Pith. sign in

REVIEW 4 major objections 6 minor 31 references

DeepSync: A Learning Framework for Pervasive Localization using Code Synchronization on Compressed Cellular Spectrum

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read DeepSync reports 2.128-meter median localization from sub-sample sync-code timing learned on folded LTE spectrum, at roughly one-fiftieth the energy of PRS-based positioning.

desk verdict A genuinely new approach to sub-sample timing in folded cellular spectrum, but the headline 'real-world' accuracy is a simulation of the receiver model, so the paper overstates what is actually demonstrated. read the letter →

arxiv 2505.05783 v1 pith:PTAWTUXL submitted 2025-05-09 eess.SY cs.SY

Aritrik Ghosh , Nakul Garg , Nirupam Roy This is my paper · ORCID
classification eess.SYcs.SY
keywords localizationcompressedspectrumsensingLTEsynchronizationsignalssub-sampletimingestimationcross-attentiontemporalCNNTDOAultra-low-powerreceiver
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

DeepSync is a deep-learning timing-estimation system that lets a low-power receiver localize from ordinary cellular broadcasts without fully decoding the LTE signal. It treats the periodic synchronization code as a template buried in the non-linearly folded, intermodulated spectrum, and learns to align that template to a fraction of a sample. The paper reports a 2.128-meter median localization error on real-world LTE traces (with folding applied in software), a 4-meter median error in ray-traced simulation, and about a 10x accuracy gain over the RSSI-based compressed-spectrum baseline, at 3.884 mJ per inference. If correct, continuous meter-scale tracking of tags, pets, and assets becomes feasible on battery budgets where GPS drains too quickly.

What carries the argument

The load-bearing object is a sub-sample offset regression model over 286-sample windows: the input is the folded-spectrum segment around the coarse correlation peak, and the target is the 276-sample concatenated PSS+SSS template. Two 1D CNN encoders with progressively smaller kernels (14, 7, 5) map the signal and the template to feature maps, per-frame signal embeddings are averaged across ten frames to raise SNR, and multi-head cross-attention treats the received signal as the query and the sync-code template as the key-value pair. A three-layer regression head then produces the predicted fractional sample offset, with training data generated by a physics-based ray-tracing digital twin built from real cell-tower records and urban geometry.

What would settle it

Build a tag with an actual diode envelope-detector receiver that physically folds three live LTE bands, run the same 286-sample extraction and trained DeepSync model on its output along the same route, and compare median localization error to 2.128 meters; the claim stands if the error stays in the meter range and fails if it degrades toward the 20-meter RSSI baseline or worse.

Watch

Extended reading notes

Core claim

On the paper's own terms, DeepSync establishes that sub-sample timing offsets can be learned from the severely degraded, non-linearly folded LTE spectrum, and that this timing precision transfers to TDOA localization. The system estimates the fractional sample offset of the PSS/SSS synchronization code in a 286-sample window: temporal CNN encoders process ten consecutive frames, the frame embeddings are averaged to exploit the 10 ms LTE periodicity, and a cross-attention layer aligns the received folded spectrum with the target sync-code template before a regression head outputs the offset. With these offsets from multiple base stations, hyperbolic TDOA positioning yields a 2.128-meter median error on real-world USRP traces and 4 meters in simulation, improving on the 20-meter RSSI compressed-spectrum baseline by a factor of about ten.

Load-bearing premise

The results assume that the software-applied squaring and spectrum-folding model used for training and for processing the USRP traces reproduces what a real envelope-detector receiver would output, even though the paper's Section 9 lists envelope-detector transfer functions, filter and amplifier responses, antenna frequency selectivity, and ADC clock drift as unhandled hardware effects.

Editorial extensions

If this is right

  • A low-power device can estimate TDOA to sub-sample precision from intermodulated LTE spectrum, so continuous asset, pet, or person tracking no longer requires GPS-class power budgets.
  • Because the method needs no infrastructure modifications and works from ambient base stations, it could be deployed anywhere with LTE coverage.
  • Temporal aggregation across ten frames exploits the 10 ms sync-code periodicity and cuts the median offset error by about 15x, pushing reliable operation below -10 dB SNR where conventional correlation-based timing fails.
  • Cross-attention alignment outperforms concatenation, element-wise addition, and correlation for combining signal and sync-code features, giving a nominal median offset error of 0.239 samples.
  • Compared with PRS-based 5G positioning, DeepSync trades some accuracy (2.128 m versus sub-3 m) for roughly 50x lower energy per inference and much lower latency.

Reading between the lines

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

  • The paper leaves implicit that the same template-matching formulation should transfer to any periodic synchronization or reference signal, such as 5G NR SSB or PRS, where accuracy could improve further.
  • Because the real-world validation applies the nonlinear fold in software to USRP captures rather than through a physical envelope-detector receiver, the decisive test of the paradigm is a hardware implementation; until then the 2.128 m figure is a projection onto ideal folding.
  • A working version would let any LTE-covered device be tracked continuously without transmitting or modifying infrastructure, which cuts both ways for privacy and surveillance.
  • The ray-tracing training pipeline could plausibly produce pre-trained models for new cities or frequency bands, but deployment would require checking how tower geometry and material differences shift the learned offsets.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. DeepSync proposes a deep learning framework for sub-sample timing-offset estimation from nonlinearly folded (compressed) cellular spectrum, targeting TDOA-based localization using LTE synchronization signals. The architecture combines temporal CNN encoders for multi-frame feature extraction with a cross-attention mechanism that aligns the received signal representation with a target sync-code template, followed by regression to predict sub-sample offsets. Training data is generated by a ray-tracing digital twin that integrates open-source cell tower data with urban geometry. The paper reports a median localization accuracy of 2.128 m in so-called real-world tests, a 10x accuracy improvement over a prior compressed-spectrum RSSI system, and a 50x power reduction relative to conventional positioning, with evaluation consisting of simulated urban scenarios and a USRP-based trace experiment in which the nonlinear folding is applied theoretically in software.

Significance. If the reported results were supported by a genuine end-to-end hardware evaluation, DeepSync would represent a meaningful advance toward ultra-low-power, meter-level localization using ambient cellular signals. The core idea—learning sub-sample offsets from folded spectrum via template matching with cross-attention and temporal aggregation—is plausible and worth exploring, and the use of a ray-traced digital twin for training data is a sensible direction. The paper also includes several useful design insights, such as the benefit of multi-frame temporal averaging in low-SNR folded spectra. However, the current evaluation does not substantiate the headline claims: the 'real-world' experiment applies the same theoretical folding model used in training, so it does not independently validate the physical receiver model, and the paper's own limitation section concedes that the results are based on simulated datasets. Given that the central contribution is a practical low-power localization system, the absence of a hardware demonstration with a real envelope-detector receiver is a critical gap.

major comments (4)
  1. [§8.3, §9] The 'real-world' evaluation is not an end-to-end hardware validation, and this is a load-bearing issue. Section 8.3 states that after collecting USRP traces, 'the non-linear transformation was theoretically applied to the frames and they were added to simulate the folded spectrum effect,' and Section 9 explicitly states that 'The results presented in this paper are based on simulated datasets' and lists unmodeled effects including envelope-detector transfer functions, low-pass filter and amplifier responses, frequency-dependent antenna loss, and clock drift. Because the same theoretical squaring/folding model is used both to generate training data and to process the USRP traces, the test is circular: it validates the model against itself rather than against a physical receiver. The manual correction of clock drift further removes a real impairment from the end-to-end path. Consequently, the claimed 2.128 m median accuracy and the 10x improvement over Garg et al. are not demonstrated for the actual low-power receiver architecture that the paper proposes.
  2. [Abstract, §1, §8.1, Table 1, §8.5] The reported quantitative values for the same claimed result are internally inconsistent. The median localization accuracy is given as 2.128 m in the abstract and Table 1, 2.182 m in the introduction, and 4 m in the simulated evaluation in Section 8.1; the per-inference energy consumption is 3.884 mJ in Table 1 and 3.825 mJ in Section 8.5. The abstract and introduction also claim '50x less power consumption compared to traditional positioning systems,' but no power measurement or principled comparison is provided in the evaluation. These inconsistencies must be reconciled, and the paper should report error bars or confidence intervals for all central metrics rather than single point values.
  3. [§8.2] The reported median sub-sample offset error of 0.239 samples is conditional on training with a maximum coarse offset error of 10 samples, yet Section 8.2 shows that with a bound of 40 samples the median error rises to about 0.8 samples. The paper does not justify that the coarse synchronization step in a real folded-spectrum receiver will indeed keep coarse errors within 10 samples, nor does it report the actual distribution of coarse offset errors observed in the USRP traces. Without this grounding, the sub-sample accuracy and the resulting meter-level localization claims are not transferable to a practical system.
  4. [§7, §9] The data generation pipeline in Section 7 includes 'diode response-based envelope correction' and transmitter nonlinearity, but it does not model the frequency-selective transfer function of a physical envelope detector, the low-pass filter and amplifier responses, or the antenna's frequency-dependent loss—effects explicitly deferred to future work in Section 9. As a result, both the simulated and the semi-simulated evaluations are likely to be optimistic relative to a real low-power receiver, and the paper provides no quantitative estimate of the performance gap introduced by these unmodeled hardware effects. The central claim of meter-level accuracy in practice therefore rests on an unvalidated model assumption.
minor comments (6)
  1. [§8.1] The sentence 'DeepSync significantly improves upon this 2, achieving a median accuracy of 4 meters' is garbled, and the surrounding figure references ('as shown in 3.') are incorrect; please rewrite and renumber the figures.
  2. [§4.3, §7, §8.5] Several citations are incomplete or malformed: '(gar, 2023)' in Section 4.3 is not in the reference list, '(Turbo)' and '(CellMapper, 2024)' in Section 7 lack full bibliographic entries, and '(r1-, 2019)' in Section 8.5 should refer to 3GPP TR 38.855 in proper form.
  3. [Table 1] The 'Data Type' column labels DeepSync as 'Real world,' but Section 9 states that the results are based on simulated datasets; this label is misleading and should be changed to 'Simulated' or 'Semi-simulated' to reflect the actual methodology.
  4. [§5.2] The segment extraction rule 'i − 5 ≤ n ≤ i + 280' is described as being 'centered around i,' but the window is asymmetric (5 samples before and 280 samples after); please clarify the intended center or the extraction rule.
  5. [§6.2] Equation (2) states that the encoder output is z_i ∈ R^{158×128}, which is not reconciled with the earlier statement that a Global Average Pooling layer 'compresses the learned representations into a fixed-size embedding'; please clarify the role and output dimension of the pooling layer.
  6. [§7] The ray-tracing simulation parameters (number of reflections, material properties, route geometry, multipath settings) are not reported, which hampers reproducibility; please provide these details or a reference to the exact simulation configuration.

Circularity Check

1 steps flagged · score 3.0 of 10

Real-world accuracy is evaluated using the same theoretical folding model used for training; the 2.128 m figure is a simulated-folding result and the 10x baseline is a self-citation, but the core regression is not equation-level circular.

  1. other [Section 8.3 (Real World Localization); confirmed in Section 9]
    "LTE traces were collected using a USRP N210 (Research) along a predefined route, as shown in Figure 6. Three Base-stations on the route were operating at 1.932 GHz , 2.115 GHz and 2.145 GHz; after that, the non-linear transformation was theoretically applied to the frames and they were added to simulate the folded spectrum effect."

    The training data were generated by applying the theoretical nonlinear (squaring/folding) spectrum model to simulated LTE frames, and Section 8.3 processes real USRP-captured LTE frames with the same theoretical transform. The claimed 2.128 m median 'real-world' accuracy therefore tests the network under the exact folding model used to create its training data; it cannot independently validate the envelope-detector folding model whose fidelity is the open hardware question. Section 9 states 'The results presented in this paper are based on simulated datasets' and lists unmodeled envelope-detector transfer functions, filter/amplifier responses, antenna loss, and clock drift.

full rationale

The core supervised-learning derivation is self-contained: offsets are known labels in synthetic data, the CNN/cross-attention regressor is trained with a Huber loss, and the synthetic evaluation measures generalization to held-out folded frames. No fitted parameter is renamed as a prediction and no result is mathematically forced by the architecture. The main flaw is the boundary of the 'real-world' claim: Section 8.3 folds USRP traces with the same theoretical nonlinear transform used in training, so the 2.128 m result validates generalization to real carrier data under the assumed folding model, not an end-to-end low-power receiver. Section 9 explicitly disclaims hardware realism, which appropriately mitigates the over-claim. The 20-30 dB SNR-degradation premise and the RSSI-baseline comparison come from the authors' own prior Litefoot paper (Garg et al., 2024); these are self-citations and give context, but the DeepSync accuracy numbers are produced by this paper's own pipeline, so the self-citation is not load-bearing for the network's derivation. Overall, there is no equation-level circularity, but the evaluation loop around the folding model and the self-cited baseline justify a modest score of 3.

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

The central claim rests on the fidelity of the simulated nonlinear folding model and on the transferability of synthetic ray-tracing training data. The model itself introduces no new physical entities. The free parameters are the coarse-offset bound, Huber delta, temporal window, and segment length, all chosen by hand.

free parameters (4)
  • Coarse offset error bound for training = 10 samples
    Section 8.2: authors train assuming a random coarse offset error of 10 samples after testing 40, 10, and 6, judging 40 unlikely. This choice directly shapes the model's robustness and the reported median error.
  • Huber loss delta = 1.0
    Section 6.2: set to 1.0 in experiments; controls the transition between quadratic and linear penalty behavior.
  • Temporal aggregation window = 10 frames
    Section 5 states N=20, but Section 6.1 and 8.2 use 10 consecutive frames for averaging; this averaging is central to SNR recovery and the reported 15x median-error reduction.
  • Input segment length = 286 samples
    Section 6.1: extracts the coarse offset plus 5 samples before and 280 after to cover the 276-sample sync-code; assumes the sync-code stays within this window.
assumptions (4)
  • domain assumption A squaring-based spectrum folding model of a passive envelope-detector receiver is faithful enough that a network trained on it transfers to real traces.
    Used in Sections 3 and 8.3; Section 9 admits real envelope detector transfer functions, filters, amplifiers, antenna frequency response, and clock drift are not accounted for.
  • domain assumption Ray-tracing synthetic data built from OpenStreetMap and CellMapper reproduces the multipath and SNR conditions of the real test route.
    Section 7 describes the data generation; Section 8.3 acknowledges the simulated route is denser and uses different bandwidth, so transfer is assumed, not validated.
  • domain assumption LTE PSS and SSS sync-codes retain enough structure after intermodulation for a CNN to learn sub-sample offsets.
    Stated in the contributions and in Section 5; it is central to the feasibility of the learning approach.
  • standard math Standard correlation-based coarse synchronization and trilateration are valid under the evaluated conditions.
    Used in Sections 2 and 5.1; no unusual assumptions beyond standard matched filtering and TDOA geometry.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepSync: A Learning Framework for Pervasive Localization using Code Synchronization on Compressed Cellular Spectrum." pith.science (2026). https://pith.science/paper/PTAWTUXL

@misc{pith2026250505783,
  author       = {Pith},
  title        = {Pith review of: DeepSync: A Learning Framework for Pervasive Localization using Code Synchronization on Compressed Cellular Spectrum},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PTAWTUXL}},
  note         = {Machine review of arXiv:2505.05783}
}
read the original abstract

Pervasive localization is essential for continuous tracking applications, yet existing solutions face challenges in balancing power consumption and accuracy. GPS, while precise, is impractical for continuous tracking of micro-assets due to high power requirements. Recent advances in non-linear compressed spectrum sensing offer low-power alternatives, but existing implementations achieve only coarse positioning through Received Signal Strength Indicator (RSSI) measurements. We present DeepSync, a deep learning framework that enables precise localization using compressed cellular spectrum. Our key technical insight lies in formulating sub-sample timing estimation as a template matching problem, solved through a novel architecture combining temporal CNN encoders for multi-frame processing with cross-attention mechanisms. The system processes non-linear inter-modulated spectrum through hierarchical feature extraction, achieving robust performance at SNR levels below -10dB -- a regime where conventional timing estimation fails. By integrating real cellular infrastructure data with physics-based ray-tracing simulations, DeepSync achieves 2.128-meter median accuracy while consuming significantly less power than conventional systems. Real-world evaluations demonstrate 10x improvement over existing compressed spectrum approaches, establishing a new paradigm for ultra-low-power localization.

Figures

Figures reproduced from arXiv: 2505.05783 by the authors.

Figure 1
Figure 1. DeepSync’s architecture ■ Cross-Attention Mechanism Attention mechanisms(Vaswani, 2017) have been widely adopted in wireless systems (Hamidi-Rad & Jain, 2021; Guo et al., 2024; Belgiovine et al., 2024). We employ a cross-attention mechanism to align the received signal rep￾resentations with the target sync-code, enhancing feature matching for offset estimation. Our model leverages cross-attention to effectively alig… view at source ↗
Figure 5
Figure 5. CDF error of sample offset: (a) With different offset size, (b) With different input sizes. Since DeepSync operates on inter-modulated and spectrum￾folded spectrum, it inherently suffers from poor SNR con￾ditions. Training with a single frame results in a median error exceeding 3 samples, which is insufficient for precise localization. However, due to the periodic nature of LTE frames, DeepSync can leverage multiple… view at source ↗
Figure 2
Figure 2. Localization performance of DeepSync vs (Garg et al., 2024) in a simulated urban setting. 8.2. Effect of Coarse Estimation Traditional correlation-based methods in inter-modulated spectrum lack precision, failing to accurately capture sub￾sample offsets and introducing spurious errors in sample off￾set estimation. In contrast, DeepSync effectively estimates sub-sample offsets with a nominal median error of just 0.23… view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: CDF of localization [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]
Figure 7
Figure 7. Figure 7: Comparison of different factors affecting DeepSync’s performance: (a) Effect of bandwidth on median sample error, (b) Median error for feature combination methods. 8.4. Feature Combination Methods To combine features from the target sync-code and the sig￾nal’s latent s…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

  1. [6]

    9 DeepSync: A Learning Framework for Pervasive Localization using Code Synchronization on Compressed Cellular Spectrum Gomes, H. C. and Carvalho, N. B. The use of intermodu- lation distortion for the design of passive rfid. In 2007 European Radar Conference, pp. 377–380. IEEE,

    work page 2007

  2. [8]

    A neural network aided approach for ldpc coded dco-ofdm with clipping distortion

    He, Y ., Jiang, M., Ling, X., and Zhao, C. A neural network aided approach for ldpc coded dco-ofdm with clipping distortion. In ICC 2019-2019 IEEE International Confer- ence on Communications (ICC), pp. 1–6. IEEE,

    work page 2019

  3. [10]

    Dwivedi, S., Shreevastav, R., Munier, F., Nygren, J., Siom- ina, I., Lyazidi, Y ., Shrestha, D., Lindmark, G., Ernstr¨om, P., Stare, E., et al

    1109/ACCESS.2020.3044812. Dwivedi, S., Shreevastav, R., Munier, F., Nygren, J., Siom- ina, I., Lyazidi, Y ., Shrestha, D., Lindmark, G., Ernstr¨om, P., Stare, E., et al. Positioning in 5g networks. IEEE Communications Magazine, 59(11):38–44,

    arXiv 2020

  4. [14]

    C., Weiss, A., Lancho, A., Polyanskiy, Y ., and Wornell, G

    Lee, G. C., Weiss, A., Lancho, A., Polyanskiy, Y ., and Wornell, G. W. On neural architectures for deep learning- based source separation of co-channel ofdm signals. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5. IEEE,

    work page 2023

  5. [15]

    Li, W., Jia, Y ., and Du, J

    doi: 10.1109/JLT.2022.3199439. Li, W., Jia, Y ., and Du, J. Toa-based cooperative localization for mobile stations with nlos mitigation. Journal of the Franklin Institute, 353(6):1297–1312,

    arXiv 2022

  6. [16]

    Network in network

    Lin, M. Network in network. arXiv preprint arXiv:1312.4400,

    arXiv

  7. [18]

    M., Chen, T., Horsmanheimo, S., and Scholliers, J

    M¨u¨ursepp, I., Kulmar, M., Elgarhy, O., Alam, M. M., Chen, T., Horsmanheimo, S., and Scholliers, J. Performance evaluation of 5g-nr positioning accuracy using time dif- ference of arrival method. In 2021 IEEE International Mediterranean Conference on Communications and Net- working (MeditCom), pp. 494–499. IEEE,

    work page 2021

  8. [19]

    In 2023 IEEE Conference on Standards for Communications and Networking (CSCN), pp. 171–176. IEEE,

    work page 2023

Show all 31 references
  1. [20]

    Fin- gerio: Using active sonar for fine-grained finger tracking

    Nandakumar, R., Iyer, V ., Tan, D., and Gollakota, S. Fin- gerio: Using active sonar for fine-grained finger tracking. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, pp. 1515–1525,

    2016

  2. [21]

    An introduction to convolutional neural net- works

    O’Shea, K. An introduction to convolutional neural net- works. arXiv preprint arXiv:1511.08458,

    arXiv

  3. [22]

    Accessed: 2024-06-28

    URL https://www.rfmw.com/products/ detail/rf2052-qorvo/326547/?gad_ source=1&gclid=EAIaIQobChMI0LTFvrT_ hgMVnWJHAR0uzQKnEAAYBCAAEgL_nfD_BwE. Accessed: 2024-06-28. Research, E. Usrp n210 software defined radio (sdr). https://www.ettus.com/all-products/ un210-kit/. Accessed: 202...

    2024

  4. [23]

    Localization in long-range ultra narrow band iot networks using rssi

    Sallouha, H., Chiumento, A., and Pollin, S. Localization in long-range ultra narrow band iot networks using rssi. In 2017 IEEE International Conference on Communications (ICC), pp. 1–6. IEEE,

    2017

  5. [24]

    Wireless interference identification with convolutional neural networks

    Schmidt, M., Block, D., and Meier, U. Wireless interference identification with convolutional neural networks. In 2017 IEEE 15th International Conference on Industrial Informatics (INDIN), pp. 180–185. IEEE,

    2017

  6. [25]

    Shahzad, A. R. and Jalal, A. A smart surveillance system for pedestrian tracking and counting using template match- ing. In 2021 International Conference on Robotics and Automation in Industry (ICRAI), pp. 1–6. IEEE,

    2021

  7. [26]

    K., Ganti, R

    Singh, R., Yerrapragada, A. K., Ganti, R. K., et al. Enhance- ments for 5g nr prach reception: An ai/ml approach. In 2024 Wireless Telecommunications Symposium (WTS), pp. 1–6. IEEE,

    2024

  8. [28]

    Unveiling energy efficiency in deep learn- ing: Measurement, prediction, and scoring across edge devices

    Tu, X., Mallik, A., Chen, D., Han, K., Altintas, O., Wang, H., and Xie, J. Unveiling energy efficiency in deep learn- ing: Measurement, prediction, and scoring across edge devices. In 2023 IEEE/ACM Symposium on Edge Com- puting (SEC), pp. 80–93. IEEE,

    2023

  9. [29]

    Overpass turbo

    Turbo, O. Overpass turbo. https:// overpass-turbo.eu/. Accessed: 2025-01-14. Varshney, A. and Corneo, L. Tunnel emitter: Tunnel diode based low-power carrier emitters for backscatter tags. In Proceedings of the 26th Annual International Conference on Mobile Computing and Netwo...

    2025

  10. [31]

    Bat: Learning to reason about spatial sounds with large language models

    Zheng, Z., Peng, P., Ma, Z., Chen, X., Choi, E., and Harwath, D. Bat: Learning to reason about spatial sounds with large language models. arXiv preprint arXiv:2402.01591,

    arXiv

  11. [1987]

    STM32N6 Se- ries

    STMicroelectronics. STM32N6 Se- ries. https://www.st.com/en/ microcontrollers-microprocessors/ stm32n6-series.html. Accessed: 2025-01-29. Sun, M., Xiao, J., Lim, E. G., Zhang, B., and Zhao, Y . Fast template matching and update for video object tracking and segmentation. In Pr...

    2025

  12. [2010]

    gov/sites/default/files/meetings/lte_ technical_characteristics_0.pdf

    URL https://www.ntia. gov/sites/default/files/meetings/lte_ technical_characteristics_0.pdf. Ac- cessed: 2025-01-14. Ninkovic, V ., Vukobratovic, D., Valka, A., and Dumic, D. Preamble-based packet detection in wi-fi: A deep learning approach. In 2020 IEEE 92nd Vehicular Techno...

    2025

  13. [2011]

    Batch normalization: Accelerating deep net- work training by reducing internal covariate shift

    Ioffe, S. Batch normalization: Accelerating deep net- work training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167,

    arXiv

  14. [2015]

    and Jagannath, J

    Jagannath, A. and Jagannath, J. Multi-task learning ap- proach for automatic modulation and wireless signal clas- sification. In ICC 2021-IEEE International Conference on Communications, pp. 1–7. IEEE,

    2021

  15. [2016]

    In-body backscatter communication and localization

    Vasisht, D., Zhang, G., Abari, O., Lu, H.-M., Flanz, J., and Katabi, D. In-body backscatter communication and localization. In Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication, pp. 132–146,

    2018

  16. [2017]

    Spatial signal strength prediction using 3d maps and deep learning

    Krijestorac, E., Hanna, S., and Cabric, D. Spatial signal strength prediction using 3d maps and deep learning. In ICC 2021-IEEE international conference on communica- tions, pp. 1–6. IEEE,

    2021

  17. [2018]

    T-prime: Transformer- based protocol identification for machine-learning at the edge

    Belgiovine, M., Groen, J., Sirera, M., Tassie, C., Trudeau, S., Ioannidis, S., and Chowdhury, K. T-prime: Transformer- based protocol identification for machine-learning at the edge. In IEEE INFOCOM 2024-IEEE Conference on Computer Communications, pp. 1031–1040. IEEE,

    2024

  18. [2019]

    F., Hoang, A

    Ensworth, J. F., Hoang, A. T., and Reynolds, M. S. A low power 2.4 ghz superheterodyne receiver architecture with external lo for wirelessly powered backscatter tags and sensors. In 2017 IEEE International Conference on RFID (RFID), pp. 149–154. IEEE,

    2017

  19. [2020]

    Agarap, A

    Available online: https://www.3gpp.org/DynaReport/22872.htm. Agarap, A. Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375,

    arXiv

  20. [2021]

    and Youssef, M

    Elbakly, R. and Youssef, M. Crescendo: An infrastructure- free ubiquitous cellular network-based localization sys- tem. In 2019 IEEE Wireless Communications and Net- working Conference (WCNC), pp. 1–6. IEEE,

    2019

  21. [2022]

    and Jain, S

    Hamidi-Rad, S. and Jain, S. Mcformer: A transformer based deep neural network for automatic modulation classifica- tion. In 2021 IEEE Global Communications Conference (GLOBECOM), pp. 1–6. IEEE,

    2021

  22. [2023]

    Aoa-toa based lo- calization for 5g cell-less communications

    Kim, S., Park, S., Ji, H., and Shim, B. Aoa-toa based lo- calization for 5g cell-less communications. In 2017 23rd Asia-Pacific Conference on Communications (APCC), pp. 1–6. IEEE,

    2017

  23. [2024]

    URL https:// www.mathworks.com/help/lte/ug/ synchronization-signals-pss-and-sss. html. Accessed: 2024-01-29. Mayer, C., Danelljan, M., Bhat, G., Paul, M., Paudel, D. P., Yu, F., and Van Gool, L. Transforming model prediction for tracking. In Proceedings of the IEEE/CVF Confere...

    2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.