Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Complying with the EU AI Act: Innovations in Explainable and User-Centric Hand Gesture Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Radar-based XentricAI claims per-user VAE thresholds flag 11.5% more anomalous gestures, SHAP explains 97.5% of them, and experience replay lifts user-calibrated accuracy from 69.56% to at least 84.73%, a design template for EU AI Act…

desk verdict A worthwhile systems paper with a strong calibration study and a useful public dataset, but the headline 97.5% explainability claim is not backed by the experiment they actually ran. read the letter →

arxiv 2503.15528 v1 pith:6TSWHEA4 submitted 2025-02-04 cs.HC cs.AI

Sarah Seifi , Tobias Sukianto , Cecilia Carbonelli , Lorenzo Servadei , Robert Wille This is my paper · ORCID
classification cs.HCcs.AI
keywords handgesturerecognitionexplainableAI(XAI)variationalautoencoderanomalydetectiontransferlearningexperiencereplayFMCWradarEUAct
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 hand-gesture recognition system used at the edge—XentricAI, built on a small gated-recurrent-unit classifier fed by five physically meaningful radar features—can be made simultaneously more accurate per user and more transparent, and that this is what the EU AI Act's transparency and robustness demands will look like in practice. Its three headline numbers are: a variational autoencoder with per-user thresholds raises correctly flagged anomalous gestures from 57.60% to 69.10%; 97.5% of flagged anomalies show feature-importance deviations that the system translates into feedback about speed or wrist execution; and experience-replay transfer learning raises mean user-calibrated gesture accuracy from a 69.56% baseline to at least 84.73%. The paper also contributes a public dataset extension of 28,000 gestures, of which 24,000 are deliberately anomalous, and maps each design module onto a table of EU AI Act compliance measures. If these results hold, the paper offers a concrete template for turning the Act's abstract obligations into measured engineering choices.

What carries the argument

The load-bearing mechanism is a four-module pipeline around a compact GRU classifier: experience-replay transfer learning for user calibration; a variational autoencoder whose reconstruction error, $e_{\mathrm{rec},m} = \|X_m - \hat{X}_m\|_2^2$, is thresholded per user at the 90th percentile of calibration-gesture errors to flag anomalies; and a SHAP-based characterization stage that builds a nominal reference envelope—per-feature minimum, maximum, and median thresholds plus the slopes between consecutive features—from nominal gestures and then diagnoses a flagged gesture by which feature deviates. The SHAP values are computed window-by-window because the explanation method cannot handle three-dimensional multivariate time series directly; the paper treats the resulting deviations (elevated Doppler for fast gestures, depressed Doppler for slow ones, reduced range for wrist executions) as the physical cause of the anomaly. The variational autoencoder is introduced as the first use of user-specific dynamic thresholding in FMCW radar-based hand gesture recognition, and the reconstruction-error threshold is what converts an unsupervised anomaly score into a per-user decision boundary.

What would settle it

Run the characterization stage on the 24,000 anomalous gestures with known labels and compare each system diagnosis (e.g., 'Doppler above the nominal upper limit' means too fast, 'Doppler below nominal' means too slow, 'range below nominal' means wrist execution) against the recorded ground-truth anomaly type. If per-class agreement does not clearly exceed the one-in-three chance level, the 97.5% deviation-detection rate is not a characterization success rate. A complementary check: feed the 4,000 newly collected nominal gestures through the VAE threshold and count false flags, since the paper reports only the increase in correctly flagged anomalies, not precision.

Watch

Extended reading notes

Core claim

On the paper's own terms, XentricAI's contribution is an integrated answer to three failures of a plain gesture classifier: it cannot tell a user why it failed, it does not adapt to a new user's way of moving, and it silently misclassifies or ignores gestures that fall outside its training distribution. The VAE module, trained on nominal gestures and thresholded at the 90th percentile of each user's calibration errors, catches the misclassified and unpredicted gestures that the previous hardcoded rules missed, lifting correctly flagged anomalies from 57.60% to 69.10%. The SHAP characterization stage then compares each flagged gesture's feature-importance values against a reference envelope built from nominal gestures and diagnoses the cause—elevated Doppler readings indicate a gesture performed too fast, depressed Doppler a too-slow gesture, reduced range a wrist execution—reporting a 97.5% success rate at identifying deviations. Meanwhile, experience-replay calibration, which mixes a subset of the original training data with the user's own recordings during retraining, raises average gesture accuracy from 69.56% to at least 84.73% while keeping performance on the original training data near 90%, in contrast to 59.93% to 85.39% without replay. The authors present the whole pipeline as a commercially viable, single-sensor system whose documented design choices map onto the EU AI Act's risk-management, data-governance, transparency, human-oversight, and robustness articles.

Load-bearing premise

The 97.5% characterization claim assumes that a SHAP deviation outside the nominal envelope reveals the true physical cause of an anomaly (too fast, too slow, or wrist execution); the paper checks that deviations occur but never verifies the diagnosis against the known anomaly labels, and it itself warns that SHAP is 'true to the model rather than true to the data.'

Editorial extensions

If this is right

  • A gesture system can flag its own failures and tell the user how to fix them: per-user VAE thresholds catch gestures the old rule-based flags missed, and SHAP deviations translate into concrete advice such as 'perform the gesture more slowly' or 'move closer to the radar'.
  • User calibration does not have to trade away prior knowledge: with experience replay, per-user accuracy rises to at least 84.73% on average while accuracy on the original training data stays near 90%, so a modest calibration effort (10 gestures per class) is enough.
  • The publicly extended dataset—24,000 labeled anomalous gestures (fast, slow, wrist) from eight users plus 4,000 nominal gestures from four new users—gives the community a reproducible benchmark for out-of-distribution testing in radar-based gesture recognition.
  • The explicit mapping of design modules to EU AI Act articles (Art. 9–15) shows one route for translating abstract compliance obligations into concrete, measurable engineering choices while formal standards are still under development.

Reading between the lines

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

  • The 97.5% figure counts how often flagged gestures show SHAP deviations outside the nominal envelope, not how often the system's stated cause matches the ground-truth anomaly label; the paper's own definition and examples make it an open question whether 'characterized' implies 'correctly diagnosed'.
  • Because SHAP is 'true to the model rather than true to the data', the characterization module inherits the GRU's biases: if the classifier has learned spurious feature relationships, the feedback it generates can point at the wrong physical cause even when the deviation detection is perfect.
  • The paper reports recall-side gains only: it does not state how often nominal gestures are wrongly flagged by the per-user VAE threshold, so the 11.50% improvement in anomaly detection is not paired with a precision or false-positive rate.
  • A natural testable extension would be to run the same VAE-plus-SHAP pipeline on other single-sensor time-series interfaces—vital-sign or activity monitoring with the same radar chip—where per-user thresholds and cause feedback are equally relevant; the EU AI Act mapping is generic, but the quantitative claims would have to be re-measured.
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 paper extends the XentricAI hand-gesture-recognition framework with three main contributions: a VAE-based anomaly-detection module with user-specific thresholding, an extension of an FMCW-radar gesture dataset with 24,000 out-of-distribution anomalous gestures and 4,000 additional nominal gestures, and an SHAP-based gesture-characterization module. The authors report that the VAE module increases the fraction of correctly flagged anomalous gestures from 57.60% to 69.10% (an 11.50% increase), that 97.5% of flagged anomalies are characterized via SHAP threshold deviations, and that experience-replay transfer learning raises mean user-calibrated gesture accuracy from a 69.56% baseline to at least 84.73% (a 15.17% gain). The paper also maps XentricAI's features onto EU AI Act requirements. The calibration experiments are reported extensively with six-run averages, but the two most novel quantitative claims—the VAE improvement and the 97.5% characterization rate—lack the statistical grounding and ground-truth validation that the rest of the experimental sections provide.

Significance. If the characterization claim were properly validated, the paper would make a useful contribution: it provides a publicly available radar-gesture dataset with 24,000 labeled anomalous gestures, demonstrates a systematic comparison of anomaly-detection methods for FMCW-radar HGR, and offers a detailed transfer-learning study with experience replay, EWC, and SI. The emphasis on user-specific thresholding and on explainability as a route to EU AI Act alignment is timely. However, the central explainability result—that SHAP deviations identify the physical cause of an anomaly—is currently only shown for a few hand-picked examples; the 97.5% figure is computed by checking whether any SHAP value falls outside a per-feature envelope, without a control condition or a comparison to the known fast/slow/wrist labels. The VAE improvement is also reported as a single number without variance or significance testing. These issues are load-bearing because the abstract and conclusion advertise the 97.5% characterization rate and the 11.50% detection gain as the main novelties. The transfer-learning contribution, by contrast, is supported by detailed tables and appears sound.

major comments (3)
  1. [Section 5.3 and Section 3.4.2, Eqs. (6)-(9)] The 97.5% gesture-characterization rate is not validated against the known anomaly labels. The paper states in Section 5.2 that "By knowing the actual types of anomalies, the results of the anomaly reasoning provided by XentricAI were compared against the ground truth," but Section 5.3 reports only that deviations occurred ("The thresholding mechanism successfully identified deviations in 97.5% of anomalously flagged gestures"). No confusion matrix, per-anomaly-type accuracy, or comparison of the predicted cause (fast, slow, wrist) with the actual cause is given. Figure 14 shows illustrative examples only. Since the paper itself cautions that SHAP is "true to the model rather than true to the data," the deviation rate cannot be equated with correct causal characterization. The authors should report the agreement between the system's assigned characterization and the ground-truth anomaly type, ideally per anomaly class and per user.
  2. [Section 5.3, Eqs. (6)-(7)] The SRV envelope construction likely over-flags nominal gestures, which undermines the interpretation of the 97.5% figure. The envelope is defined as the minimum and maximum SHAP values over only 10 nominal gestures per class and per feature (Eqs. 6 and 7). With D=5 features, the union of per-feature "outside the envelope" events can occur for a nominal gesture with substantial probability purely by chance, especially since the min/max of a small sample underestimates the true support of the distribution. No control experiment is reported that passes nominal gestures through the characterization pipeline to measure the false-positive rate. The authors should report what fraction of nominal gestures trigger at least one feature deviation under this envelope; without that baseline, 97.5% is not interpretable as a characterization success rate.
  3. [Section 5.2, Table 5] The headline 11.50% improvement in correctly flagged gestures (57.60% to 69.10%) is presented as an aggregate with no error bars, standard deviation, or significance test, despite the text stating that each user's results were averaged over six experimental runs. Given that the VAE's advantage over GAN in Appendix B is only 1.40 percentage points, it is essential to know whether the difference is within run-to-run variability. The authors should provide per-user breakdowns and a significance test (e.g., paired t-test or Wilcoxon signed-rank test across users) for the VAE versus the condition-flagged baseline and versus GAN. Additionally, the denominator for "correctly flagged gestures" should be defined explicitly: is it all anomalous gestures, or only those that the GRU model misclassified or failed to predict?
minor comments (5)
  1. [Abstract] The abstract contains a typo: "sucess rate" should be "success rate."
  2. [Section 1 and throughout] The manuscript contains several typographical errors that should be corrected, including "adresses" (Introduction), "Suport vector machines" (Section 3.1.1), and "data distribution shits" (Section 5.4).
  3. [Section 5.1.1, paragraph after Table 3] The text refers to "Table 5.1.1" when discussing retraining configurations; the actual table is numbered "Table 3." The cross-reference should be fixed, and the same applies to other internal references such as "Table Appendix C" and "Table B" in Appendices B and C.
  4. [Section 4.4, Eq. (14)] In Eq. (14), the indicator function I(·) is used before it is defined; the sentence defining I(·) appears after the equation and should be moved before it. Also, the variables sm and em are defined after being used in the equation; reordering would improve readability.
  5. [Section 5.4 and Table 7] The claim in the abstract and conclusion that XentricAI offers "a commercially viable solution" is not supported by any cost, energy, latency, or deployment analysis in the paper. Consider softening this claim or adding a brief discussion of edge-deployment feasibility.

Circularity Check

1 steps flagged · score 6.0 of 10

The 97.5% gesture-characterization claim is the method's own decision rule: a gesture is counted as characterized exactly when its SHAP values deviate from the SRV envelope, with no validation against known anomaly causes.

  1. self definitional [Section 5.3 (Gesture Characterization Results) and Section 3.4.2, Eqs. (6)-(7)]
    "For each user, the method's effectiveness in characterizing gestures performed at slow and fast speeds, as well as wrist executions, was assessed. The analysis focused on how many flagged anomalous gestures exhibited deviations in feature rankings that enabled characterization, and how well these deviations contributed to understanding the anomalies. ... The thresholding mechanism successfully identified deviations in 97.5% of anomalously flagged gestures, indicating a well-functioning system."

    The paper defines "gesture characterization" as reasoning about the potential cause of an anomalous gesture, and implements it by comparing anomalous SHAP values to SHAP reference values (SRVs), where the SRVs are computed from nominal gestures as per-feature min/max bounds (Eqs. 6-7). The reported 97.5% success rate counts only whether a flagged anomalous gesture's SHAP values deviated from those SRV bounds. Since "characterized" is operationalized as "deviation detected," the success metric is the method's own decision rule evaluated on itself. The paper states that ground-truth anomaly types were known and compared, but no per-cause accuracy, confusion matrix, or precision/recall for fast/slow/wrist identification is provided.

full rationale

The transfer learning and VAE anomaly-detection claims are empirically grounded: the 15.17% improvement compares held-out user calibration accuracy against a baseline, and the 11.50% increase in correctly flagged gestures compares condition-flagged and VAE-flagged detection on anomalous gestures. Those results do not reduce to their inputs by construction. The central circularity is confined to the gesture-characterization result. The SRV envelope (Eqs. 6-7) is fitted to nominal SHAP values, and the characterization phase declares success when an anomalous gesture's SHAP values fall outside that envelope. Because the paper's aggregate success metric is defined as the occurrence of those deviations, the 97.5% rate is equivalent to the method's own trigger condition. No independent validation against the known fast/slow/wrist labels is reported, so the claim that 97.5% of anomalies are "accurately characterized" is not supported by the presented evidence. This weakens the explainability and EU AI Act alignment arguments, but the calibration and VAE detection contributions remain independently defensible. Score 6 reflects partial circularity: one central quantitative claim reduces to its operational definition, while other major claims are self-contained. No load-bearing self-citation or uniqueness-importation issue was found.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rely on several hand-chosen hyperparameters (90th percentile threshold, Lwindow=22, preprocessing amplitude threshold), inherited design choices from prior work, and an unvalidated causal interpretation of SHAP deviations. No new physical entities are introduced.

free parameters (4)
  • VAE user-specific threshold percentile = 90th percentile
    Threshold for flagging anomalies is set at the 90th percentile of calibration reconstruction errors; no sensitivity analysis is reported.
  • SHAP window length Lwindow = 22 frames
    Hand-chosen to balance context against recency; affects all SHAP explanations and no ablation is reported.
  • Radar amplitude threshold for label refinement = not reported
    Called an 'empirically determined threshold' in Section 4.3; affects which frames are labeled as gestures.
  • Operating point ntrain-nuser,i for anomaly experiments = 50 train, 10 user per class
    Selected for data minimization and because performance plateaus; not justified with significance tests.
assumptions (5)
  • domain assumption The closest moving target to the radar is the user's hand.
    Used in Section 4.3 to select the range bin for feature extraction; if false, all five input features are corrupted.
  • domain assumption The GRU architecture and feature set from Strobel et al. (2024) are adequate for this task.
    Adopted without re-derivation in Section 3.1.2; all results inherit this design choice.
  • domain assumption Reconstruction error from a VAE trained only on nominal gestures is a valid anomaly score.
    Section 3.3.2 and 4.5 assume OOD gestures yield high reconstruction error, with no theoretical guarantee for this data distribution.
  • standard math SHAP values satisfy local accuracy, missingness, and consistency and are meaningful for a GRU with windowed inputs.
    Section 3.4.2 relies on Shapley axioms, though SHAP is an approximation and is true-to-model rather than true-to-data, as the authors note.
  • domain assumption The paper's mapping of abstract EU AI Act obligations to technical measures is a valid compliance interpretation.
    Section 5.4 asserts compliance through a self-assessed mapping, not through certification or external audit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Complying with the EU AI Act: Innovations in Explainable and User-Centric Hand Gesture Recognition." pith.science (2026). https://pith.science/paper/6TSWHEA4

@misc{pith2026250315528,
  author       = {Pith},
  title        = {Pith review of: Complying with the EU AI Act: Innovations in Explainable and User-Centric Hand Gesture Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6TSWHEA4}},
  note         = {Machine review of arXiv:2503.15528}
}
read the original abstract

The EU AI Act underscores the importance of transparency, user-centricity, and robustness in AI systems, particularly for high-risk systems. In response, we present advancements in XentricAI, an explainable hand gesture recognition (HGR) system designed to meet these regulatory requirements. XentricAI adresses fundamental challenges in HGR, such as the opacity of black-box models using explainable AI methods and the handling of distributional shifts in real-world data through transfer learning techniques. We extend an existing radar-based HGR dataset by adding 28,000 new gestures, with contributions from multiple users across varied locations, including 24,000 out-of-distribution gestures. Leveraging this real-world dataset, we enhance XentricAI's capabilities by integrating a variational autoencoder module for improved gesture anomaly detection, incorporating user-specific thresholding. This integration enables the identification of 11.50% more anomalous gestures. Our extensive evaluations demonstrate a 97.5% sucess rate in characterizing these anomalies, significantly improving system explainability. Furthermore, the implementation of transfer learning techniques has shown a substantial increase in user adaptability, with an average improvement of at least 15.17%. This work contributes to the development of trustworthy AI systems by providing both technical advancements and regulatory compliance, offering a commercially viable solution that aligns with the EU AI Act requirements.

Figures

Figures reproduced from arXiv: 2503.15528 by the authors.

Figure 1
Figure 1. Overview of the EU AI Act’s framework. A) Pyramid illustrating the risk-based approach of the [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Comparison of AI models. Explainable AI involves opening the black-box to identify the factors [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Overview of the XentricAI methodology. This comprehensive framework for HGR and analysis is [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Exemplary illustration of dataset allocation for model training, calibration, and assessment. (A) [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Windowing and preprocessing of gesture sequences for SHAP analysis. [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Flowchart illustrating the two-phase process for gesture characterization using SHAP values. The [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Illustration of a single chirp emitted by the radar device. (a) Time-domain representation of the [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: Data collection and processing pipeline. The 60 GHz FMCW radar captures gestures, producing [PITH_FULL_IMAGE:figures/full_fig_p028_8.png]
Figure 9
Figure 9. Figure 9: Lightweight preprocessing algorithm. A) Target detection for preprocessing: The range profile [PITH_FULL_IMAGE:figures/full_fig_p030_9.png]
Figure 10
Figure 10. Figure 10: Visualization of the dynamic gesture accuracy metric dg [PITH_FULL_IMAGE:figures/full_fig_p033_10.png]
Figure 11
Figure 11. Figure 11: Comparison of gesture accuracy and model stability with (blue) and without ER (green) for user [PITH_FULL_IMAGE:figures/full_fig_p037_11.png]
Figure 12
Figure 12. Figure 12: Error bar plot illustrating the mean performance of models across varying combinations of training [PITH_FULL_IMAGE:figures/full_fig_p040_12.png]
Figure 13
Figure 13. Figure 13: Error bar plot illustrating the mean performance of two gesture recognition models (simple setup [PITH_FULL_IMAGE:figures/full_fig_p041_13.png]
Figure 14
Figure 14. Figure 14: Gesture characterization of a SwipeLeft gesture. A) The gesture characterization of a nominal [PITH_FULL_IMAGE:figures/full_fig_p045_14.png]
Figure 15
Figure 15. Figure 15: Mapping of XentricAI’s conformance with the European AI Act’s principles, including trans [PITH_FULL_IMAGE:figures/full_fig_p048_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 17 canonical work pages

  1. [6]

    Exploiting transfer learning for personalized view invariant gesture recognition, in: 2014 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 1250–1254. Cˆ ot´ e-Allard, U., Fall, C.L., Drouin, A., Campeau-Lecours, A., Gosselin, C., Glette, K., Lavi- olette, F., Gosselin, B.,

    work page 2014

  2. [8]

    arXiv:1702.08608

    Towards a rigorous science of interpretable machine learning. arXiv:1702.08608. Fatemifar, S., Arashloo, S.R., Awais, M., Kittler, J.,

    arXiv

  3. [9]

    arXiv preprint arXiv:2009.13996

    Explainable ai without interpretable model. arXiv preprint arXiv:2009.13996 . Frasca, M., La Torre, D., Pravettoni, G., Cutica, I.,

    arXiv 2009

  4. [11]

    Radar gesture recognition system in presence of interference using self-attention neural network, in: 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), IEEE. pp. 1409–1414. Hedstr¨ om, A., Weber, L., Lapuschkin, S., H¨ ohne, M.,

    work page 2019

  5. [12]

    arXiv preprint arXiv:2401.06465

    Sanity checks revisited: An explo- ration to repair the model parameter randomisation test. arXiv preprint arXiv:2401.06465 . Huang, H., Liang, Z., Sun, F., Dong, M., et al.,

    arXiv

  6. [13]

    URL: https: //arxiv.org/abs/1412.6980, arXiv:1412.6980

    Adam: A method for stochastic optimization. URL: https: //arxiv.org/abs/1412.6980, arXiv:1412.6980. Kingma, D.P., Welling, M.,

    arXiv

  7. [16]

    URL: https://arxiv.org/abs/1705.07874, arXiv:1705.07874

    A unified approach to interpreting model predictions. URL: https://arxiv.org/abs/1705.07874, arXiv:1705.07874. Malysa, G.J., Ali, M., Wang, D., Netsch, L.P.,

    arXiv

  8. [18]

    https://eur-lex.europa.eu/eli/reg/2024/1689/oj

    Proposal for a regulation of the european parliament and of the council laying down harmonised rules on artificial intelli- gence (artificial intelligence act). https://eur-lex.europa.eu/eli/reg/2024/1689/oj. Accessed: 2024-09-03. Pruksachatkun, Y., Mcateer, M., Majumdar, S.,

    work page 2024

Show all 33 references
  1. [19]

    Hand gesture recognition using continuous wave (cw) radar based on hybrid pca-knn, in: 2021 IEEE Symposium on Wireless Technology & Applications (ISWTA), IEEE. pp. 88–92. Reinhardt, K.,

    2021

  2. [20]

    Seifi, S., Sukianto, T., Carbonelli, C., 2024a

    Detection of sensor-to- sensor variations using explainable ai , 1–6. Seifi, S., Sukianto, T., Carbonelli, C., 2024a. 60 ghz fmcw radar gesture dataset URL: https://dx.doi.org/10.21227/s12w-cc46, doi:10.21227/s12w-cc46. Seifi, S., Sukianto, T., Carbonelli, C., Servadei, L., Wi...

    doi

  3. [22]

    Gesture recognition for fmcw radar on the edge, in: 2024 IEEE Topical Conference on Wireless Sensors and Sensor Networks (WiSNeT), pp. 45–48. doi: 10.1109/WiSNeT59910.2024.10438579. Sundararajan, M., Taly, A., Yan, Q.,

    2024

  4. [23]

    arXiv preprint arXiv:2006.05714

    Optilime: Optimized lime explanations for diagnostic computer algorithms. arXiv preprint arXiv:2006.05714 . Vollert, S., Atzmueller, M., Theissler, A.,

    2006 arXiv

  5. [24]

    Interpretable machine learning: A brief survey from the predictive maintenance perspective, in: 2021 26th IEEE International Conference on Emerging Technologies and Factory Automation (ETF A ), pp. 01–08. doi: 10.1109/ ETFA45728.2021.9613467. Von Eschenbach, W.J.,

    2021

  6. [25]

    Gesture recognition for smart home applications using portable radar sensors, in: 2014 36th annual international conference of the IEEE engineering in medicine and biology society, IEEE. pp. 6414–6417. Wang, C., Zhao, X., Li, Z.,

    2014

  7. [26]

    IEEE Internet of Things Journal 10, 17680–17693

    Dcs-ctn: Subtle gesture recognition based on td-cnn- transformer via millimeter-wave radar. IEEE Internet of Things Journal 10, 17680–17693. Wang, F., Ao, X., Wu, M., Kawata, S., She, J., 2024a. Explainable deep learning for semg- based similar gesture recognition: A shapley-v...

    2022

  8. [27]

    ”do you trust me?”: Increasing user-trust by integrating virtual agents in explainable ai interaction design, in: Proceedings of the 19th ACM International Conference on Intelligent Virtual Agents, Association for Computing Machinery, New York, NY, USA. p. 7–9. URL: https://do...

  9. [28]

    IEEE Transactions on Geoscience and Remote Sensing 59, 4749–4764

    Multidimensional feature representation and learning for robust hand-gesture recognition on commercial millimeter-wave radar. IEEE Transactions on Geoscience and Remote Sensing 59, 4749–4764. doi: 10.1109/TGRS.2020. 3010880. Xiao, R., Liu, J., Han, J., Ren, K.,

    2020 doi

  10. [29]

    Gesture recognition for home automation using trans- fer learning, in: 2018 International Conference on Intelligent Informatics and Biomedical Sciences (ICIIBMS), IEEE. pp. 136–138. Ying, Z., Bourgeois, D., You, J., Zitnik, M., Leskovec, J.,

    2018

  11. [30]

    arXiv preprint arXiv:2111.10602

    Unsupervised domain adaptation for device-free gesture recognition. arXiv preprint arXiv:2111.10602 . Zhang, G., Lan, S., Zhang, K., Ye, L.,

    arXiv

  12. [31]

    Temporal-range-doppler features interpretation and recognition of hand gestures using mmw fmcw radar sensors, in: 2020 14th European conference on antennas and propagation (EuCAP), IEEE. pp. 1–4. Zhang, G., Zhang, K., Yun, Y., Lu, G., Lan, S.,

    2020

  13. [32]

    5 decision tree in human gesture recognition based on doppler radars, in: 2019 International Symposium on Antennas and Propagation (ISAP), IEEE

    Implementation of c4. 5 decision tree in human gesture recognition based on doppler radars, in: 2019 International Symposium on Antennas and Propagation (ISAP), IEEE. pp. 1–3. Zhu, G., Zhang, L., Shen, P., Song, J.,

    2019

  14. [33]

    URL: https://arxiv.org/abs/1911.02685, arXiv:1911.02685

    A comprehensive survey on transfer learning. URL: https://arxiv.org/abs/1911.02685, arXiv:1911.02685. 60 Appendix A. Quantitative Evaluation of Catastrophic Forgetting Mitigation Approaches To further validate the effectiveness of Experience Replay (ER) in mitigating catastrop...

    1911 arXiv

  15. [2000]

    Lof: identifying density-based local outliers, in: Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pp. 93–104. Bridle, J.,

    2000

  16. [2008]

    Isolation forest, in: 2008 eighth ieee international conference on data mining, IEEE. pp. 413–422. Liu, H., Cui, K., Hu, K., Wang, Y., Zhou, A., Liu, L., Ma, H., 2022a. mtranssee: Enabling environment-independent mmwave sensing based gesture recognition via transfer learning. ...

    2008

  17. [2014]

    arXiv preprint arXiv:1406.1078

    Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 . Choi, J.W., Park, C.W., Kim, J.H.,

    arXiv

  18. [2017]

    Breland, D.S., Dayal, A., Jha, A., Yalavarthy, P.K., Pandey, O.J., Cenkeramaddi, L.R.,

    arXiv preprint arXiv:1704.07911 . Breland, D.S., Dayal, A., Jha, A., Yalavarthy, P.K., Pandey, O.J., Cenkeramaddi, L.R.,

    arXiv

  19. [2018]

    Simulation of the radar cross-section of dynamic human motions using virtual reality data and ray tracing, in: 2018 IEEE Radar Conference (RadarConf18), IEEE. pp. 1555–1560. Skolnik, M.I., et al.,

    2018

  20. [2019]

    URL: https: //www.aepd.es/sites/default/files/2019-12/ai-ethics-guidelines.pdf

    Ethics guideline for trustworthy ai. URL: https: //www.aepd.es/sites/default/files/2019-12/ai-ethics-guidelines.pdf . Bach, S., Binder, A., Montavon, G., Klauschen, F., M¨ uller, K.R., Samek, W.,

    2019

  21. [2020]

    Chen, X., Wang, S., Fu, B., Long, M., Wang, J.,

    True to the model or true to the data? arXiv preprint arXiv:2006.16234 . Chen, X., Wang, S., Fu, B., Long, M., Wang, J.,

    2006 arXiv

  22. [2021]

    Electromyography signal-based gesture recognition for human-machine interaction in real-time through model calibration, in: Advances in Information and Communication: Proceedings of the 2021 Future of Information and Communication Conference (FICC), Volume 2, Springer. pp. 898...

    2021

  23. [2022]

    URL: https://arxiv

    Auto-encoding variational bayes. URL: https://arxiv. org/abs/1312.6114, arXiv:1312.6114. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.,

    arXiv

  24. [2023]

    1139–1150

    The role of explainable ai in the context of the ai act, in: Proceedings of the 2023 ACM conference on fairness, accountability, and transparency, pp. 1139–1150. 56 Parliament, T.E., the Council of the European Union,

    2023

  25. [2024]

    arXiv preprint arXiv:2410.07959

    Compl-ai framework: A technical interpretation and llm benchmarking suite for the eu artificial intelligence act. arXiv preprint arXiv:2410.07959 . Han, M., Yang, H., Ni, T., Duan, D., Ruan, M., Chen, Y., Zhang, J., Xu, W.,

    arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.