional neural networks for massive MIMO fingerprint-based positioning,

    J. Vieira, E. Leitinger, M. Sarajlic, X. Li, and F. Tufvesson, “Deep convolutional neural networks for massive MIMO fingerprint-based positioning,” inIEEE Int. Symp. Personal, Indoor, Mobile Radio Commun. (PIMRC), Oct. 2017, pp. 1–6

    2017

  2. [10]

    Direct localization for massive MIMO,

    N. Garcia, H. Wymeersch, E. G. Larsson, A. M. Haimovich, and M. Coulon, “Direct localization for massive MIMO,”IEEE Trans. Signal Process., vol. 65, no. 10, pp. 2475–2487, May 2017

    2017

  3. [11]

    CSI-based positioning in massive MIMO systems using convolutional neural networks,

    S. D. Bast, A. P. Guevara, and S. Pollin, “CSI-based positioning in massive MIMO systems using convolutional neural networks,” inIEEE Veh. Technol. Conf. Spring (VTC-Spring), May 2020, pp. 1–5

    2020

  4. [12]

    CSI-based fingerprinting for indoor localization: A deep learning approach,

    X. Wang, L. Gao, S. Mao, and S. Pandey, “CSI-based fingerprinting for indoor localization: A deep learning approach,”IEEE Trans. Veh. Technol., vol. 66, no. 1, pp. 763–776, Jan. 2017

    2017

  5. [13]

    CSI-based indoor localization,

    K. Wu, J. Xiao, Y . Yi, D. Chen, X. Luo, and L. M. Ni, “CSI-based indoor localization,”IEEE Trans. Parallel Distrib. Syst., vol. 24, no. 7, pp. 1300–1309, Jul. 2013

    2013

  6. [14]

    Deep learning-based indoor localization using received signal strength and channel state information,

    C.-H. Hsieh, J.-Y . Chen, and B.-H. Nien, “Deep learning-based indoor localization using received signal strength and channel state information,” IEEE Access, vol. 7, pp. 33 256–33 267, Mar. 2019

    2019

  7. [15]

    Channel charting: Locating users within the radio environment using channel state information,

    C. Studer, S. Medjkouh, E. Gönülta¸ s, T. Goldstein, and O. Tirkkonen, “Channel charting: Locating users within the radio environment using channel state information,”IEEE Access, vol. 6, pp. 47 682–47 698, Aug. 2018

    2018

  8. [16]

    Channel charting in real-world coordinates with distributed MIMO,

    S. Taner, V . Palhares, and C. Studer, “Channel charting in real-world coordinates with distributed MIMO,”IEEE Trans. Wireless Commun., pp. 1–1, Apr. 2025

    2025

  9. [17]

    Angle-delay profile-based and timestamp-aided dissimilarity metrics for channel charting,

    P. Stephan, F. Euchner, and S. Ten Brink, “Angle-delay profile-based and timestamp-aided dissimilarity metrics for channel charting,”IEEE Trans. Commun., vol. 72, no. 9, pp. 5611–5625, Sep. 2024

    2024

  10. [18]

    Federated learning for multipoint channel charting,

    P. Agostini, Z. Utkovski, and S. Sta ´nczak, “Federated learning for multipoint channel charting,” inIEEE Int. Workshop Signal Process. Advances Wireless Commun. (SPAWC), Jul. 2022, pp. 1–5. 13

    2022

  11. [19]

    Multipoint channel charting for wireless networks,

    J. Deng, S. Medjkouh, N. Malm, O. Tirkkonen, and C. Studer, “Multipoint channel charting for wireless networks,” inAsilomar Conf. Signals, Syst., Comput., Oct. 2018, pp. 286–290

    2018

  12. [20]

    Improving channel charting with representation -constrained autoencoders,

    P. Huang, O. Castañeda, E. Gönülta¸ s, S. Medjkouh, O. Tirkkonen, T. Gold- stein, and C. Studer, “Improving channel charting with representation -constrained autoencoders,” inProc. IEEE Int. Workshop Signal Process. Advances Wireless Commun. (SPAWC), Jul. 2019, pp. 1–5

    2019

  13. [21]

    Large wireless model (LWM): A foundation model for wireless channels,

    S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (LWM): A foundation model for wireless channels,” Apr. 2025. [Online]. Available: https://arxiv.org/abs/2411.08872

    2025 arXiv

  14. [22]

    Radio foundation models: Pre-training transformers for 5G-based indoor localization,

    J. Ott, J. Pirkl, M. Stahlke, T. Feigl, and C. Mutschler, “Radio foundation models: Pre-training transformers for 5G-based indoor localization,” Oct

  15. [23]

    Multi-environment based meta-learning with CSI fingerprints for radio based positioning,

    A. Foliadis, M. H. Castañeda Garcia, R. A. Stirling-Gallacher, and R. S. Thomä, “Multi-environment based meta-learning with CSI fingerprints for radio based positioning,” inIEEE Wireless Commun. Netw. Conf. (WCNC), Mar. 2023, pp. 1–6

    2023

  16. [24]

    Self-supervised and invariant representations for wireless localization,

    A. Salihu, M. Rupp, and S. Schwarz, “Self-supervised and invariant representations for wireless localization,”IEEE Trans. Wireless Commun., vol. 23, no. 8, pp. 8281–8296, Aug. 2024

    2024

  17. [25]

    A MIMO wireless channel foundation model via CIR-CSI consistency,

    J. Jiang, W. Yu, Y . Li, Y . Gao, and S. Xu, “A MIMO wireless channel foundation model via CIR-CSI consistency,” Mar. 2025. [Online]. Available: https://arxiv.org/abs/2502.11965

    2025 arXiv

  18. [26]

    6G WavesFM: A foundation model for sensing, communication, and localization,

    A. Aboulfotouh, E. Mohammed, and H. Abou-Zeid, “6G WavesFM: A foundation model for sensing, communication, and localization,” Apr

  19. [27]

    CSI dataset dichasus-cf0x: Distributed antenna setup in industrial environment, day 1,

    F. Euchner and M. Gauger, “CSI dataset dichasus-cf0x: Distributed antenna setup in industrial environment, day 1,” 2022. [Online]. Available: https://doi.org/doi:10.18419/darus-2854

    2022 doi

  20. [28]

    RFCam: Uncertainty-aware fusion of camera and Wi-Fi for real-time human identification with mobile devices,

    H. Chen, S. Munir, and S. Lin, “RFCam: Uncertainty-aware fusion of camera and Wi-Fi for real-time human identification with mobile devices,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 6, no. 2, Jul. 2022. [Online]. Available: https://doi.org/10.1145/3534588

    2022 doi

  21. [29]

    A multi-task foundation model for wireless channel representation using contrastive and masked autoencoder learning,

    B. Guler, G. Geraci, and H. Jafarkhani, “A multi-task foundation model for wireless channel representation using contrastive and masked autoencoder learning,” May 2025. [Online]. Available: https://arxiv.org/abs/2505.09160

    2025

  22. [30]

    DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,

    A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,” Feb. 2019. [Online]. Available: https://arxiv.org/abs/1902.06435

    2019 arXiv

  23. [31]

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

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

    2022

  24. [32]

    Deep CSI compression for dual-polarized massive MIMO channels with disentangled representation learning,

    S. Fan, W. Xu, R. Xie, S. Jin, D. W. K. Ng, and N. Al-Dhahir, “Deep CSI compression for dual-polarized massive MIMO channels with disentangled representation learning,”IEEE Trans. Commun., vol. 72, no. 9, pp. 5564–5580, Sep. 2024

    2024

  25. [33]

    CSI-GPT: Integrating generative pre-trained transformer with federated-tuning to acquire downlink massive MIMO channels,

    Y . Zeng, L. Qiao, Z. Gao, T. Qin, Z. Wu, E. Khalaf, S. Chen, and M. Guizani, “CSI-GPT: Integrating generative pre-trained transformer with federated-tuning to acquire downlink massive MIMO channels,” Sep. 2024. [Online]. Available: https://arxiv.org/abs/2406.03438

    2024 arXiv

  26. [34]

    Deep learning for massive MIMO CSI feedback,

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

    2018

  27. [35]

    Deep learning-based CSI feedback approach for time-varying massive MIMO channels,

    T. Wang, C.-K. Wen, S. Jin, and G. Y . Li, “Deep learning-based CSI feedback approach for time-varying massive MIMO channels,”IEEE Wireless Commun. Lett., vol. 8, no. 2, pp. 416–419, Apr. 2019

    2019

  28. [36]

    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,” inIEEE Int. Conf. Commun. (ICC), Jun. 2020, pp. 1–6

    2020

  29. [37]

    Universal auto-encoder framework for MIMO CSI feedback,

    J. So and H. Kwon, “Universal auto-encoder framework for MIMO CSI feedback,” inIEEE Global Telecommun. Conf. (GLOBECOM), Dec. 2023, pp. 01–07

    2023

  30. [38]

    WiFo: Wireless foundation model for channel prediction,

    B. Liu, S. Gao, X. Liu, X. Cheng, and L. Yang, “WiFo: Wireless foundation model for channel prediction,” Mar. 2025. [Online]. Available: https://arxiv.org/abs/2412.08908

    2025 arXiv

  31. [39]

    Dual- propagation-feature fusion enhanced neural CSI compression for massive MIMO,

    S. Zhang, W. Xu, S. Jin, X. You, D. W. K. Ng, and L.-C. Wang, “Dual- propagation-feature fusion enhanced neural CSI compression for massive MIMO,”IEEE Trans. Commun., vol. 71, no. 9, pp. 5182–5198, Sep. 2023

    2023

  32. [40]

    CSI compression using channel charting,

    B. Chatelier, V . Corlay, M. Crussière, and L. L. Magoarou, “CSI compression using channel charting,” Dec. 2024. [Online]. Available: https://arxiv.org/abs/2501.01431

    2024

  33. [41]

    Dynamic model fine-tuning for extreme MIMO CSI compression,

    M. Sattari, D. Gündüz, and T. Svensson, “Dynamic model fine-tuning for extreme MIMO CSI compression,” Jan. 2025. [Online]. Available: https://arxiv.org/abs/2501.18250

    2025 arXiv

  34. [42]

    C2S-AE: CSI to sensing enabled by an auto-encoder-based framework,

    J. Jiang, S. Xu, W. Yu, and Y . Gao, “C2S-AE: CSI to sensing enabled by an auto-encoder-based framework,” Mar. 2025. [Online]. Available: https://arxiv.org/abs/2503.00941

    2025 arXiv

  35. [43]

    Leveraging self-supervised learning for MIMO-OFDM channel representation and generation,

    Z. Liu, J. Chen, Y . Xu, T. Ma, J. Liu, H. Zhou, and D. Niyato, “Leveraging self-supervised learning for MIMO-OFDM channel representation and generation,” May 2024. [Online]. Available: https://arxiv.org/abs/2407.07702

    2024 arXiv

  36. [44]

    Localization in dynamic indoor MIMO-OFDM wireless systems using domain adaptation,

    R. Ismayilov, R. L. G. Cavalcante, and S. Stanczak, “Localization in dynamic indoor MIMO-OFDM wireless systems using domain adaptation,” Aug. 2024. [Online]. Available: https://arxiv.org/abs/2408.13017

    2024 arXiv

  37. [45]

    Learning latent wireless dynamics from channel state information,

    C. B. Chaaya, A. M. Girgis, and M. Bennis, “Learning latent wireless dynamics from channel state information,” Sep. 2024. [Online]. Available: https://arxiv.org/abs/2409.10045

    2024 arXiv

  38. [46]

    Transfer learning for CSI-based positioning with multi-environment meta-learning,

    A. Foliadis, M. H. Castañeda Garcia, R. A. Stirling-Gallacher, and R. S. Thomä, “Transfer learning for CSI-based positioning with multi-environment meta-learning,” May 2024. [Online]. Available: https://arxiv.org/abs/2405.11816

    2024 arXiv

  39. [47]

    Deep learning-based positioning with multi-task learning and uncertainty-based fusion,

    ——, “Deep learning-based positioning with multi-task learning and uncertainty-based fusion,”IEEE Trans. Mach. Learn. Commun. Netw., vol. 2, pp. 1127–1141, Aug. 2024

    2024

  40. [48]

    Siamese neural networks for wireless positioning and channel charting,

    E. Lei, O. Castañeda, O. Tirkkonen, T. Goldstein, and C. Studer, “Siamese neural networks for wireless positioning and channel charting,” inProc. Allerton Conf. Commun., Contr., Comput., Sep. 2019, pp. 200–207

    2019

  41. [49]

    Triplet- based wireless channel charting: Architecture and experiments,

    P. Ferrand, A. Decurninge, L. G. Ordoñez, and M. Guillaud, “Triplet- based wireless channel charting: Architecture and experiments,”IEEE J. Sel. Areas Commun., vol. 39, no. 8, pp. 2361–2373, Jun. 2021

    2021

  42. [50]

    Unveiling the power of complex-valued transformers in wireless communications,

    Y . Leng, Q. Lin, L.-Y . Yung, J. Lei, Y . Li, and Y .-C. Wu, “Unveiling the power of complex-valued transformers in wireless communications,” Feb. 2025. [Online]. Available: https://arxiv.org/abs/2502.11151

    2025 arXiv

  43. [51]

    Facenet: A unified embedding for face recognition and clustering,

    F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified embedding for face recognition and clustering,” inIEEE Conf. Comput. Vision Pattern Recognit. (CVPR), Jun. 2015, pp. 815–823

    2015

  44. [52]

    On lines and planes of closest fit to systems of points in space,

    K. Pearson, “On lines and planes of closest fit to systems of points in space,”The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, vol. 2, no. 11, pp. 559–572, 1901

    1901

  45. [53]

    Analysis of a complex of statistical variables into principal components,

    H. Hotteling, “Analysis of a complex of statistical variables into principal components,”J. of Edu. Psychology, vol. 24, no. 6, p. 417, 1933

    1933

  46. [54]

    Signature verification using a

    J. Bromley, I. Guyon, Y . LeCun, E. Säckinger, and R. Shah, “Signature verification using a "siamese" time delay neural network,” inProc. Intl. Conf. Neural Info. Proc. Systems (NIPS), Nov. 1993, p. 737–744

    1993

  47. [55]

    A nonlinear mapping for data structure analysis,

    J. Sammon, “A nonlinear mapping for data structure analysis,”IEEE Trans. Comput., vol. C-18, no. 5, pp. 401–409, May 1969

    1969

  48. [56]

    Understanding the difficulty of training deep feedforward neural networks,

    X. Glorot and Y . Bengio, “Understanding the difficulty of training deep feedforward neural networks,” inProc. Thirteenth Int. Conf. Artificial Intelligence Stat., vol. 9, May 2010, pp. 249–256

    2010

  49. [57]

    Learning representations by back-propagating errors,

    D. Rumelhart, G. Hinton, and R. Williams, “Learning representations by back-propagating errors,”Nature, vol. 323, p. 533–536, Oct. 1986

    1986

  50. [58]

    Reducing the dimensionality of data with neural networks,

    G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,”Science, vol. 313, no. 5786, pp. 504–507, Jul. 2006

    2006

  51. [59]

    Dimensionality reduction: A comparative review,

    L. J. P. Van der Maaten, E. O. Postma, and H. J. Van Den Herik, “Dimensionality reduction: A comparative review,”J. of Mach. Learn. Res., vol. 10, pp. 66–71, Oct. 2009

    2009

  52. [60]

    Autoencoders, unsupervised learning and deep architectures,

    P. Baldi, “Autoencoders, unsupervised learning and deep architectures,” inProc. ICML on Unsupervised and Transfer Learn., ser. UTLW’11. JMLR.org, Jul. 2012, p. 37–50

    2012

  53. [2024]

    Available: https://arxiv.org/abs/2410.00617

    [Online]. Available: https://arxiv.org/abs/2410.00617

    arXiv

  54. [2025]

    Available: https://arxiv.org/abs/2504.14100

    [Online]. Available: https://arxiv.org/abs/2504.14100

    arXiv

Pith tools

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