Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Efficient and Robust Spiking Neural Networks for sEMG-Based Muscle Fatigue Detection

T0 review · 3 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Quantized spiking networks detect muscle fatigue from sEMG at up to 201 imes lower energy while holding accuracy under noise.

desk verdict Solid applied SNN pipeline for sEMG fatigue with real F1/robustness numbers; the 201× energy headline is overstated relative to their own full-system sim. read the letter →

arxiv 2607.11065 v1 pith:JTAPXVYH submitted 2026-07-13 cs.NE

Kaiwen Tang , Jiaqi Dong , Zhanglu Yan , Weng-Fai Wong This is my paper
classification cs.NE
keywords SpikingNeuralNetworksMuscleFatigueDetectionSurfaceElectromyographyEdgeComputingQuantizationNoiseRobustnessWearableSystems
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

Muscle fatigue detection from surface electromyography is useful in sports, rehab, and safety monitoring, but ordinary deep models are too energy-hungry and data-hungry for continuous use on a watch or band. This paper shows that a compact spiking neural network, obtained by converting a small trained ANN and then quantizing weights to 3 or 4 bits, can match or beat strong baselines on two public sEMG datasets while cutting estimated energy by as much as two hundred times. A training objective called SDH (safe-haven activation quantization, defensive weight regularization, and Huber loss) keeps performance stable under seven kinds of noise, including feature-level models of motion artifacts and electrode drift. A lightweight post-filter (QAF) further cleans the label stream so the system can react quickly when fatigue appears. The result is a practical path to always-on, privacy-preserving fatigue monitoring on battery-powered wearables.

What carries the argument

SDH: an additive training loss that combines binary cross-entropy with SHAQ activation centering, defensive weight-orthogonality regularization, and Huber loss, keeping the model quantization-compatible and noise-resistant; together with average-spike-generation conversion and post-training low-bit weight quantization.

What would settle it

Measure the same models on raw sEMG streams recorded with genuine motion artifacts, electrode lift, and power-line interference, or run the quantized SNN on actual neuromorphic silicon and compare measured joules against the claimed 201× saving; a large drop in F1 or energy advantage would falsify the central claim.

Watch

Extended reading notes

Core claim

A four-layer fully connected network converted from ANN to rate-coded SNN, trained with the composite SDH loss and quantized to 3- or 4-bit weights, reaches QAF F1 scores of 87.93 % on SUE and 90.14 % on SPE—matching or exceeding the best non-spiking neural baselines—while remaining more stable across seven noise types and reducing estimated energy by up to 201.77× relative to the full-precision ANN.

Load-bearing premise

That the feature-level synthetic noises used in testing truly stand in for real wearable sEMG artifacts, and that the paper’s simple spike-rate energy model correctly predicts what a real chip would consume.

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 a compact quantized SNN pipeline for on-device sEMG muscle-fatigue detection. An ANN is trained with clamp-and-quantize activations and converted via average-spike-generation IF neurons; robustness is pursued with a composite SDH objective (SHAQ activation centering + defensive weight regularization + Huber loss); weights are post-training quantized to 3/4 bits; and a QAF post-processor (median filter + EMA adaptive threshold + fast channel) is applied. On the public SUE and SPE datasets the resulting SNN/QSN models report QAF F1 scores of 87.39–87.93 % (SUE) and 89.13–90.14 % (SPE), remain comparatively stable under seven feature-level noise types, and claim energy reductions of up to 201.77 imes versus a full-precision MLP baseline, with a supporting CU micro-architecture description.

Significance. If the accuracy, robustness and efficiency claims hold under fairer accounting, the work supplies a concrete, low-parameter (21 k) edge-friendly recipe for a practically important wearable bio-signal task. Strengths that should be credited include: evaluation on two public datasets against a broad classical-ML and neural baseline suite (including SpikeFormer), multi-seed averaging for the neural models, an explicit ablation of the SDH terms, a fully specified feature pipeline, and a concrete CU datapath/scheduler that goes beyond pure algorithmic claims. The combination of ANN-to-SNN conversion, aggressive quantization and a noise-oriented training objective is a useful engineering contribution for neuromorphic wearables even if the absolute energy ratio needs recalibration.

major comments (3)
  1. [Section IV-D] Section IV-D (Energy Consumption): the headline 201.77 imes figure is obtained by comparing a pure-AC estimate for the quantized SNN (22 nm low-bit AC costs scaled by an unexplained factor of 2, multiplied by T·r with r≈17 %) against a 45 nm MAC/AC estimate for the MLP (96.85 nJ). The same section later reports a full CU simulation (core + control + memory + leakage) of 19.55 nJ for the T=15 4-bit design—already an order of magnitude higher—yet never recomputes the MLP baseline under an identical full-system model. Without a matched accounting the 201× claim is an optimistic lower bound on SNN cost rather than a fair comparison; the smartwatch lifetime projection inherits the same gap. Either recompute both models under the same full-system energy model or replace the 201× claim with the more conservative CU-based figure and clearly label the pure-AC numbers as operation-count estimates
  2. [Section IV-C2, Table II] Section IV-C2 and Table II: the four “physically motivated” artifacts (multiplicative gain, Bernoulli-masked bursts, cumulative-walk drift, fixed power-line offset) are injected at the already-extracted 132-dimensional feature level rather than on raw sEMG. While the synthetic Gaussian/Poisson/uniform results remain informative, the claim of robustness to “motion-induced artifacts” and “electrode-skin impedance variations” therefore rests on feature-space proxies whose fidelity to real wearable corruption is unvalidated. A short raw-signal noise experiment (or at least a quantitative justification that the chosen feature-level maps preserve the dominant spectral/temporal signatures of the cited artifacts) is needed before the robustness narrative can be considered load-bearing for unconstrained wearable deployment.
  3. [Section III-C, Eq. (5)] Section III-C / Eq. (5): SDH coefficients λ1,λ2,λ3 and Huber δ are selected by grid search on validation data under uniform noise only and then frozen. The paper correctly notes that this avoids noise-specific tuning, yet no sensitivity analysis or confidence intervals on the chosen λ’s are reported, nor is it shown that the same coefficients remain near-optimal when the validation noise type is changed. Because SDH is presented as a central methodological contribution, a brief leave-one-noise-out or coefficient-perturbation study would strengthen the claim that the composite objective generalizes rather than merely fitting the single noise used for selection.
minor comments (5)
  1. [Table I, Section III-E] Table I: the “Post” column uses the earlier method of [9]; it would help the reader if the exact differences between that post-processor and QAF (Algorithm 1) were summarized in one sentence in the caption or in §III-E.
  2. [Figure 5] Figure 5c,f: energy bars are shown on a log scale without explicit numerical labels for the QSN3/QSN4 bars; adding the exact nJ values (or a table) would make the 52–201× ratios easier to verify.
  3. [Section III-B, Eq. (4)] Eq. (4) and the surrounding text use both T (spike-train length) and θ (threshold); later the same symbol T is reused for short-term window duration. A consistent subscript convention (e.g., T_s, T_m, T_spike) would remove ambiguity.
  4. [Section III-F] The CU description in §III-F is valuable but sits somewhat disconnected from the experimental energy numbers; a short sentence linking the simulated 19.55 nJ back to the FSM stages and memory organization would improve cohesion.
  5. Minor typographical issues: “emulating a reset” (p. 2), occasional missing spaces around citations, and “QAFDENOTES” in the Table I caption (formatting artifact).

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity; empirical SNN/QSN results on public sEMG sets are measured against external baselines, with only minor non-forcing self-citations of the authors' prior conversion and SHAQ machinery.

  1. self citation load bearing [Sec. III-B / Eq. (4) and Sec. III-C]
    "we adopt an ANN-to-SNN conversion method. This method first trains a traditional ANN and then converts it into an SNN with the same structure for inference. During training, we use clamp and quantization functions as activation function to optimize the weight transfer... [24]. ... we introduce SDH... which integrates safe haven activation quantization (SHAQ) [25], defensive quantization (DQ) [26], and Huber loss."

    The conversion pipeline and the Lact term of SDH are taken from the authors' own prior papers. This is ordinary reuse of tooling rather than a load-bearing uniqueness claim that forces the fatigue F1 or energy numbers; those remain measured on external data. Flagged only as minor self-citation, not as a reduction of the central result.

full rationale

The paper is an empirical methods contribution: ANN-to-SNN conversion, 3/4-bit PTQ, SDH composite loss, and QAF post-processing are evaluated by F1 on the public SUE/SPE datasets (clean + seven noise types) against classical ML, MLP, CNN, LSTM, Transformer and SpikeFormer baselines, plus an independent energy model. No quantity is defined in terms of the quantity it is said to predict; SDH lambdas are grid-searched once on validation uniform noise and then frozen, so the reported robustness numbers are not forced by construction. Self-citations ([24] CQ+ conversion, [25] SHAQ, [28] ASG) supply reusable components but are not uniqueness theorems that close the loop on the fatigue-detection claim; the performance and energy figures remain externally falsifiable measurements. The ad-hoc 22 nm-to-45 nm scale factor of 2 and incomplete full-system ANN baseline affect correctness of the 201.77x claim but do not constitute circular derivation. Hence only a minimal score for ordinary self-citation of prior tooling.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The central empirical claims rest on standard SNN conversion assumptions, public datasets, hand-chosen architectural and regularization hyper-parameters, and an energy model whose technology scaling factor is ad-hoc. No new physical entities are postulated; free parameters are the usual ML knobs plus the energy scaling constant.

free parameters (6)
  • SDH coefficients λ1, λ2, λ3 and Huber δ
    Selected by grid search on validation set under uniform noise only, then frozen for all other noise types; values not reported numerically but control the robustness claim.
  • timestep T (15 for SUE, 6 for SPE)
    Chosen after ablation; directly multiplies the addition count and therefore the energy ratio.
  • weight bit-width n_bit = 3 or 4
    Post-training choice that trades accuracy for the large energy saving.
  • technology scaling factor of 2 (22 nm → 45 nm)
    Ad-hoc multiplier applied to low-bit AC energies to enable comparison with the 45 nm MAC baseline; not derived from a process model.
  • spike rate ≈ 17.1 %
    Measured on the trained models; enters the energy formula N_ADD = T · r · N_MUL.
  • QAF parameters (window M=3, α=0.2, fast threshold θf)
    Hand-set post-processing constants that improve reported F1.
assumptions (4)
  • domain assumption ANN-to-SNN conversion with clamp-and-quantize activations (CQ) preserves classification accuracy for the chosen architecture and T.
    Invoked throughout Section III-B; relies on prior conversion theory without re-deriving error bounds for this task.
  • domain assumption Feature-level noise models (multiplicative, motion bursts, drift, power-line offset) are faithful proxies for real wearable sEMG artifacts.
    Stated in Section IV-C2; no raw-signal validation is provided.
  • domain assumption A single MAC costs 4.6 pJ and an AC 0.9 pJ in 45 nm; low-bit ACs scale as given in 22 nm literature.
    Taken from Horowitz 2014 and later papers; used for all energy claims in Section IV-D.
  • ad hoc to paper Binary cross-entropy plus the three SDH regularizers is a valid training objective for quantized IF networks.
    Composite loss introduced in Eq. (5); coefficients chosen empirically.
invented entities (2)
  • SDH composite training objective
    purpose: Combine SHAQ activation centering, DQ weight regularization and Huber loss to improve noise robustness under quantization.
    New combination of existing terms; no independent theoretical guarantee outside the reported experiments.
  • QAF (quick adapt filter) post-processing
    purpose: Adaptive thresholding plus fast-channel to reduce latency and false transitions.
    Engineering post-filter; performance gains are empirical.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient and Robust Spiking Neural Networks for sEMG-Based Muscle Fatigue Detection." pith.science (2026). https://pith.science/paper/JTAPXVYH

@misc{pith2026260711065,
  author       = {Pith},
  title        = {Pith review of: Efficient and Robust Spiking Neural Networks for sEMG-Based Muscle Fatigue Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JTAPXVYH}},
  note         = {Machine review of arXiv:2607.11065}
}
read the original abstract

Detecting muscle fatigue via surface electromyography (sEMG) is essential for applications in sports, rehabilitation, and wearable health monitoring. Accurate and timely detection of fatigue is crucial for preventing injuries, optimizing physical performance, and ensuring user safety during prolonged activity. However, existing deep learning models are often unsuitable for this task due to their high computational cost and dependence on large-scale data. In this work, we propose an energy-efficient framework for muscle fatigue detection based on Spiking Neural Networks (SNNs), which exploit sparse, event-driven computation and temporal modeling. We further introduce a quantization-compatible training scheme (SDH) that combines multiple regularization terms to improve robustness under noisy conditions. Evaluated on two public sEMG datasets against a broad set of baselines and under seven noise conditions including physically motivated perturbations, our quantized SNNs match or exceed strong baselines while remaining more stable under diverse noise and reducing estimated energy consumption by up to 201.77x. These results demonstrate the framework's strong potential for real-time deployment in low-power wearable systems.

Figures

Figures reproduced from arXiv: 2607.11065 by the authors.

Figure 1
Figure 1. Overview of the proposed muscle fatigue detection framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. shows the CU datapath and scheduler in our SNN. To minimize power and area, our SNN setting uses a single CU for all computations; for the EMG classification workloads, this CU sustains the required output rate. The CU provides dedicated datapaths for accumulation, bias addition, Algorithm 1 Fatigue Detection Post-Processing Require: Fatigue predictions P, window size w, step size ∆, base threshold θb, smoothing fac… view at source ↗
Figure 3
Figure 3. Ablation study comparing our methods with previous work in terms [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Ablation study on timestep and energy analysis. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 4
Figure 4. Figure 4: Comparison of differential features and all features. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 5 linked inside Pith

  1. [1]

    Muscle fatigue: gen- eral understanding and treatment,

    J.-j. Wan, Z. Qin, P.-y. Wang, Y . Sun, and X. Liu, “Muscle fatigue: gen- eral understanding and treatment,”Experimental & molecular medicine, vol. 49, no. 10, pp. e384–e384, 2017

    2017

  2. [2]

    Sports fatigue detection based on deep learning,

    X. Guan, Y . Lin, Q. Wang, Z. Liu, and C. Liu, “Sports fatigue detection based on deep learning,” in2021 14th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP- BMEI). IEEE, 2021, pp. 1–6

    2021

  3. [3]

    Automated and continuous fatigue monitoring in construction workers using forearm emg and imu wearable sensors and recurrent neural network,

    S. S. Bangaru, C. Wang, and F. Aghazadeh, “Automated and continuous fatigue monitoring in construction workers using forearm emg and imu wearable sensors and recurrent neural network,”Sensors, vol. 22, no. 24, p. 9729, 2022

    2022

  4. [4]

    A review of non-invasive techniques to detect and predict localised muscle fatigue,

    M. R. Al-Mulla, F. Sepulveda, and M. Colley, “A review of non-invasive techniques to detect and predict localised muscle fatigue,”Sensors, vol. 11, no. 4, pp. 3545–3594, 2011

    2011

  5. [5]

    Real-time forecasting of semg features for trunk muscle fatigue using machine learning,

    A. Moniri, D. Terracina, J. Rodriguez-Manzano, P. H. Strutton, and P. Georgiou, “Real-time forecasting of semg features for trunk muscle fatigue using machine learning,”IEEE Transactions on Biomedical Engineering, vol. 68, no. 2, pp. 718–727, 2020

    2020

  6. [6]

    Wearable sensors for monitoring the internal and external workload of the athlete,

    D. R. Seshadri, R. T. Li, J. E. V oos, J. R. Rowbottom, C. M. Alfes, C. A. Zorman, and C. K. Drummond, “Wearable sensors for monitoring the internal and external workload of the athlete,”NPJ digital medicine, vol. 2, no. 1, p. 71, 2019

    2019

  7. [7]

    A real-time, practical sensor fault-tolerant module for robust emg pattern recognition,

    X. Zhang and H. Huang, “A real-time, practical sensor fault-tolerant module for robust emg pattern recognition,”Journal of neuroengineering and rehabilitation, vol. 12, pp. 1–16, 2015

    2015

  8. [8]

    Surface elec- tromyography spectral parameters for the study of muscle fatigue in swimming,

    L. Puce, I. Pallecchi, L. Marinelli, L. Mori, M. Bove, D. Diotti, P. Ruggeri, E. Faelli, F. Cotellessa, and C. Trompetto, “Surface elec- tromyography spectral parameters for the study of muscle fatigue in swimming,”Frontiers in Sports and Active Living, vol. 3, p. 644765, 2021

    2021

Show all 40 references
  1. [9]

    Physical fatigue detection through emg wearables and subjective user reports: a machine learning approach towards adaptive rehabilitation,

    M. Papakostas, V . Kanal, M. Abujelala, K. Tsiakas, and F. Makedon, “Physical fatigue detection through emg wearables and subjective user reports: a machine learning approach towards adaptive rehabilitation,” inProceedings of the 12th ACM international conference on pervasive ...

    2019

  2. [10]

    Psychophysiological data-driven multi-feature information fusion and recognition of miner fatigue in high-altitude and cold areas,

    S. Chen, K. Xu, X. Yao, S. Zhu, B. Zhang, H. Zhou, X. Guo, and B. Zhao, “Psychophysiological data-driven multi-feature information fusion and recognition of miner fatigue in high-altitude and cold areas,” Computers in biology and medicine, vol. 133, p. 104413, 2021

    2021

  3. [11]

    A muscle fatigue classification model based on lstm and improved wavelet packet threshold,

    J. Wang, S. Sun, and Y . Sun, “A muscle fatigue classification model based on lstm and improved wavelet packet threshold,”Sensors, vol. 21, no. 19, p. 6369, 2021

    2021

  4. [12]

    Mffnet: Multi-dimensional feature fusion network based on attention mechanism for semg analysis to detect muscle fatigue,

    Y . Zhang, S. Chen, W. Cao, P. Guo, D. Gao, M. Wang, J. Zhou, and T. Wang, “Mffnet: Multi-dimensional feature fusion network based on attention mechanism for semg analysis to detect muscle fatigue,”Expert Systems with Applications, vol. 185, p. 115639, 2021

    2021

  5. [13]

    Design of analog and digital filter of electromyography,

    L. Rozaqi, A. Nugroho, K. H. Sanjaya, and A. I. Simbolon, “Design of analog and digital filter of electromyography,” in2019 International Conference on Sustainable Energy Engineering and Application (IC- SEEA). IEEE, 2019, pp. 186–192

    2019

  6. [14]

    Estimation of mean and median frequency from synthetic semg signals: Effects of different spectral shapes and noise on estimation methods,

    G. Corvini, C. D’Anna, and S. Conforto, “Estimation of mean and median frequency from synthetic semg signals: Effects of different spectral shapes and noise on estimation methods,”Biomedical Signal Processing and Control, vol. 73, p. 103420, 2022

    2022

  7. [15]

    Evaluation of data processing and artifact removal approaches used for physiological signals captured using wearable sensing devices during construction tasks,

    S. Anwer, H. Li, M. F. Antwi-Afari, A. M. Mirza, M. A. Rahman, I. Mehmood, R. Guo, and A. Y . L. Wong, “Evaluation of data processing and artifact removal approaches used for physiological signals captured using wearable sensing devices during construction tasks,”Journal of Co...

    2024

  8. [16]

    Training spiking neural networks using lessons from deep learning,

    J. K. Eshraghian, M. Ward, E. O. Neftci, X. Wang, G. Lenz, G. Dwivedi, M. Bennamoun, D. S. Jeong, and W. D. Lu, “Training spiking neural networks using lessons from deep learning,”Proceedings of the IEEE, vol. 111, no. 9, pp. 1016–1054, 2023

    2023

  9. [17]

    Onespike: Ultra-low latency spiking neural networks,

    K. Tang, Z. Yan, and W.-F. Wong, “Onespike: Ultra-low latency spiking neural networks,” in2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024, pp. 1–8

    2024

  10. [18]

    Eescn: A novel spiking neural network method for eeg-based emotion recog- nition,

    F. Xu, D. Pan, H. Zheng, Y . Ouyang, Z. Jia, and H. Zeng, “Eescn: A novel spiking neural network method for eeg-based emotion recog- nition,”Computer methods and programs in biomedicine, vol. 243, p. 107927, 2024

    2024

  11. [19]

    An ultra- low power time-domain based snn processor for ecg classification,

    H. Fan, L. Chang, J. Zhou, X. Yang, S. Lin, and J. Zhou, “An ultra- low power time-domain based snn processor for ecg classification,” in 2024 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 2024, pp. 1–5

    2024

  12. [20]

    An energy-efficient and robust snn classifier for lc-adc sampled ecg signals,

    H. Chu, Y . Yan, Y . Zhou, Z. Zhu, Y . Huan, L. Zheng, and Z. Zou, “An energy-efficient and robust snn classifier for lc-adc sampled ecg signals,” in2023 8th International Conference on Integrated Circuits and Microsystems (ICICM). IEEE, 2023, pp. 502–506

    2023

  13. [21]

    Hr-snn: An end-to-end spiking neural network for four-class classification motor imagery brain-computer interface,

    Y . Li, L. Fan, H. Shen, and D. Hu, “Hr-snn: An end-to-end spiking neural network for four-class classification motor imagery brain-computer interface,”IEEE Transactions on Cognitive and Developmental Systems, 2024

    2024

  14. [22]

    A novel event-driven spiking convolutional neural network for electromyography pattern recognition,

    M. Xu, X. Chen, A. Sun, X. Zhang, and X. Chen, “A novel event-driven spiking convolutional neural network for electromyography pattern recognition,”IEEE Transactions on Biomedical Engineering, vol. 70, no. 9, pp. 2604–2615, 2023

    2023

  15. [23]

    Spgesture: Source-free domain-adaptive semg-based gesture recognition with jac- card attentive spiking neural network,

    W. Guo, Y . Sun, Y . Xu, Z. Qiao, Y . Yang, and H. Xiong, “Spgesture: Source-free domain-adaptive semg-based gesture recognition with jac- card attentive spiking neural network,”arXiv preprint arXiv:2405.14398, 2024

    2024 arXiv

  16. [24]

    Cq+ training: Minimizing accuracy loss in conversion from convolutional neural networks to spiking neural networks,

    Z. Yan, J. Zhou, and W.-F. Wong, “Cq+ training: Minimizing accuracy loss in conversion from convolutional neural networks to spiking neural networks,”IEEE Transactions on Pattern Analysis and Machine Intelli- gence, vol. 45, no. 10, pp. 11 600–11 611, 2023

    2023

  17. [25]

    Improving model robustness against noise with safe haven activations,

    Z. Yan, D. Y . Polat, S. Wang, K. Tang, and W.-F. Wong, “Improving model robustness against noise with safe haven activations,” 2024. [Online]. Available: https://openreview.net/forum?id=PoSq0B0ffE

    2024

  18. [26]

    Defensive quantization: When efficiency meets robustness,

    J. Lin, C. Gan, and S. Han, “Defensive quantization: When efficiency meets robustness,”arXiv preprint arXiv:1904.08444, 2019

    1904 arXiv

  19. [27]

    Optimal ann- snn conversion for high-accuracy and ultra-low-latency spiking neural networks,

    T. Bu, W. Fang, J. Ding, P. Dai, Z. Yu, and T. Huang, “Optimal ann- snn conversion for high-accuracy and ultra-low-latency spiking neural networks,”arXiv preprint arXiv:2303.04347, 2023

    2023 arXiv

  20. [28]

    Low latency conversion of artificial neural network models to rate-encoded spiking neural networks,

    Z. Yan, J. Zhou, and W.-F. Wong, “Low latency conversion of artificial neural network models to rate-encoded spiking neural networks,”arXiv preprint arXiv:2211.08410, 2022

    2022 arXiv

  21. [29]

    Reducing noise, artifacts and interference in single-channel emg signals: A review,

    M. Boyer, L. Bouyer, J.-S. Roy, and A. Campeau-Lecours, “Reducing noise, artifacts and interference in single-channel emg signals: A review,” Sensors, vol. 23, no. 6, p. 2927, 2023

    2023

  22. [30]

    Application of an emg interference filtering method to dynamic ecgs based on an adaptive wavelet-wiener filter and adaptive moving average filter,

    Y . Li, Z. Su, K. Chen, W. Zhang, and M. Du, “Application of an emg interference filtering method to dynamic ecgs based on an adaptive wavelet-wiener filter and adaptive moving average filter,”Biomedical Signal Processing and Control, vol. 72, p. 103344, 2022

    2022

  23. [31]

    Assessment of self-report, palpation, and surface electromyography dataset during isometric muscle contraction,

    J. Lim, L. Lu, K. Goonewardena, J. Z. Liu, and Y . Tan, “Assessment of self-report, palpation, and surface electromyography dataset during isometric muscle contraction,”Scientific Data, vol. 11, no. 1, p. 208, 2024

    2024

  24. [32]

    Pressure-controlled palpation: a new technique which increases the reliability of manual palpation,

    L. Bendtsen, R. Jensen, N. Jensen, and J. Olesen, “Pressure-controlled palpation: a new technique which increases the reliability of manual palpation,”Cephalalgia, vol. 15, no. 3, pp. 205–210, 1995

    1995

  25. [33]

    Application of surface electromyography in exercise fatigue: a review,

    J. Sun, G. Liu, Y . Sun, K. Lin, Z. Zhou, and J. Cai, “Application of surface electromyography in exercise fatigue: a review,”Frontiers in Systems Neuroscience, vol. 16, p. 893275, 2022

    2022

  26. [34]

    Advances and disturbances in semg-based intentions and movements recognition: A review,

    H. Xu and A. Xiong, “Advances and disturbances in semg-based intentions and movements recognition: A review,”IEEE Sensors Journal, vol. 21, no. 12, pp. 13 019–13 028, 2021

    2021

  27. [35]

    A laguerre–volterra network model based on ant colony optimization applied to evaluate emg-force relationship in the muscle fatigue state,

    M. Ma, X. Luo, S. Xiahou, and X. Shan, “A laguerre–volterra network model based on ant colony optimization applied to evaluate emg-force relationship in the muscle fatigue state,”Review of Scientific Instruments, vol. 95, no. 6, 2024

    2024

  28. [36]

    Machine learning and regression analysis for diagnosis of bruxism by using emg signals of jaw muscles,

    T. Sonmezocak and S. Kurt, “Machine learning and regression analysis for diagnosis of bruxism by using emg signals of jaw muscles,” Biomedical Signal Processing and Control, vol. 69, p. 102905, 2021

    2021

  29. [37]

    1.1 computing’s energy problem (and what we can do about it),

    M. Horowitz, “1.1 computing’s energy problem (and what we can do about it),” in2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC). IEEE, 2014, pp. 10–14

    2014

  30. [38]

    Weardrive: Fast and energy-efficient storage for wearables,

    J. Huang, A. Badam, R. Chandra, and E. B. Nightingale, “Weardrive: Fast and energy-efficient storage for wearables,” in2015 USENIX Annual Technical Conference (USENIX ATC 15), 2015, pp. 613–625

    2015

  31. [39]

    A neuromorphic processing system with spike-driven snn processor for wearable ecg classification,

    H. Chu, Y . Yan, L. Gan, H. Jia, L. Qian, Y . Huan, L. Zheng, and Z. Zou, “A neuromorphic processing system with spike-driven snn processor for wearable ecg classification,”IEEE Transactions on Biomedical Circuits and Systems, vol. 16, no. 4, pp. 511–523, 2022

    2022

  32. [40]

    Reconsidering the energy efficiency of spiking neural networks,

    Z. Yan, Z. Bai, and W.-F. Wong, “Reconsidering the energy efficiency of spiking neural networks,”arXiv preprint arXiv:2409.08290, 2024

    2024 arXiv

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.