Pith. sign in

REVIEW 4 major objections 4 minor 129 references

SigMap, a wireless localization foundation model, uses 3D maps as prompts and cycle-aware CSI masking to transfer across unseen cities with only a tiny fine-tuned head.

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

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

T0 review · deepseek-v4-flash

2026-08-01 22:30 UTC pith:RAYW5EGA

load-bearing objection Interesting cycle-adaptive masking and map-prompt ideas, but the abstract's zero-shot claim is not supported by the few-shot fine-tuning protocol; needs correction before the numbers can be trusted. the 4 major comments →

arxiv 2607.15713 v1 pith:RAYW5EGA submitted 2026-07-17 eess.SP cs.AIcs.LG

Map as a Prompt: Learning Multi-Modal Spatial-Signal Foundation Models for Cross-scenario Wireless Localization

Yong Chu , Xun Zhou , Zenglin Xu , Hui Wang , Yue Yu This is my paper
classification eess.SP cs.AIcs.LG
keywords wireless localizationchannel state informationfoundation modelself-supervised pretrainingmasked autoencodingprompt tuningcross-scenario generalizationgeometric prompt
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Wireless localization with deep learning usually fails when the environment changes: models trained in one urban scene need extensive labeled data to work in another. SigMap claims to break that cycle by pre-training a transformer on unlabeled channel state information (CSI) with a masking scheme that deliberately disrupts periodic patterns, so the network cannot cheat by interpolating repeated rows. At fine-tuning time, the frozen backbone is guided by a lightweight prompt generated from the 3D building geometry and base station positions via a graph network, so only a 0.085M-parameter head is updated. On two unseen ray-traced city suites, the method reports mean absolute errors near one to two meters using only about 100 labeled target samples, beating supervised and self-supervised baselines by large margins. A sympathetic reader would take away that environment-aware prompting, not more data or bigger models, is the lever for cross-scenario wireless localization.

Core claim

The paper's central discovery is that two obstacles to a transferable wireless localization model—periodicity-driven shortcut learning in masked pre-training and the absence of environmental geometry in downstream adaptation—can be addressed with explicit structural inductive biases. First, by computing cross-correlation between rows of the channel frequency response and generating shift-aware masks, the pre-training objective forces representations to capture global propagation semantics instead of local repetition. Second, by rendering the environment's 3D map and base station layout as a heterogeneous graph and pooling it into a soft prompt token prepended to the frozen transformer's inpu

What carries the argument

The key machinery is a pair of coupled mechanisms. The cycle-adaptive mask is a binary pattern generated from the row-wise cross-correlation of the CSI matrix; it detects the dominant shift period d and removes diagonal bands of width w, so masked reconstruction cannot be solved by copying neighboring rows. The map-as-prompt token is a single D-dimensional vector produced by a two-layer graph convolution over a Delaunay-triangulated heterogeneous graph of building vertices and base station positions, pooled and projected; it is prepended to the frozen transformer's input so that the environment's geometry modulates attention without updating backbone weights. Together they deliver the paper'

Load-bearing premise

The cross-scenario generalization claim rests on the assumption that representations learned on a single ray-traced urban scenario (DeepMIMO O1, with base stations 3/4/9/10) are general enough that fine-tuning only a 0.085M-parameter head on about 100 labeled samples of a new scene yields the reported accuracy; the paper does not ablate pre-training diversity.

What would settle it

Pre-train on the same O1 data but replace cycle-adaptive masking with random masking (keeping the map prompt) and evaluate on O2 and WAIR-D with 100 fine-tuning samples; if the MAE gap to the full method nearly disappears, the masking mechanism is not the source of generalization. Alternatively, pre-train on O1 and fine-tune on a ray-traced city whose building heights and street widths fall far outside O1's distribution; if CDF@1m collapses below 30% or errors exceed a CNN trained from scratch on the same 100 samples, the cross-scenario claim fails.

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

If this is right

  • Fine-tuning only a 0.085M-parameter head on ~100 labeled samples of an unseen city yields MAEs of 1.03 m (DeepMIMO O2) and 1.88 m (WAIR-D Scenario-2), implying a single pre-trained backbone can serve multiple deployment scenarios.
  • Removing the map prompt raises MAE from 1.564 m to 2.275 m (single-BS) and from 0.673 m to 0.789 m (multi-BS), showing the geometric prompt is a direct accuracy source.
  • Replacing the cycle-adaptive mask with fixed grid or strip masks degrades multi-BS MAE from 0.673 m to 0.770/0.753 m, indicating masking strategy meaningfully shapes the learned representation.
  • Swapping the full 3D mesh for a 2D bird's-eye polygon costs only 8% MAE (1.564 m to 1.692 m), suggesting most of the prompt's benefit comes from topological/LoS structure rather than vertical detail.

Where Pith is reading between the lines

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

  • If the transfer results hold, the most consequential implication is that pre-training diversity may matter less than prompt-side geometry: a model trained on one urban layout can adapt to another with almost no labeled data, shifting the field's focus from bigger pre-training corpora to better environment encoders.
  • The 2-D map result hints at a cheaper deployment path: since most of the benefit survives dropping height and facade normals, a satellite or street-level image could serve as a proxy prompt when 3D meshes are unavailable—an extension the paper itself flags for future work.
  • The paper reports only the ~100-sample fine-tuning regime, so the prompt mechanism's sample-efficiency curve remains uncharacterized; a natural test is to vary the target-set size from a handful to a few thousand labeled samples.
  • Because the map prompt is generated per scene rather than per user, the same frozen backbone and prompt mechanism may transfer to other environment-aware wireless tasks—beam prediction, channel estimation, or coverage mapping—by swapping the task head.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes SigMap, a self-supervised wireless-localization foundation model. Pre-training uses a transformer-based masked autoencoder with a proposed "cycle-adaptive" masking strategy that adjusts mask patterns to the periodicity of CSI. For downstream localization, a GNN encodes a Delaunay-triangulated graph of 3D building vertices and base-station positions into a soft prompt token, which is appended to the frozen backbone; only the GNN, projection MLP, and task head are fine-tuned. Experiments on DeepMIMO O1 report strong single-BS and multi-BS localization results, ablations study the masking and map prompt, and cross-scenario results on DeepMIMO O2 and WAIR-D claim strong generalization with minimal fine-tuning. The abstract and conclusion claim state-of-the-art performance and strong zero-shot generalization to unseen environments.

Significance. If the claims were fully supported, SigMap would be a useful contribution: a parameter-efficient, map-conditioned foundation model for wireless localization that transfers across ray-traced environments. The paper has genuine strengths: it ships a reproducibility repository, the cycle-adaptive masking idea is novel and well motivated by CSI structure, and fine-tuning only ~0.085 M parameters while keeping the backbone frozen is a practically meaningful efficiency result. However, the central "zero-shot" claim is not tested by the experiments, several numeric inconsistencies weaken trust in the reported transfer numbers, and the state-of-the-art claim is not backed by comparisons to the foundation-model baselines cited in the introduction. These issues are load-bearing because they concern the paper's headline contribution.

major comments (4)
  1. [§4.5, Abstract, §5] The abstract claims "strong zero-shot generalization in unseen environments," but the only cross-scenario experiment in §4.5 fine-tunes downstream task heads on approximately 100 labeled samples per target scenario and uses the target scenario's 3D map and base-station positions to generate prompts. This is few-shot adaptation with environment oracle information, not zero-shot inference. No experiment evaluates the frozen encoder on O2 or WAIR-D without any target labels or target map. The claims in the abstract, §1.2, and §5 must be reconciled with the actual protocol, either by adding a true zero-shot experiment or by removing/rephrasing the zero-shot claim throughout.
  2. [§4.5, Table 4.5] There is a direct numeric inconsistency in the WAIR-D results. The text states that SigMap reaches 1.580 m MAE on WAIR-D Scenario-2 and outperforms LWLM by 44.3%, while the table reports 1.880 m MAE for SigMap. The improvement over LWLM (3.375 m) is 44.3% for 1.880 m but 53.2% for 1.580 m. Both numbers cannot be correct. Additionally, §4.5 says only 0.4% of parameters are updated while §4.6 and Table 5 say 0.7% (0.085 M of 11.815 M ≈ 0.72%). These discrepancies must be fixed before the transfer results can be considered reliable.
  3. [§4.2, §4.5] The state-of-the-art claim is not adequately supported by the baseline set. The introduction cites LWM, WirelessGPT, CrowdBERT, Signal-guided MAE, and MetaLoc as related foundation-model/meta-learning approaches, but none of these are evaluated. The experiments compare only OMP, CNN, SWiT, and LWLM. For a paper whose contribution is a foundation model, omitting the cited foundation-model baselines is a significant gap. Moreover, no standard deviations or confidence intervals are reported despite §4.1 saying results are averaged over 5 runs, and the cross-scenario table has no from-scratch training baseline under the same 100-sample protocol, so the transfer gain cannot be attributed to pre-training.
  4. [§3.3, Eq. (6), Appendix B.4] The cycle-adaptive masking mechanism depends on the periodicity shift d_final, but the paper does not specify how d_final is estimated from data. Eq. (6) takes d_final as an input, and Appendix B.4 gives illustrative shifts (d = 8, -3, 0) but provides no detection algorithm, no selection criterion on validation data, and no sensitivity analysis for the mask parameters N_a, N_s, and w. If d_final is tuned on labeled target data, the self-supervised advantage is compromised; if it is estimated from unlabeled CSI, the estimator needs to be described precisely. This is load-bearing for the first innovation claim.
minor comments (4)
  1. [Table 7] The Train/Val/Test row reads "10,000/1,000/10,00" — likely a typo for "10,000/1,000/1,000" or "10,000/1,000/10,000". Please correct.
  2. [§2.1] The phrase "each path in equation equation 1" contains a duplicated word. Also, Eq. (11) appears in §4.2 after Table 1 without being referenced in the surrounding text; it should be introduced or removed.
  3. [Appendix B.2 / §3.2] The notation for the input tensor is inconsistent: Eq. (4) defines H ∈ C^{Nr×Nt×Ns}, Eq. (5) uses X = [Re(H), Im(H)], but Appendix B.2 uses H_s = [|Hs|, angle(Hs)] with dimensions (2, N_ant, N_subc). Please unify the notation and clarify which representation is actually used.
  4. [§4.5] The table under §4.5 is referred to as "Table 4.5" in the text but is not numbered. Please number it and reference it consistently.

Circularity Check

1 steps flagged

Advertised zero-shot transfer is evaluated by fine-tuning task heads on target labels; the cross-scenario 'prediction' is therefore few-shot by construction.

specific steps
  1. fitted input called prediction [Abstract & Section 1.2 vs Section 4.5 (Generalization to New Environments)]
    "Abstract: 'exhibiting strong zero-shot generalization in unseen environments.' Contribution 3: 'demonstrates strong zero-shot generalization to unseen environments and base station configurations.' Section 4.5: 'In all experiments, only the downstream task heads are fine-tuned using limited target samples (approximately 100 instances per scenario), while the self-supervised backbone remains frozen.'"

    The advertised 'zero-shot generalization' is evaluated by fine-tuning the task heads on roughly 100 labeled samples from each target scenario (DeepMIMO O2 and WAIR-D). The reported cross-scenario MAEs in Table 4.5 are therefore produced with target labels in the loop, not by a frozen encoder alone. No experiment omits target data, so the 'zero-shot' result reduces by construction to few-shot adaptation: the claimed prediction is partly fitted to the target scenario. This is load-bearing because the paper's central contribution is cross-scenario transfer without per-scenario labels.

full rationale

The base localization results (Tables 1 and 2) are not circular: they compare SigMap against external baselines on the same benchmark, using standard supervised evaluation, and the masking/map ablations (Tables 3 and 4) are internal comparisons of design choices, not predictions masquerading as fits. There is also no load-bearing self-citation chain: the cited prior methods are external, and no 'uniqueness theorem' is imported from the authors' own work. The significant circularity issue is confined to the headline zero-shot claim. The abstract and contribution list promise 'strong zero-shot generalization to unseen environments,' but Section 4.5 explicitly fine-tunes the task heads on about 100 target instances per unseen scenario and reports only those fine-tuned numbers. Because target labels and target-region map prompts are used, the evidence does not demonstrate zero-shot transfer; it demonstrates few-shot adaptation. That is a real reduction of the advertised prediction by construction, and since it is the paper's central selling point, it warrants a score of 6. The additional text/table discrepancy in Section 4.5 (1.580 m vs. 1.880 m MAE for WAIR-D) further weakens confidence in the reported transfer numbers, though it is a correctness issue rather than a separate circular step.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claims rest primarily on domain assumptions about simulation realism, transferability of a single-scenario pretraining, and the informativeness of GNN map prompts. No new physical entities are introduced. The main unaccounted free parameter is the periodicity shift in the masking mechanism.

free parameters (2)
  • d_final / d (periodicity shift in cycle-adaptive mask) = unspecified; examples d=8, -3, 0 in App. B.4
    Eq. 6 defines the mask via a 'detected periodicity shift' but no detection procedure is given; Appendix B.4 treats d as a slope parameter with random start columns, so the core masking pattern is effectively a hand-chosen/randomized parameter rather than a derived quantity.
  • Mask counts and width (N_a, N_s, w) = N_a=8, N_s=32, w unspecified
    These control the masking geometry and are chosen without sensitivity analysis; the central pre-training objective depends on them.
axioms (4)
  • domain assumption Ray-traced CSI from DeepMIMO/WAIR-D faithfully represents real-world wireless channels for localization
    All training and evaluation use simulations (§4.1, §4.5); if sim-to-real gap is large, the absolute accuracy claims do not carry to deployed systems.
  • domain assumption Row-wise cross-correlation between CFR rows can reliably detect CSI periodicity with boundary effects controlled
    App. B.4 asserts this without giving the detection rule for d_final; the cycle-adaptive masking claim depends on this.
  • domain assumption Delaunay triangulation over building vertices and base-station positions captures the spatial-topological cues relevant to localization
    §3.4 constructs the prompt graph this way; the map-prompt contribution rests on this surrogate for 3D geometry.
  • domain assumption A backbone pre-trained only on DeepMIMO O1 3p5 transfers to other ray-traced scenarios after ~100-sample head fine-tuning
    §4.5 tests exactly this; no pre-training data diversity ablations are provided, so transferability is assumed rather than established.

pith-pipeline@v1.3.0-alltime-deepseek · 11091 in / 15066 out tokens · 116170 ms · 2026-08-01T22:30:52.034959+00:00 · methodology

0 comments
read the original abstract

Accurate and robust wireless localization is a critical enabler for emerging 5G/6G applications, including autonomous driving, extended reality, and smart manufacturing. Despite its importance, achieving precise localization across diverse environments remains challenging due to the complex nature of wireless signals and their sensitivity to environmental changes. Existing data-driven approaches often suffer from limited generalization capability, requiring extensive labeled data and struggling to adapt to new scenarios. To address these limitations, we propose SigMap, a multimodal foundation model that introduces two key innovations: (1) A cycle-adaptive masking strategy that dynamically adjusts masking patterns based on channel periodicity characteristics to learn robust wireless representations; (2) A novel "map-as-prompt" framework that integrates 3D geographic information through lightweight soft prompts for effective cross-scenario adaptation. Extensive experiments demonstrate that our model achieves state-of-the-art performance across multiple localization tasks while exhibiting strong zero-shot generalization in unseen environments, significantly outperforming both supervised and self-supervised baselines by considerable margins.

Figures

Figures reproduced from arXiv: 2607.15713 by Hui Wang, Xun Zhou, Yong Chu, Yue Yu, Zenglin Xu.

Figure 1
Figure 1. Figure 1: Wireless propagation paths in urban environments. LoS represents direct propagation, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of our wireless localization foundation model, showing the two-stage [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of our cycle-adaptive masking strategy. The mask pattern (right) is dynamically [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The pipeline of geographic prompt generation. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comprehensive performance comparison across metrics. Our method shows consistent [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Two typical complex scenes of WAIR-D. Generalization performance on unseen scenarios with minimal fine-tuning. Method MAE (m) RMSE (m) CDF@1m (%) DeepMIMO O2 outdoor SIGMAP (w/ map) 1.026 1.551 66.4 SIGMAP (w/o map) 1.282 5.824 63.9 LWLM 2.213 11.837 63.2 WAIR-D Scenario-2 (100 cities) SIGMAP (w/ map) 1.880 3.717 58.0 SIGMAP (w/o map) 2.578 4.650 51.5 LWLM 3.375 6.921 50.3 Equipped with geographic prompts,… view at source ↗
Figure 7
Figure 7. Figure 7: CSI Amplitude Heatmaps for Different Shift Patterns [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Single-Base Station Localization Performance Comparison [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Multi-Base Station Localization Performance Comparison [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

129 extracted references · 5 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2212.02159 , year=

    Wair-d: Wireless ai research dataset , author=. arXiv preprint arXiv:2212.02159 , year=

    Pith/arXiv arXiv

  2. [2]

    IEEE Journal on Selected Areas in Communications , volume=

    MetaLoc: Learning to learn wireless localization , author=. IEEE Journal on Selected Areas in Communications , volume=. 2023 , publisher=

    2023

  3. [3]

    IEEE Sensors Journal , volume=

    Robust TDOA-based indoor localization using improved clock-sync-scheme and multilevel constrained ARPF , author=. IEEE Sensors Journal , volume=. 2023 , publisher=

    2023

  4. [4]

    2019 28th Wireless and Optical Communications Conference (WOCC) , pages=

    A Cross-region Wireless-synchronization-based TDOA Method for Indoor Positioning Applications , author=. 2019 28th Wireless and Optical Communications Conference (WOCC) , pages=. 2019 , organization=

    2019

  5. [5]

    IEEE Transactions on Vehicular Technology , year=

    Signal-guided masked autoencoder for wireless positioning with limited labeled samples , author=. IEEE Transactions on Vehicular Technology , year=

  6. [6]

    2024 14th International Conference on Indoor Positioning and Indoor Navigation (IPIN) , pages=

    Radio foundation models: Pre-training transformers for 5G-based indoor localization , author=. 2024 14th International Conference on Indoor Positioning and Indoor Navigation (IPIN) , pages=. 2024 , organization=

    2024

  7. [7]

    arXiv preprint arXiv:2505.10134 , year=

    Large Wireless Localization Model (LWLM): A Foundation Model for Positioning in 6G Networks , author=. arXiv preprint arXiv:2505.10134 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2405.17053 , year=

    Wirelessllm: Empowering large language models towards wireless intelligence , author=. arXiv preprint arXiv:2405.17053 , year=

    Pith/arXiv arXiv

  9. [9]

    2024 IEEE 6th Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC) , volume=

    Map-assisted Wireless TDOA Localization Enhancement Based On CNN , author=. 2024 IEEE 6th Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC) , volume=. 2024 , organization=

    2024

  10. [10]

    IEEE Transactions on Vehicular Technology , year=

    CSI-GPT: Integrating generative pre-trained transformer with federated-tuning to acquire downlink massive MIMO channels , author=. IEEE Transactions on Vehicular Technology , year=

  11. [11]

    arXiv preprint arXiv:1902.06435 , year=

    DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications , author=. arXiv preprint arXiv:1902.06435 , year=

    Pith/arXiv arXiv 1902

  12. [12]

    2013 , publisher=

    Auto-encoding variational bayes , author=. 2013 , publisher=

    2013

  13. [13]

    , title =

    Alkhateeb, A. , title =. Proc. ITA'2019 , year =

    2019

  14. [14]

    2023 , publisher=

    Decarli, Nicol. 2023 , publisher=

    2023

  15. [15]

    2023 , publisher=

    Integrated localization and communication for efficient millimeter wave networks , author=. 2023 , publisher=

    2023

  16. [16]

    IEEE Trans

    TransLoc: A heterogeneous knowledge transfer framework for fingerprint-based indoor localization , author=. IEEE Trans. Wireless Commun. , volume=. 2021 , publisher=

    2021

  17. [17]

    IEEE Commun

    On the ground and in the sky: A tutorial on radio localization in ground-air-space networks , author=. IEEE Commun. Surveys Tuts. , year=

  18. [18]

    Location-Dependent Performance Analysis for

    Chen, Changshan and Jiang, Hao and Pan, Cunhua , journal=. Location-Dependent Performance Analysis for. 2024 , publisher=

    2024

  19. [19]

    Location based joint spectrum sensing and radio resource allocation in cognitive radio enabled

    Yang, Qi and Huang, Yi-Feng and Yen, Yin-Chen and Chen, Ling-Yu and Chen, Hsiao-Hwa and Hong, Xue-Min and Shi, Jiang-Hong and Wang, Liangmin , journal=. Location based joint spectrum sensing and radio resource allocation in cognitive radio enabled. 2020 , publisher=

    2020

  20. [20]

    Orientation and location tracking of

    Talvitie, Jukka and S. Orientation and location tracking of. 2023 , publisher=

    2023

  21. [21]

    2017 , publisher=

    Wymeersch, Henk and Seco-Granados, Gonzalo and Destino, Giuseppe and Dardari, Davide and Tufvesson, Fredrik , journal=. 2017 , publisher=

    2017

  22. [22]

    arXiv preprint arXiv:1807.03748 , year=

    Representation learning with contrastive predictive coding , author=. arXiv preprint arXiv:1807.03748 , year=

    Pith/arXiv arXiv

  23. [23]

    2024 , publisher=

    A survey on information bottleneck , author=. 2024 , publisher=

    2024

  24. [24]

    and Svensson, Tommy and Schindhelm, Kim and Wolfgang, Andreas and Lindberg, Simon and Wymeersch, Henk , journal=

    Behravan, Ali and Yajnanarayana, Vijaya and Keskin, Musa Furkan and Chen, Hui and Shrestha, Deep and Abrudan, Traian E. and Svensson, Tommy and Schindhelm, Kim and Wolfgang, Andreas and Lindberg, Simon and Wymeersch, Henk , journal=. Positioning and Sensing in. 2023 , volume=

    2023

  25. [25]

    2024 , volume=

    Chen, Hui and Keskin, Musa Furkan and Sakhnini, Adham and Decarli, Nicolò and Pollin, Sofie and Dardari, Davide and Wymeersch, Henk , journal=. 2024 , volume=

    2024

  26. [26]

    and Moghaddam, Mohammad H

    Wymeersch, Henk and Chen, Hui and Guo, Hao and Keskin, Musa Furkan and Khorsandi, Bahare M. and Moghaddam, Mohammad H. and Ramirez, Alejandro and Schindhelm, Kim and Stavridis, Athanasios and Svensson, Tommy and Yajnanarayana, Vijaya , journal=. 2025 , volume=

    2025

  27. [27]

    Pan, Guangjin and Gao, Yuan and Gao, Yilin and Zhong, Zhiyong and Yang, Xiaoyu and Guo, Xinyu and Xu, Shugong , journal=

  28. [28]

    Bommasani, Rishi and Hudson, Drew A and Adeli, Ehsan and Altman, Russ and Arora, Simran and von Arx, Sydney and Bernstein, Michael S and Bohg, Jeannette and Bosselut, Antoine and Brunskill, Emma and others , journal=

  29. [29]

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , booktitle=

  30. [30]

    Language models are few-shot learners , author=. Proc. NIPS'2020 , volume=

    2020

  31. [31]

    Attention is all you need , author=. Proc. NIPS'2017 , volume=

    2017

  32. [32]

    Hong, Danfeng and Zhang, Bing and Li, Xuyang and Li, Yuxuan and Li, Chenyu and Yao, Jing and Yokoya, Naoto and Li, Hao and Ghamisi, Pedram and Jia, Xiuping and Plaza, Antonio and Gamba, Paolo and Benediktsson, Jon Atli and Chanussot, Jocelyn , journal=

  33. [33]

    Foundation Model-Based Multimodal Remote Sensing Data Classification , year=

    He, Xin and Chen, Yushi and Huang, Lingbo and Hong, Danfeng and Du, Qian , journal=. Foundation Model-Based Multimodal Remote Sensing Data Classification , year=

  34. [34]

    Nature Communications , year=

    Towards Artificial General Intelligence via a Multimodal Foundation Model , author=. Nature Communications , year=

  35. [35]

    Deep learning-based multi-user positioning in wireless

    Chen, Zirui and Zhang, Zhaoyang and Xiao, Zhuoran and Yang, Zhaohui and Jin, Richeng , journal=. Deep learning-based multi-user positioning in wireless. 2023 , publisher=

    2023

  36. [36]

    Klus, Roman and Talvitie, Jukka and Equi, Julia and Fodor, Gabor and Torsner, Johan and Valkama, Mikko , journal=. Robust. 2024 , publisher=

    2024

  37. [37]

    A tutorial on terahertz-band localization for

    Chen, Hui and Sarieddeen, Hadi and Ballal, Tarig and Wymeersch, Henk and Alouini, Mohamed-Slim and Al-Naffouri, Tareq Y , journal=. A tutorial on terahertz-band localization for. 2022 , publisher=

    2022

  38. [38]

    A Tutorial on

    Italiano, Lorenzo and Tedeschini, Bernardo Camajori and Brambilla, Mattia and Huang, Huiping and Nicoli, Monica and Wymeersch, Henk , journal=. A Tutorial on. 2024 , volume=

    2024

  39. [39]

    Learning to localize: A

    Wu, Chi and Yi, Xinping and Wang, Wenjin and You, Li and Huang, Qing and Gao, Xiqi and Liu, Qing , journal=. Learning to localize: A. 2021 , publisher=

    2021

  40. [40]

    2021 , publisher=

    Keskin, Musa Furkan and Wymeersch, Henk and Koivunen, Visa , journal=. 2021 , publisher=

    2021

  41. [41]

    and Crouse, Michael B

    Comiter, Marcus Z. and Crouse, Michael B. and Kung, H. T. , booktitle=. A Data-Driven Approach to Localization for High Frequency Wireless Mobile Networks , year=

  42. [42]

    High Accurate Time-of-Arrival Estimation With Fine-Grained Feature Generation for Internet-of-Things Applications , year=

    Pan, Guangjin and Wang, Tao and Zhang, Shunqing and Xu, Shugong , journal=. High Accurate Time-of-Arrival Estimation With Fine-Grained Feature Generation for Internet-of-Things Applications , year=

  43. [43]

    A transformer-based signal denoising network for

    Liu, Junchen and Wang, Tianyu and Li, Yuxiao and Li, Cheng and Wang, Yi and Shen, Yuan , journal=. A transformer-based signal denoising network for. 2022 , publisher=

    2022

  44. [44]

    2023 , publisher=

    Li, Wei and Meng, Xiangxu and Zhao, Zheng and Liu, Zhihan and Chen, Chuhao and Wang, Huiqiang , journal=. 2023 , publisher=

    2023

  45. [45]

    Deep learning based fingerprint positioning for multi-cell massive

    Gong, Xinrui and Lu, An'an and Liu, Xiaofeng and Fu, Xiao and Gao, Xiqi and Xia, Xiang-Gen , journal=. Deep learning based fingerprint positioning for multi-cell massive. 2023 , publisher=

    2023

  46. [46]

    Swin-Loc: Transformer-based

    Xu, Xiaodong and Zhu, Fangzhou and Han, Shujun and Yu, Zhongyao and Zhao, Hangyu and Wang, Bizhu and Zhang, Ping , journal=. Swin-Loc: Transformer-based. 2024 , publisher=

    2024

  47. [47]

    Positioning and Sensing in

    Behravan, Ali and Yajnanarayana, Vijaya and Keskin, Musa Furkan and Chen, Hui and Shrestha, Deep and Abrudan, Traian E and Svensson, Tommy and Schindhelm, Kim and Wolfgang, Andreas and Lindberg, Simon and others , journal=. Positioning and Sensing in

  48. [48]

    Wymeersch, Henk and P. Proc. EuCNC (Eur. Conf. Netw. Commun.) & 6G Summit , year=

  49. [49]

    A simple framework for contrastive learning of visual representations , author=. Proc. ICML‘2020 , pages=. 2020 , organization=

    2020

  50. [50]

    IEEE Transactions on Communications , year=

    Angle-delay profile-based and timestamp-aided dissimilarity metrics for channel charting , author=. IEEE Transactions on Communications , year=

  51. [51]

    Tokens-to-token vit: Training vision transformers from scratch on imagenet , author=. Proc. IEEE/CVF ICCV'2021 , pages=

    2021

  52. [52]

    2023 , volume=

    Ruan, Yanlin and Chen, Liang and Zhou, Xin and Liu, Zhaoliang and Liu, Xiaoyan and Guo, Guangyi and Chen, Ruizhi , journal=. 2023 , volume=

    2023

  53. [53]

    2022 , volume=

    Chen, Xi and Li, Hang and Zhou, Chenyi and Liu, Xue and Wu, Di and Dudek, Gregory , journal=. 2022 , volume=

    2022

  54. [54]

    Enhancing Indoor Localization With Semi-Crowdsourced Fingerprinting and

    Junoh, Suhardi Azliy and Pyun, Jae-Young , journal=. Enhancing Indoor Localization With Semi-Crowdsourced Fingerprinting and. 2024 , volume=

    2024

  55. [55]

    2024 , publisher=

    A survey on self-supervised learning: Algorithms, applications, and future trends , author=. 2024 , publisher=

    2024

  56. [56]

    2024 , volume=

    Han, Yu and Li, Zan and Zhao, Zhongliang and Braun, Torsten , journal=. 2024 , volume=

    2024

  57. [57]

    Self-Supervised and Invariant Representations for Wireless Localization , year=

    Salihu, Artan and Rupp, Markus and Schwarz, Stefan , journal=. Self-Supervised and Invariant Representations for Wireless Localization , year=

  58. [58]

    Signal-Guided Masked Autoencoder for Wireless Positioning With Limited Labeled Samples , year=

    Wang, Ji and Fang, Wei and Xiao, Jian and Zheng, Yi and Zheng, Le and Liu, Fan , journal=. Signal-Guided Masked Autoencoder for Wireless Positioning With Limited Labeled Samples , year=

  59. [59]

    Radio Foundation Models: Pre-training Transformers for

    Ott, Jonathan and Pirkl, Jonas and Stahlke, Maximilian and Feigl, Tobias and Mutschler, Christopher , booktitle=. Radio Foundation Models: Pre-training Transformers for. 2024 , organization=

    2024

  60. [60]

    Yang, Tingting and Zhang, Ping and Zheng, Mengfan and Shi, Yuxuan and Jing, Liwen and Huang, Jianbo and Li, Nan , journal=

  61. [61]

    Liu, Boxun and Gao, Shijian and Liu, Xuanyu and Cheng, Xiang and Yang, Liuqing , journal=

  62. [62]

    Large wireless model

    Alikhani, Sadjad and Charan, Gouranga and Alkhateeb, Ahmed , journal=. Large wireless model

  63. [63]

    Liu, Xuanyu and Gao, Shijian and Liu, Boxun and Cheng, Xiang and Yang, Liuqing , journal=

  64. [64]

    2024 , publisher=

    Liu, Boxun and Liu, Xuanyu and Gao, Shijian and Cheng, Xiang and Yang, Liuqing , journal=. 2024 , publisher=

    2024

  65. [65]

    Wireless communications , author=. Aug. 2005 , publisher=

    2005

  66. [66]

    2006 , publisher=

    Numerical optimization , author=. 2006 , publisher=

    2006

  67. [67]

    2022 , publisher=

    Yassine, Taha and Le Magoarou, Luc , journal=. 2022 , publisher=

    2022

  68. [68]

    Coverage Analysis of Joint Localization and Communication in

    Zheng, Pinjun and Ballal, Tarig and Chen, Hui and Wymeersch, Henk and Al-Naffouri, Tareq Y , year=. Coverage Analysis of Joint Localization and Communication in

  69. [69]

    Ozturk, Cuneyd and Keskin, Musa Furkan and Wymeersch, Henk and Gezici, Sinan , journal=

  70. [70]

    Mohammadian, Amirhossein and Tellambura, Chintha , journal=. Aug. 2021 , publisher=

    2021

  71. [71]

    Chen, Hui and Aghdam, Sina Rezaei and Keskin, Musa Furkan and Wu, Yibo and Lindberg, Simon and Wolfgang, Andreas and Gustavsson, Ulf and Eriksson, Thomas and Wymeersch, Henk , booktitle=

  72. [72]

    IEEE Trans

    Performance limits and geometric properties of array localization , author=. IEEE Trans. Inf. Theory , volume=. Dec. 2015 , publisher=

    2015

  73. [73]

    Tensor decomposition based beamspace

    Wen, Fuxi and Garcia, Nil and Kulmer, Josef and Witrisal, Klaus and Wymeersch, Henk , booktitle=. Tensor decomposition based beamspace. Dec. 2018 , organization=

    2018

  74. [74]

    The Journal of Machine Learning Research , volume=

    Manopt, a Matlab toolbox for optimization on manifolds , author=. The Journal of Machine Learning Research , volume=. Jan. 2014 , publisher=

    2014

  75. [75]

    Beamspace Multidimensional

    Jiang, Fan and Wen, Fuxi and Ge, Yu and Zhu, Meifang and Wymeersch, Henk and Tufvesson, Fredrik , journal=. Beamspace Multidimensional

  76. [76]

    Nazari, Mohammad A and Seco-Granados, Gonzalo and Johannisson, Pontus and Wymeersch, Henk , booktitle=. Jun. 2021 , organization=

    2021

  77. [77]

    Localization error bounds for

    Ghaseminajm, Fariba and Abu-Shaban, Zohair and Ikki, Salama S and Wymeersch, Henk and Benson, Craig R , journal=. Localization error bounds for. Apr. 2020 , publisher=

    2020

  78. [78]

    End-to-end learning for integrated sensing and communication , author=. Proc. IEEE Int. Conf. Commun. (ICC) , pages=. May. 2022 , organization=

    2022

  79. [79]

    No radio left behind: Radio fingerprinting through deep learning of physical-layer hardware impairments , author=. Oct. 2019 , publisher=

    2019

  80. [80]

    Joint compensation of transmitter and receiver impairments in

    Tarighat, Alireza and Sayed, Ali H , journal=. Joint compensation of transmitter and receiver impairments in. Feb. 2007 , publisher=

    2007

Showing first 80 references.