Pith. sign in

REVIEW 3 major objections 5 minor 30 references

Gating a plane-wave phase loss to LoS samples cuts AoA error by up to 6° under multipath domain shift.

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 · grok-4.5

2026-07-12 12:56 UTC pith:OXIY7772

load-bearing objection Solid engineering paper: real multipath campaign + LoS-gated plane-wave loss inside DIL beats standard continual baselines by a few degrees under small replay; the unsupervised gate is the softest link, not a collapse of the claim. the 3 major comments →

arxiv 2607.02537 v1 pith:OXIY7772 submitted 2026-06-21 eess.SP cs.LGstat.ML

Physics-Informed Domain-Invariant Feature Learning with Autoencoder-Driven Gaussian Clustering for Robust Non-line-of-Sight Scenarios

Nisha L. Raichur , Lucas Heublein , Dominik Seu{\ss} , Frank Deinzer , Felix Ott This is my paper
classification eess.SP cs.LGstat.ML
keywords angle-of-arrival estimationphysics-informed neural networksdomain-incremental learningLoS/NLoS classificationGNSS interferencemultipathantenna arrays
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.

Direction finding of jamming or spoofing signals is reliable in clear line-of-sight but collapses when walls and scatterers create multipath. This paper shows that a deep regressor for azimuth and elevation can be made far more transferable if a physics loss—matching predicted angles to inter-antenna phase differences under a plane-wave model—is applied only to samples that look LoS. An autoencoder plus Gaussian clustering on a handful of statistical IQ features supplies that unsupervised LoS mask. Combined with domain-incremental training across environments with one to four walls, the gated physics regularizer reduces average angle error by as much as six degrees when only a few exemplars from earlier domains can be replayed. A sympathetic reader cares because real interference localization almost always faces changing multipath and scarce labels; a cheap physical prior that fires only when it is valid is a practical way to keep earlier knowledge without over-constraining the model in NLoS.

Core claim

On real indoor multipath recordings with sequential 1→2→3→4 wall domains, selectively applying a plane-wave phase-difference loss only to LoS-dominant samples (identified by autoencoder latent GMM clustering) yields domain-invariant features that improve cross-domain AoA estimation by up to roughly 6° versus standard DIL baselines, especially when the replay buffer is small.

What carries the argument

Selective physics-informed loss L_p: the mean-squared circular error between observed and geometry-predicted inter-antenna phase differences, multiplied by a hard LoS mask (GMM posterior > τ) so the plane-wave constraint never fights multipath samples.

Load-bearing premise

The unsupervised LoS detector must correctly flag the samples for which the plane-wave phase model is actually valid; if that mask is noisy, the physics loss either does nothing useful or injects false structure.

What would settle it

Hold out a labeled LoS/NLoS subset with independent geometric ground truth and recompute the full sequential 1–4 wall table after ablating the LoS gate (or after injecting controlled mask errors); if the 6° gap disappears, the claim fails.

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

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

3 major / 5 minor

Summary. The manuscript proposes a hybrid AoA estimator for GNSS interference localization on a four-element array. An XceptionTime regressor predicts relative position and unit-vector azimuth/elevation; a plane-wave physics loss on inter-antenna phase differences (Eqs. 6–8) is applied only to samples whose LoS posterior from an AE+GMM pipeline exceeds τ. Domain-incremental learning across four wall-induced multipath domains with replay buffers of size N_t ∈ {100,500,1000} is used for sequential adaptation. On a real indoor USRP campaign, the method reports lower average angular error than Regular, Distillation, EWC, and MAS baselines, with the largest gains (up to ~6°) in the low-exemplar regime, and shows that SHAP-reduced four-feature inputs preserve LoS/NLoS separation and localization accuracy.

Significance. If the LoS-gated physics mechanism is the true source of the reported gains, the work is a solid systems contribution for RF interference direction finding under multipath: it combines a real multi-domain measurement campaign, selective physics regularization grounded in known array geometry, unsupervised LoS/NLoS masking, and standard DIL baselines with low-exemplar replay. The SHAP-guided feature reduction and efficiency comparison are practical strengths. The result would be of interest to GNSS anti-jamming and indoor RF localization communities, provided the gating premise is substantiated.

major comments (3)
  1. [§III-C–D, §V-f, Table I] §III-C–D and §V-f: The central claim that LoS-gated L_p yields domain-invariant features (and the headline ~6° gain in Table I at N_t=100) depends on the AE+GMM posterior correctly selecting samples for which the plane-wave model (Eqs. 6–7) is approximately valid. The GMM is trained only on the two-wall training split; no external LoS/NLoS ground truth, precision/recall, or τ-sensitivity is reported. Without that validation, gains could be ordinary replay plus mild regularization rather than physics-driven invariance. Please add a quantitative gate evaluation (e.g., against geometry-based LoS labels or controlled free-space/blocked subsets) and a τ ablation.
  2. [§V-f, Fig. 6] §V-f and Fig. 6: In the four-wall domain, natural LoS counts collapse and the authors inject an ad-hoc floor of five LoS replay samples per batch so that L_p does not vanish. This intervention is load-bearing for training stability yet is not ablated against pure random replay or against disabling L_p on that domain. Report results without the floor (or with alternative stabilization) so readers can separate the contribution of the physics term from the forced LoS replay policy.
  3. [Table I, Abstract, Fig. 7] Table I: Average azimuth improvements at N_t=100 (20.95° vs 22.95–26.84°) are modest relative to residual error (~21°) and domain-shift magnitudes in Table II (tens of degrees). Report mean±std over seeds for all methods (Fig. 7 already uses five seeds for λ only), and clarify whether the “up to 6°” abstract claim is average-azimuth vs a specific baseline or a peak pairwise difference. Without variance and a clear definition, the significance of the low-exemplar claim is hard to judge.
minor comments (5)
  1. [§III-A] Notation: α/β vs θ_α/θ_β and ê_α vs e_α are introduced with slight inconsistency; unify angle encoding early in §III-A.
  2. [Eq. (8)] Eq. (8) uses arctan2(sin δ, cos δ) then squares; state explicitly that this is a circular phase error and that Δϕ is unwrapped or modulo-2π consistent with the plane-wave construction.
  3. [Fig. 2] Fig. 2 color scale is azimuth error but the caption does not state units or the exact metric; add a colorbar label and clarify whether values are absolute degrees.
  4. [§II] Related work cites several PINN/AoA and DIL papers; a short explicit contrast with Bayesian physics-informed RFID AoA [17] and transfer-learning indoor DOA [13] on what is new (gated plane-wave + DIL on wall domains) would help.
  5. [Abstract, title block] Typographical: “6{\deg}” / “6 ◦” spacing and “N ¨urnberg” encoding; ensure consistent degree symbols and author affiliations in the camera-ready version.

Circularity Check

0 steps flagged

No circular derivation chain; empirical DIL gains rest on an external plane-wave regularizer and a separately trained unsupervised gate, not on any quantity forced by construction from its own inputs.

full rationale

The paper is a supervised hybrid method paper whose headline numbers (Table I, up to ~6° average azimuth reduction under small replay) are measured comparisons against Regular/Distillation/EWC/MAS baselines on real sequential wall-domain data, not first-principles predictions. The physics loss (Eq. 8) is the circular distance between phase differences obtained by applying the known plane-wave map (Eqs. 6–7, fixed λ=0.1903 m, d=0.09 m array geometry) to predicted versus label angles, gated by an independently trained AE+GMM LoS posterior (Eqs. 3–4, features from Wu et al., SHAP-reduced). This is ordinary physics-informed regularization of an already-supervised angle/position MSE (Eq. 2); it does not redefine the targets in terms of the loss, fit a free parameter on a subset and then “predict” a near-identical quantity, or import a uniqueness theorem. The LoS/NLoS clustering is unsupervised and interpreted post-hoc via spatial maps (Fig. 3); its lack of external ground-truth accuracy is a correctness/assumption risk, not a definitional loop that makes L_p or the 6° gain tautological. Self-citations ([14]–[16]) concern prior DIL sampling work and are not load-bearing for the physics claim. No ansatz is smuggled, no known empirical pattern is merely renamed, and the derivation is self-contained against the reported real-world benchmarks.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The load-bearing story is empirical: plane-wave consistency on GMM-selected LoS samples plus replay DIL improves cross-wall AoA. Free knobs (λ, τ, buffer size, forced LoS count, latent dim) and domain assumptions (plane wave under LoS; LoS phases domain-invariant; wall count as domain proxy) carry the claim. No new physical particle or force is invented; the ‘entities’ are methodological constructs (gated physics loss, AE–GMM LoS mask) whose only evidence is this paper’s tables.

free parameters (6)
  • physics loss weight λ = 0.3 (selected)
    Balances L_l and L_p; grid-searched in Fig. 7; λ=0.3 chosen as best trade-off especially at low N_t.
  • LoS probability threshold τ = >0.9
    Gates which batch samples enter L_p; set high (τ>0.9) without independent calibration against labeled LoS.
  • replay buffer size N_t = 100 / 500 / 1000
    Number of stored exemplars per past domain; central low-exemplar claim depends on N_t∈{100,500,1000}.
  • minimum LoS samples per batch (4-wall) = 5
    Ad-hoc floor of five LoS replay samples enforced when L_p vanished; stabilizes training but is not derived.
  • AE latent dimension d = 4
    Fixed to 4 for GMM clustering of LoS/NLoS; affects mask quality.
  • GMM components K = 2
    Hard-coded two-component full-covariance GMM for LoS vs NLoS.
axioms (5)
  • domain assumption Under LoS, inter-antenna phase differences obey the plane-wave model ϕ = (2π/λ) u R^⊤ for the known 2×2 geometry.
    §III-D; standard array processing assumption; used to define L_p.
  • domain assumption LoS phase structure is comparatively domain-invariant across wall configurations, so constraining only LoS samples yields transferable features.
    Stated in abstract and §III-D; motivates selective application of L_p.
  • ad hoc to paper A two-component GMM on AE latent statistical features separates LoS from NLoS well enough to gate the physics loss.
    §III-C; unsupervised labels never scored against external LoS truth.
  • domain assumption Wall count (1–4) defines sequential domains with shifted multipath distributions suitable for DIL evaluation.
    §IV DIL setup; experimental design choice, not a universal multipath taxonomy.
  • standard math MSE on position and unit-circle angle encodings is an adequate primary localization objective.
    Eq. (2); standard regression choice.
invented entities (2)
  • LoS-probability-gated plane-wave physics loss L_p no independent evidence
    purpose: Regularize AoA predictions only when GMM says LoS, to learn domain-invariant features without over-constraining NLoS.
    Core methodological construct of the paper; evidence is internal ablation vs DIL baselines only.
  • AE + GMM unsupervised LoS/NLoS mask pipeline (with SHAP-reduced 4-feature variant) no independent evidence
    purpose: Provide M_LoS for selective physics loss and reduce feature dimensionality.
    Composition of known tools; no external labeled LoS benchmark validates the mask.

pith-pipeline@v1.1.0-grok45 · 18172 in / 3939 out tokens · 42917 ms · 2026-07-12T12:56:10.316775+00:00 · methodology

0 comments
read the original abstract

Jamming and spoofing pose significant threats to wireless and satellite navigation by disrupting radio-frequency (RF) signals and compromising availability and integrity. Robust RF interference direction finding through angle-of-arrival (AoA) estimation is therefore essential for detecting and localizing anomalous signals. Although data-driven methods perform well under line-of-sight (LoS) conditions, their performance degrades in practical environments due to non-line-of-sight (NLoS) multipath propagation. In this work, we propose a hybrid learning framework that incorporates physics-informed constraints into deep neural networks to improve the robustness of AoA estimation. A neural network is trained to estimate the azimuth and elevation of incoming signals received by a four-element antenna array, while a physics-informed loss enforces consistency between the predicted angles and inter-antenna phase differences under a plane-wave model. We further introduce a latent-space classifier to distinguish LoS from NLoS samples. Since inter-antenna phase differences under LoS propagation exhibit domain-invariant structure across environments, the physics-based loss is applied only to LoS samples, promoting physically consistent and domain-invariant representations without over-constraining the model in NLoS scenarios. In addition, domain-incremental learning (DIL) across NLoS environments with varying scatterer distributions improves cross-domain generalization. Evaluations on real-world datasets show that the proposed method reduces AoA estimation error by up to 6{\deg} in low-exemplar settings compared with DIL baselines.

Figures

Figures reproduced from arXiv: 2607.02537 by Dominik Seu{\ss}, Felix Ott, Frank Deinzer, Lucas Heublein, Nisha L. Raichur.

Figure 1
Figure 1. Figure 1: Method overview. IQ samples are encoded into a latent space and clustered to identify LoS/NLoS. A LoS mask activates the physics-informed loss during AoA training, while DIL adapts the model across environments using limited exemplars. is designed to learn a sequence of domains corresponding to distinct environmental scenarios. In this study, each domain is defined by a different number of walls, which ind… view at source ↗
Figure 2
Figure 2. Figure 2: Evaluation of our physics-informed DIL method with [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Histogram of LoS prob￾abilities by the AE + GMM. 0 10 20 30 40 LoS count in each batch 0 50 100 150 Number of Batches 1 wall 2 walls 3 walls 4 walls [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of phase differences between antennas [PITH_FULL_IMAGE:figures/full_fig_p006_8.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

30 extracted references · 2 linked inside Pith

  1. [1]

    Angle-of-Arrival Localization Based on Antenna Ar- rays for Wireless Sensor Networks,

    P. Kułakowski, J. Vales-Alonso, E. Egea-L ´opez, W. Ludwin, and J. Garc ´ıa-Haro, “Angle-of-Arrival Localization Based on Antenna Ar- rays for Wireless Sensor Networks,” inComputers & Electrical Engi- neering, vol. 36(6), Nov. 2010

    2010

  2. [2]

    Multiple Emitter Location and Signal Parameter Estima- tion,

    R. Schmidt, “Multiple Emitter Location and Signal Parameter Estima- tion,” inIEEE Trans. on Antennas and Propagation (TAP), vol. 34(3), Mar. 1986, pp. 276–280

    1986

  3. [3]

    ESPRIT – Estimation of Signal Parameters via Rotational Invariance Techniques,

    R. Roy and T. Kailath, “ESPRIT – Estimation of Signal Parameters via Rotational Invariance Techniques,” inIEEE Trans. on Signal Processing (TSP), vol. 37(7), Jul. 2002, pp. 984–995

    2002

  4. [4]

    Improving the Resolution Performance of Eigenstructure- Based Direction-Finding Algorithms,

    A. Barabell, “Improving the Resolution Performance of Eigenstructure- Based Direction-Finding Algorithms,” inIEEE Intl. Conf. on Acoustics, Speech, and Signal Processing (ICASSP), vol. 8, Boston, MA, 1983, pp. 336–339

    1983

  5. [5]

    Deep Networks for Direction-of-Arrival Estimation in Low SNR,

    G. K. Papageorgiou, M. Sellathurai, and Y . C. Eldar, “Deep Networks for Direction-of-Arrival Estimation in Low SNR,” inIEEE Trans. on Signal Processing (TSP), vol. 69, Jun. 2021, pp. 3714–3729

    2021

  6. [6]

    GNSS Jammer Direction Finding in Dynamic Scenarios Using an Inertial-based Multi- Antenna System,

    L. Heublein, T. Nowak, T. Feigl, J. Pahl, and F. Ott, “GNSS Jammer Direction Finding in Dynamic Scenarios Using an Inertial-based Multi- Antenna System,” inIEEE DGON Inertial Sensors and Applications (ISA), Braunschweig, Germany, Oct. 2025

    2025

  7. [7]

    Towards an AI-Driven Universal Anti- Jamming Solution with Convolutional Interference Cancellation Net- work,

    H. N. Nguyen and G. Noubir, “Towards an AI-Driven Universal Anti- Jamming Solution with Convolutional Interference Cancellation Net- work,” inarXiv preprint arXiv:2203.09717, Mar. 2022

    Pith/arXiv arXiv 2022

  8. [8]

    Source Localization and Sensing: A Nonparametric Iterative Adaptive Approach Based on Weighted Least Squares,

    T. Yardibi, J. Li, P. Stoica, M. Xue, and A. B. Baggeroer, “Source Localization and Sensing: A Nonparametric Iterative Adaptive Approach Based on Weighted Least Squares,” inIEEE Trans. on Aerospace and Electronic Systems (TAES), vol. 46(1), Jan. 2010, pp. 425–443

    2010

  9. [9]

    Attention-Based Fusion of IQ and FFT Spectrograms with AoA Features for GNSS Jammer Localization,

    L. Heublein, C. Wielenberg, T. Nowak, T. Feigl, C. Mutschler, and F. Ott, “Attention-Based Fusion of IQ and FFT Spectrograms with AoA Features for GNSS Jammer Localization,” inIEEE Radar Conference (RadarConf), Krakow, Poland, Oct. 2025

    2025

  10. [10]

    DeepAoANet: Learning Angle-of-Arrival From Software Defined Radios with Deep Neural Networks,

    Z. Dai, Y . He, V . Tran, N. Trigoni, and A. Markham, “DeepAoANet: Learning Angle-of-Arrival From Software Defined Radios with Deep Neural Networks,” inIEEE Access, vol. 10, Jan. 2022, pp. 3164–3176

    2022

  11. [11]

    Direction-of-Arrival Estimation of Multipath Signals Using Independent Component Analysis and Compressive Sensing,

    L. Zhao, J. Xu, J. Ding, A. Liu, and L. Li, “Direction-of-Arrival Estimation of Multipath Signals Using Independent Component Analysis and Compressive Sensing,” inPLoS One, vol. 12(7), Jul. 2017

    2017

  12. [12]

    Estimating Angle-of-Arrival and Time-of-Flight for Mul- tipath Components Using WiFi Channel State Information,

    A. U. Ahmed, R. Arablouei, F. D. Hoog, B. Kusy, R. Jurdak, and N. Bergmann, “Estimating Angle-of-Arrival and Time-of-Flight for Mul- tipath Components Using WiFi Channel State Information,” inMDPI Sensors, vol. 18(6), May 2018

    2018

  13. [13]

    Domain-Adaptive Direction of Arrival (DOA) Estimation in Complex Indoor Environments Based on Convolutional Autoencoder and Transfer Learning,

    L. Shen, J. Li, J. Pan, J. Shi, R. Xu, H. Wang, and W. Deng, “Domain-Adaptive Direction of Arrival (DOA) Estimation in Complex Indoor Environments Based on Convolutional Autoencoder and Transfer Learning,” inMDPI Sensors, vol. 25, no. 10, 2025, p. 2959

    2025

  14. [14]

    Bayesian Learning-driven Prototypical Contrastive Loss for Class-Incremental Learning,

    N. L. Raichur, L. Heublein, T. Feigl, A. R ¨ugamer, C. Mutschler, and F. Ott, “Bayesian Learning-driven Prototypical Contrastive Loss for Class-Incremental Learning,” inTrans. on Machine Learning Research (TMLR), Apr. 2025

    2025

  15. [15]

    5G-DIL: Domain Incremental Learning with Similarity-Aware Sampling for Dynamic 5G Indoor Localization,

    N. L. Raichur, L. Heublein, C. Mutschler, and F. Ott, “5G-DIL: Domain Incremental Learning with Similarity-Aware Sampling for Dynamic 5G Indoor Localization,” inIEEE Intl. Conf. on Localization and GNSS (ICL-GNSS), Rome, Italy, Jun. 2025

    2025

  16. [16]

    Adaptive & Efficient Sample Selection for Domain Incremental 5G Indoor Localization in Dynamic Multipath Environments,

    N. L. Raichur, L. Heublein, J. Pahl, D. Seuß, C. Mutschler, and F. Ott, “Adaptive & Efficient Sample Selection for Domain Incremental 5G Indoor Localization in Dynamic Multipath Environments,” inIEEE Journal of Indoor and Seamless Positioning and Navigation (J-ISPIN), vol. 4, Jan. 2026, pp. 23–40

    2026

  17. [17]

    Physics-Informed Bayesian Angle-of-Arrival Estimation for Commercial Off-The-Shelf RFID with Sub-Degree Accuracy,

    N. M. Benelmekki, A. Morell, G. Zamora, J. Bonache, and J. L. Vicario, “Physics-Informed Bayesian Angle-of-Arrival Estimation for Commercial Off-The-Shelf RFID with Sub-Degree Accuracy,” inEl- sevier Internet of Things (IOT), vol. 37, May 2026

    2026

  18. [18]

    Physics-Informed Neural Networks for Wireless Channel Estimation with Limited Pilot Signals,

    A. Javid and N. G. Prelcic, “Physics-Informed Neural Networks for Wireless Channel Estimation with Limited Pilot Signals,” inNeurIPS Workshop, Sep. 2025

    2025

  19. [19]

    Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward and Inverse Problems Involving Nonlinear Partial Differential Equations,

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward and Inverse Problems Involving Nonlinear Partial Differential Equations,” in Elsevier Journal of Computational Physics, vol. 378, Feb. 2019, pp. 686–707

    2019

  20. [20]

    Physics-Informed Neural Networks (PINNs) for Wave Propagation and Full Waveform Inversions,

    M. Rasht-Behesht, C. Huber, K. Shukla, and G. E. Karniadakis, “Physics-Informed Neural Networks (PINNs) for Wave Propagation and Full Waveform Inversions,” inWILEY Journal of Geophysical Research: Solid Earth, vol. 127(5), Apr. 2022

    2022

  21. [21]

    Phase-Resolved Wave Prediction with Linear Wave Theory and Physics-Informed Neural Networks,

    Y . Liu, X. Zhang, Q. Dong, G. Chen, and X. Li, “Phase-Resolved Wave Prediction with Linear Wave Theory and Physics-Informed Neural Networks,” inElsevier Applied Energy, vol. 355, Feb. 2024

    2024

  22. [22]

    Physics-Informed Neural Networks with Architectural Physics Embedding for Large-Scale Wave Field Reconstruction,

    H. Zhang, F. Ye, and C. Ma, “Physics-Informed Neural Networks with Architectural Physics Embedding for Large-Scale Wave Field Reconstruction,” inarXiv preprint arXiv:2603.02231, Feb. 2026

    arXiv 2026

  23. [23]

    Overcoming Catastrophic Forgetting in Neural Networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran, and R. Hadsell, “Overcoming Catastrophic Forgetting in Neural Networks,” inApplied Mathematics, vol. 114(13), Mar. 2017, pp. 3521–3526

    2017

  24. [24]

    Memory Aware Synapses: Learning What (not) to Forget,

    R. Aljundi, F. Babiloni, M. Elhoseiny, M. Rohrbach, and T. Tuytelaars, “Memory Aware Synapses: Learning What (not) to Forget,” inEurop. Conf. on Computer Vision (ECCV), Munich, Germany, Sep. 2018, pp. 144–161

    2018

  25. [25]

    Learning Without Forgetting,

    Z. Li and D. Hoiem, “Learning Without Forgetting,” inIEEE Trans. on Pattern Analysis and Machine Intelligence (TPAMI), vol. 40(12), Dec. 2018, pp. 2935–2947

    2018

  26. [26]

    XceptionTime: A Novel Deep Architecture based on Depthwise Sepa- rable Convolutions for Hand Gesture Classification,

    E. Rahimian, S. Zabihi, S. F. Atashzar, A. Asif, and A. Mohammadi, “XceptionTime: A Novel Deep Architecture based on Depthwise Sepa- rable Convolutions for Hand Gesture Classification,” inarXiv preprint arXiv:1911.03803, Nov. 2019

    Pith/arXiv arXiv 1911

  27. [27]

    tsai – A State-of-the-art Deep Learning Library for Time Series and Sequential Data,

    I. Oguiza, “tsai – A State-of-the-art Deep Learning Library for Time Series and Sequential Data,” Github, 2023. [Online]. Available: https://github.com/timeseriesAI/tsai

    2023

  28. [28]

    TransAoA: Transformer-Based Angle of Arrival Estimation for BLE Indoor Local- ization,

    W. Wu, D. Zhou, L. Shen, Z. Zhao, C. Li, and G. Q. Huang, “TransAoA: Transformer-Based Angle of Arrival Estimation for BLE Indoor Local- ization,” inIEEE Trans. on Instrumentation and Measurements (TIM), vol. 74, 2025

    2025

  29. [29]

    Gaussian Mixture Models

    D. A. Reynolds, “Gaussian Mixture Models.” inEncyclopedia of Bio- metrics, vol. 741, 2009, pp. 659–663

    2009

  30. [30]

    A Unified Approach to Interpreting Model Predictions,

    S. M. Lundberg and S.-I. Lee, “A Unified Approach to Interpreting Model Predictions,” inAdvances in Neural Information Processing Systems (NIPS), Dec. 2017, pp. 4768–4777

    2017