Pith. sign in

REVIEW 3 major objections 5 minor 36 references

WiFuse: An Attention Mechanism for Human Activity Recognition using Fused CSI Amplitude and Delay-Doppler Channel Features

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read WiFuse claims that fusing denoised amplitude with sanitized-phase Delay–Doppler features, processed by a ResNet–TCN with attention and two-stage transfer learning, lifts Wi-Fi CSI-only activity recognition to 95.28% (XRF55) and 98.20% (Wi-M

desk verdict Useful dual-stream architecture for CSI-based HAR, but the headline accuracy gains over DML are not yet supported because the comparison uses mismatched train/test splits and single-run numbers. read the letter →

arxiv 2608.00642 v1 pith:VCJH6CDO submitted 2026-08-01 cs.CV

Alison M. Fernandes , Hermes I. Del Monego , Bruno S. Chang , Anelise Munaretto , Hélder M. Fontes , Rui L. Campos This is my paper · ORCID
classification cs.CV
keywords HumanActivityRecognitionChannelStateInformationDelay–DopplerDual-StreamFusionResNetTemporalConvolutionalNetworkTransferLearningWi-FiSensing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a Wi-Fi sensing system can recognize human activities more reliably by fusing two complementary views of the channel state: the denoised signal strength over time and a Delay–Doppler map of motion-induced velocity, derived from sanitized channel phase. The two views are concatenated and processed by a ResNet–TCN hybrid with channel and spatio-temporal attention, trained in two stages. On the 55-class XRF55 benchmark the framework reports up to 95.28% accuracy across four rooms using Wi-Fi only, and up to 98.20% on the 17-class multi-person Wi-MIR dataset. The paper argues that each stream alone is weak (neither exceeds 90%) but their fusion, plus phase sanitization and transfer learning, produces the gains.

What carries the argument

The central object is the dual-stream input tensor of shape (2×F×T): channel 0 is the denoised time-domain CSI amplitude, channel 1 is the Delay–Doppler spectrum obtained by IFFT across subcarriers followed by FFT across packets on the sanitized phase, then log-compressed and resampled to 1000 Doppler bins. The architecture is a ResNet (1D residual blocks with channel attention over the 540 fused channels) whose weights initialize a second stage in which a dilated causal TCN with spatio-temporal attention models long-range temporal dependencies, with layer-specific learning rates. The two-stage transfer learning (train ResNet alone, then fine-tune the hybrid with the ResNet layers nearly fro

What would settle it

Run the full WiFuse pipeline on XRF55 under the official 7:3 train–test split with subject-disjoint partitions: if the accuracy advantage over amplitude-only training shrinks to the noise level, the generalization claim fails. Alternatively, replace the sanitized phase input with random phase while keeping everything else identical; if accuracy stays near 95%, the Delay–Doppler stream is not carrying motion information.

Watch

Extended reading notes

Core claim

WiFuse establishes that feature-level fusion of a denoised time-domain amplitude stream and a 2D-FFT Delay–Doppler stream computed from sanitized CSI phase, fed into a ResNet backbone with channel attention followed by a dilated TCN with spatio-temporal attention, improves CSI-only human activity recognition under multipath, noise, and multi-user interference. In the paper's experiments the full pipeline reaches up to 95.28% average accuracy on XRF55's four environments and up to 98.20% on Wi-MIR, outperforming an amplitude-only CSI baseline by 2.77–4.45 percentage points and the multi-user reference by about 3 percentage points. Ablations attribute the gains to the combination of phase sani

Load-bearing premise

The Delay–Doppler stream is only physically informative if phase sanitization removes all hardware-induced phase offsets and random distortions; if residual phase noise survives, that stream is largely noise and the fusion gain would not be what the paper claims.

Editorial extensions

If this is right

  • CSI-only systems could close much of the gap to multimodal (RFID+mmWave+Wi-Fi) recognition: WiFuse reports within ~1–2 points of full DML while using one modality.
  • Neither the amplitude stream nor the Delay–Doppler stream alone reaches 90% accuracy in the paper's ablations, while their concatenation exceeds 93%, implying the fusion itself is the decisive ingredient.
  • Phase sanitization contributes a consistent but small gain (0.4–1.1 percentage points across scenes), suggesting that residual phase artifacts are a secondary but real source of error.
  • The TCN and transfer learning add roughly 8 percentage points on Wi-MIR compared to a standalone ResNet, indicating that temporal modeling matters most for multi-person interaction classes.
  • Zero-shot cross-dataset transfer fails (8.99–34.11%), so the learned representations are dataset-specific unless joint training or adaptation is applied.

Reading between the lines

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

  • If the Delay–Doppler stream's value is tied to strong motion, the same framework could be made adaptive by gating the two streams by estimated Doppler energy, potentially improving low-motion classes the paper notes benefit least.
  • Because the Delay–Doppler stream alone scores only ~60% on Scene 4, a skeptic could test whether the fusion gain comes from genuine physical information or from regularization/ensemble effects of a second input; replacing the phase-derived stream with a random or shuffled version would separate these.
  • The paper's own planned re-evaluation under the official 7:3 protocol will be the sharper test: the current 80/20 sample-wise split may overstate generalization if the same subjects appear in training and test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript proposes WiFuse, a dual-stream CSI-based HAR framework that concatenates denoised time-domain amplitude features with Delay–Doppler features computed from sanitized phase, feeds the fused representation to a ResNet–TCN with channel and spatio-temporal attention, and trains it with a two-stage transfer-learning strategy. Experiments on XRF55 and Wi-MIR report accuracies up to 95.28% and 98.20%, respectively, and compare against DML, multiple hybrid baselines, ablations, domain-adaptation settings, and computational-cost metrics. The central claim is that feature-level fusion of amplitude and Delay–Doppler representations improves recognition under multipath, noise, and multi-user interaction.

Significance. If the central claim were established, WiFuse would be a useful contribution: it combines two physically complementary CSI views in a single architecture and reports substantially lower parameter and GFLOP counts than a three-modality DML baseline. The breadth of the evaluation—public datasets, hybrid-architecture benchmarks, ablations, cross-dataset tests, and complexity analysis—is a strength, and the authors are transparent about several caveats. However, the headline outperformance over DML and over the closest hybrid baseline is not yet supported because of a train/test protocol mismatch and an unresolved statistical-reporting contradiction. These issues are fixable, so the work is promising but currently requires major revision.

major comments (3)
  1. [§IV.D, Table IV, Fig. 3] Table IV juxtaposes WiFuse with DML and DML-Only-CSI, and Fig. 3 concludes 'substantial gains' of 2.77–4.45 pp. The comparison is not on equal footing: Section III.C states that WiFuse uses an 80/20 sample-wise split, and Section IV.D acknowledges that the official DML figures come from the 7:3 protocol and that the sample-wise split 'can put the same subjects in both the training and test sets.' With Scene 1 containing 30 participants, a sample-wise split can inflate accuracy through subject-specific memorization. The authors themselves state that the XRF55 results characterize 'within-scene robustness rather than cross-subject generalization,' yet the table presents these numbers alongside the official 7:3 DML numbers as if comparable. A matched-protocol comparison—ideally a subject-wise 7:3 or cross-subject split—with error bars is required before the headline improvement is establish
  2. [§IV opening paragraph; §V.M; Table V] Section IV states that 'a total of ten simulations were performed for all experiments ... and the average values were calculated,' but Section V.M states that 'the reported accuracies are single-run point estimates' with no significance testing. These statements contradict each other and the issue is load-bearing: Table V shows WiFuse exceeding ResNet–BiGRU by only 0.46 pp in accuracy, a margin plausibly within run-to-run variation; the same concern applies to the 3.06 pp Wi-MIR gain over CSI-IRNet. The manuscript should report mean and standard deviation over multiple seeds (or otherwise reconcile the ten-simulation claim), state how significance was assessed, and indicate whether the differences in Table V are statistically reliable. Without this, the claimed superiority over the closest baselines is not supported.
  3. [§II.B–II.C, Fig. 4] The Delay–Doppler stream is a core contribution, but its validity rests on the assumption that phase sanitization (unwrapping, centralization, linear regression, reconstruction) removes hardware-induced offsets and random distortions sufficiently that the residual phase is dominated by motion-induced Doppler. That assumption is imported from prior work and is not validated on XRF55 or Wi-MIR. The ablation shows that concatenating the two streams improves over either stream alone, but it does not isolate whether the benefit comes from clean Doppler physics or simply from providing the network with additional learned features. A validation of the sanitized phase (e.g., residual phase statistics, Doppler SNR, or an oracle experiment) would strengthen the fusion claim and rule out the possibility that the Delay–Doppler stream is largely adding noise that the network happens to ignore or over
minor comments (5)
  1. [Table IV] The table note mentions an 'Impact (%)' column, but no such column appears in the table; remove the note or add the column.
  2. [§II.C] The notation uses N for both the packet count and the number of Doppler bins; Table II uses T for the time dimension. Please make the notation consistent and clarify the resampling step that fixes N=1000.
  3. [Tables VI and VII] Wi-MIR is described as a single indoor environment, yet the domain-adaptation and cross-domain tables list 'Scene 1–4' for Wi-MIR. Clarify what the four rows represent (repeated runs, different train/test partitions, or something else), and align the row labels with the dataset description.
  4. [§IV.H] The explanation of the 5.39% gap between the authors' ResNet and the reference CSI-IRNet, both nominally ResNet-based, is somewhat speculative ('reflects the limited reproducibility of the reference'). A brief statement of the implementation differences would be more informative.
  5. [General] There are minor typographical inconsistencies (e.g., 'Wi-MiR' in Section IV, 'Scenario' vs. 'Scene') and some references listed in Table I are not discussed in the related-work narrative. A final copyedit would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: WiFuse's results rest on public datasets and external baselines; the only self-citation (IBIS) is a balanced comparison, not load-bearing support.

full rationale

The claimed derivation chain is self-contained. The dual-stream input is computed by explicit signal-processing equations—amplitude via Eq. (3) and Delay–Doppler via Eq. (4) from the sanitized CSI—and the network architecture is fully specified in Table II with training described in Section III.C. The headline accuracies are measured on public XRF55 and Wi-MIR datasets against published external baselines (DML [4], CSI-IRNet [3]) rather than against any parameter fitted to the target results. The ablations in Section IV.F compare alternative inputs, preprocessing variants, and training regimes, and none of the reported 'gains' is defined in terms of a fitted value or a renamed prediction. The only self-citation is to the authors' prior IBIS [13], used in Section IV.K as an additional comparison baseline; that comparison is even-handed—IBIS wins some domain-adaptation rows in Table IX—so it is not load-bearing evidence for the central claim. The zero-shot cross-domain experiment is explicitly reported as a diagnostic negative result, further showing that the authors do not rename favorable outcomes as predictions. The paper's own stated limitations—the sample-wise 80/20 split versus the official 7:3 protocol (Section IV.D) and single-run point estimates with small margins (Section V.M)—are methodological validity concerns about protocol matching and statistical significance, not circularity: none of the equations or claims reduces to its own input. Therefore, the derivation is not circular.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities, forces, or conserved quantities. The dual-stream Delay-Doppler representation is a feature-engineering construct, not a postulated entity.

free parameters (6)
  • Moving-average smoothing window = 30 packets (150 ms)
    Applied to the amplitude stream along the packet axis; chosen to suppress rapid fluctuations while preserving motion envelopes; no sensitivity analysis in the paper (§II-B).
  • Doppler resampling length = 1000 bins / 5 s effective observation
    Delay-Doppler spectrum and amplitude stream are resampled to a fixed T=1000; sets Doppler resolution 0.2 Hz and velocity resolution ≈5.3 mm/s (§II-C).
  • Per-sample z-score normalization = per-sample mean and std
    Normalization strategy for both streams before concatenation; affects feature scaling and generalization; chosen without comparison to dataset-level normalization (§II-B).
  • 80/20 sample-wise train/test split = 80% train / 20% test
    Diverges from the official XRF55 7:3 protocol; the paper acknowledges the same subjects may appear in both train and test, so direct comparison with DML is not matched (§IV-D).
  • Stage-2 layer-specific learning rates = 2e-6 / 2e-5 / 2e-4
    Fine-tuning learning rates for pre-trained CNN layers, channel-attention layers, and new TCN/classifier layers; chosen without reported sensitivity analysis (§III-C).
  • Data augmentation strengths = Gaussian noise 0.005/0.008, time masking
    Regularization schedule for ResNet and ResNet-TCN stages; picked by hand; no ablation for augmentation strength (§III-C, Table III).
assumptions (5)
  • domain assumption Phase sanitization via unwrapping, index centralization, linear regression, and reconstruction removes hardware phase offsets and random distortions, leaving motion-induced phase.
    Imported from prior work [17], [18] in §II-B; central to the Delay-Doppler stream. If residual phase noise dominates, the Delay-Doppler stream is not a motion representation.
  • standard math The 2D-FFT (IFFT over subcarriers, FFT over packets) maps CSI to a Delay-Doppler domain where human motion appears as compact, resolvable peaks.
    Standard OTFS/signal-processing result cited via [20], [21]; used to justify Eq. (4).
  • domain assumption After digital resampling, the Delay-Doppler grid can be reshaped to F=270 features and concatenated with the amplitude stream along a shared spatial-frequency axis.
    The paper does not specify how M delay bins map to N_sc subcarriers when forming F=270 features; if the axes are incompatible, the feature-level concatenation is ill-posed (§II-B).
  • domain assumption The public datasets XRF55 and Wi-MIR are correctly labeled and preprocessed; Wi-MIR's released data omits the steady-state class as stated.
    All reported accuracies depend on dataset integrity; the paper notes the omitted class in §IV-H.
  • standard math Residual networks, dilated causal convolutions, attention, and transfer learning behave as described in the cited literature.
    Architectural components are taken from prior work [22]–[26]; no new theory is introduced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WiFuse: An Attention Mechanism for Human Activity Recognition using Fused CSI Amplitude and Delay-Doppler Channel Features." pith.science (2026). https://pith.science/paper/VCJH6CDO

@misc{pith2026260800642,
  author       = {Pith},
  title        = {Pith review of: WiFuse: An Attention Mechanism for Human Activity Recognition using Fused CSI Amplitude and Delay-Doppler Channel Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VCJH6CDO}},
  note         = {Machine review of arXiv:2608.00642}
}
read the original abstract

Recently, Wi-Fi sensing has played a significant role in Human Activity Recognition (HAR), as it enables the detection of various activities using only Wi-Fi signals, ensuring privacy and remaining non-intrusive for the user. However, environmental characteristics such as reflective surfaces, hardware offsets, and other physical impairments affect recognition by the neural network, subsequently causing errors and significantly reducing model accuracy. To overcome this problem we present the WiFuse framework, a dual-stream Channel State Information (CSI) framework for human activity recognition (HAR) that pairs denoised time-domain amplitude variations with 2D-FFT-derived Delay-Doppler motion representations computed from the sanitized channel phase. The fused representation feeds a hybrid ResNet-Temporal Convolutional Network (TCN) neural architecture augmented with channel and spatio-temporal attention, where the ResNet extracts spatial-spectral features and the TCN models long-range temporal dependencies; a decoupled two-stage transfer learning strategy is employed to improve optimization stability and feature reuse. We conduct extensive experiments on two public datasets, including comparisons against state-of-the-art methods and alternative hybrid architectures, ablation studies, and cross-dataset and domain-adaptation evaluations. The proposed framework reaches an overall accuracy of up to 95.28% across the four environments of the XRF55 dataset and up to 98.20% on the multi-user Wi-MIR dataset. Overall, the results indicate that combining amplitude and Delay-Doppler representations within a dual-stream strategy, enhanced by transfer learning, improves recognition performance under conditions that typically degrade deep neural networks, such as class overlap, multipath propagation, noise, and interference.

Figures

Figures reproduced from arXiv: 2608.00642 by the authors.

Figure 1
Figure 1. Illustration of the WiFuse framework showing the preprocessing stages used for feature extraction, as well as the [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Performance evaluation across multiple environments. The confusion matrices validate the model’s robustness in (a) [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Classification accuracy: comparison between the pro [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparative performance graph considering multi [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Confusion matrix representations for ResNet and hybrid [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Confusion matrix representations for the original CSI [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 6 linked inside Pith

  1. [1]

    A tutorial-cum-survey on self-supervised learning for wi-fi sensing: Trends, challenges, and outlook,

    A. Y . Radwan, M. Yildirim, N. Hasanzadeh, H. Tabassum, and S. Valaee, “A tutorial-cum-survey on self-supervised learning for wi-fi sensing: Trends, challenges, and outlook,”IEEE Communications Surveys & Tutorials, vol. 28, p. 2345–2376, 2026. [Online]. Available: http://dx.doi.org/10.1109/COMST.2025.3586212

    arXiv 2026

  2. [2]

    Analysis and research on performance of sleep safety monitoring system based on wifi sensing,

    Y . Wang, H. Yang, and J. Wang, “Analysis and research on performance of sleep safety monitoring system based on wifi sensing,” in2022 IEEE 4th Eurasia Conference on IOT, Communication and Engineering (ECICE), 2022, pp. 144–147

    2022

  3. [3]

    Wi- mir: A csi dataset for wi-fi based multi-person interaction recognition,

    M. Shafiqul Islam, M. Humayun Kabir, M. Ali Hasan, and W. Shin, “Wi- mir: A csi dataset for wi-fi based multi-person interaction recognition,” IEEE Access, vol. 12, pp. 67 256–67 272, 2024

    2024

  4. [4]

    Xrf55: A radio frequency dataset for human indoor action analysis,

    F. Wang, Y . Lv, M. Zhu, H. Ding, and J. Han, “Xrf55: A radio frequency dataset for human indoor action analysis,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 8, 2024

    2024

  5. [5]

    Two-stream convolution augmented transformer for human activity recognition,

    B. Li, W. Cui, W. Wang, L. Zhang, Z. Chen, and M. Wu, “Two-stream convolution augmented transformer for human activity recognition,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, 2021, pp. 286–293

    2021

  6. [6]

    Wimans: A benchmark dataset for wifi- based multi-user activity sensing,

    S. Huang, K. Li, D. You, Y . Chen, A. Lin, S. Liu, X. Li, and J. A. McCann, “Wimans: A benchmark dataset for wifi- based multi-user activity sensing,” 2024. [Online]. Available: https: //arxiv.org/abs/2402.09430

    arXiv 2024

  7. [7]

    Moric: Csi delay-doppler decomposition for robust wi-fi-based human activity recognition,

    N. Hasanzadeh and S. Valaee, “Moric: Csi delay-doppler decomposition for robust wi-fi-based human activity recognition,” 2025. [Online]. Available: https://arxiv.org/abs/2506.12997

    arXiv 2025

  8. [8]

    Dorf: Doppler radiance fields for robust human activity recognition using wi-fi,

    Navid Hasanzadeh and Shahrokh Valaee, “Dorf: Doppler radiance fields for robust human activity recognition using wi-fi,” 2025. [Online]. Available: https://arxiv.org/abs/2507.12132

    arXiv 2025

Show all 36 references
  1. [9]

    Towards siso bistatic sensing for isac,

    Z. Wang, J. A. Zhang, K. Wu, M. Xu, and Y . J. Guo, “Towards siso bistatic sensing for isac,” 2025. [Online]. Available: https://arxiv.org/abs/2508.12614

    2025 arXiv

  2. [10]

    Dual-stream contrastive learning for channel state information based human activity recognition,

    K. Xu, J. Wang, L. Zhang, H. Zhu, and D. Zheng, “Dual-stream contrastive learning for channel state information based human activity recognition,”IEEE Journal of Biomedical and Health Informatics, vol. 27, no. 1, pp. 329–338, 2023

    2023

  3. [11]

    Csi-glstn: A location-independent csi human activity recognition method based on spatio-temporal and channel feature fusion,

    Y . Zhang, D. Wu, Y . Wang, Y . Zhang, G. Ji, and J. Ai, “Csi-glstn: A location-independent csi human activity recognition method based on spatio-temporal and channel feature fusion,”IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1–12, 2024

    2024

  4. [12]

    Wifi- based cross-domain gesture recognition using attention mechanism,

    R. Liu, C. Pan, J. Zeng, H. Ren, K. Wang, L. Kong, and J. Wang, “Wifi- based cross-domain gesture recognition using attention mechanism,”

  5. [13]

    Ibis: A hybrid inception-bilstm and svm ensemble for robust doppler-based human activity recognition,

    A. M. Fernandes, H. I. D. Monego, B. S. Chang, A. Munaretto, H. M. Fontes, and R. L. Campos, “Ibis: A hybrid inception-bilstm and svm ensemble for robust doppler-based human activity recognition,” 2026. [Online]. Available: https://arxiv.org/abs/2510.24936

    2026 arXiv

  6. [14]

    Human activity recognition via score level fusion of Wi-Fi CSI signals,

    G. Lim, B. Oh, D. Kim, and K.-A. Toh, “Human activity recognition via score level fusion of Wi-Fi CSI signals,”Sensors, vol. 23, no. 16, p. 7292, 2023

    2023

  7. [15]

    Enhanced human activity recognition using Wi-Fi sensing: Leveraging phase and amplitude with attention mechanisms,

    T. D. Quy, C.-Y . Lin, and T. K. Shih, “Enhanced human activity recognition using Wi-Fi sensing: Leveraging phase and amplitude with attention mechanisms,”Sensors, vol. 25, no. 4, p. 1038, 2025

    2025

  8. [16]

    From RSSI to CSI: Indoor localization via channel response,

    Z. Yang, Z. Zhou, and Y . Liu, “From RSSI to CSI: Indoor localization via channel response,”ACM Computing Surveys, vol. 46, no. 2, pp. 25:1–25:32, Nov. 2013

    2013

  9. [17]

    Indoor localization with irregular antenna deployment,

    Y . Zheng, J. Liu, M. Sheng, and J. Li, “Indoor localization with irregular antenna deployment,” in2017 IEEE 86th Vehicular Technology Conference (VTC-Fall), 2017, pp. 1–5

    2017

  10. [18]

    Channel phase processing in wireless networks for human activity recognition,

    G. Diaz, I. Sobron, I. Eizmendi, I. Landa, J. Coyote, and M. Velez, “Channel phase processing in wireless networks for human activity recognition,”Internet of Things, vol. 24, p. 100960, Dec. 2023. [Online]. Available: http://dx.doi.org/10.1016/j.iot.2023.100960

    2023

  11. [19]

    A device-free indoor localization method using csi with wi-fi signals,

    X. Dang, X. Tang, Z. Hao, and Y . Liu, “A device-free indoor localization method using csi with wi-fi signals,”Sensors, vol. 19, no. 14, 2019. [Online]. Available: https://www.mdpi.com/1424-8220/19/14/3233

    2019

  12. [20]

    Delay doppler transform,

    X.-G. Xia, “Delay doppler transform,”IEEE Wireless Communications Letters, vol. 13, no. 6, pp. 1636–1639, 2024

    2024

  13. [21]

    Orthogonal time frequency space modu- lation,

    R. Hadani, S. Rakib, M. Tsatsanis, A. Monk, A. J. Goldsmith, A. F. Molisch, and R. Calderbank, “Orthogonal time frequency space modu- lation,” in2017 IEEE Wireless Communications and Networking Con- ference (WCNC), 2017, pp. 1–6

    2017

  14. [22]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  15. [23]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,”arXiv preprint arXiv:1803.01271, 2018

    2018 arXiv

  16. [24]

    Temporal convolutional networks for action segmentation and detection,

    C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager, “Temporal convolutional networks for action segmentation and detection,” in2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1003–1012

    2017

  17. [25]

    CBAM: Convolutional block attention module,

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “CBAM: Convolutional block attention module,” inProceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 3–19

    2018

  18. [26]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  19. [27]

    Tool release: Gath- ering 802.11n traces with channel state information,

    D. Halperin, W. Hu, A. Sheth, and D. Wetherall, “Tool release: Gath- ering 802.11n traces with channel state information,”ACM SIGCOMM Computer Communication Review, vol. 41, no. 1, p. 53, 2011

    2011

  20. [28]

    Learning phrase representations using RNN encoder–decoder for statistical machine translation,

    K. Cho, B. van Merri ¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using RNN encoder–decoder for statistical machine translation,” inProceed- ings of the 2014 Conference on Empirical Methods in Natural Language Pro...

    2014

  21. [29]

    Deep transfer learning for detection of upper and lower body move- ments: Transformer with convolutional neural network,

    K. Lacroix, D. Gholamiangonabadi, A. Luisa Trejos, and K. Grolinger, “Deep transfer learning for detection of upper and lower body move- ments: Transformer with convolutional neural network,”IEEE Sensors Journal, vol. 24, no. 20, pp. 33 778–33 790, 2024

    2024

  22. [30]

    K-core based temporal graph convolutional network for dynamic graphs,

    J. Liu, C. Xu, C. Yin, W. Wu, and Y . Song, “K-core based temporal graph convolutional network for dynamic graphs,”IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 8, pp. 3841–3853, 2022

    2022

  23. [31]

    Going deeper with convolutions,

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 1–9

    2015

  24. [32]

    Domain adap- tation: Challenges, methods, datasets, and applications,

    P. Singhal, R. Walambe, S. Ramanna, and K. Kotecha, “Domain adap- tation: Challenges, methods, datasets, and applications,”IEEE Access, vol. 11, pp. 6973–7020, 2023

    2023

  25. [33]

    Transferability assessment of wifi sensing through csi-based doppler,

    J. R. Merino Bernaola, I. Sobron, I. Landa, I. n. Eizmendi, and M. V´elez, “Transferability assessment of wifi sensing through csi-based doppler,” inProceedings of the 14th International Conference on Information Communication and Applications, ser. ICICA ’25. New York, NY , U...

    2026

  26. [34]

    A comparison of methods for multiclass support vector machines,

    C.-W. Hsu and C.-J. Lin, “A comparison of methods for multiclass support vector machines,”IEEE Transactions on Neural Networks, vol. 13, no. 2, pp. 415–425, 2002

    2002

  27. [35]

    A study of the bilstm model based on woa optimized attention mechanism for power load forecasting,

    X. Li and X. Zheng, “A study of the bilstm model based on woa optimized attention mechanism for power load forecasting,” in2023 IEEE International Conference on Sensors, Electronics and Computer Engineering (ICSECE), 2023, pp. 237–242

    2023

  28. [2025]

    Available: https://arxiv.org/abs/2512.04521

    [Online]. Available: https://arxiv.org/abs/2512.04521

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.