Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Understanding of Task-specific and Subject-specific Components in Surface EMG

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that surface EMG can be split into task-specific components shared across people and subject-specific components unique to each person, and that this separation substantially boosts both gesture recognition and user identi

desk verdict Solid inter-day accuracy gains from disentangled sEMG latents, but the physiological interpretation rests on out-of-distribution reconstructions and needs a control before the overclaims can stand. read the letter →

arxiv 2607.14744 v1 pith:Y6EH36HK submitted 2026-07-16 cs.HC

Yangyang Yuan , Jionghui Liu , Xinyu Jiang , ChihHong Chou , Chenyun Dai , Jiahao Fan This is my paper
classification cs.HC
keywords surfaceelectromyographysEMGfeaturedisentanglementhandgesturerecognitionuseridentificationhigh-densityinter-daygeneralizationinterpretableneuralnetwork
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 surface EMG recorded during hand gestures contains two separable kinds of information: a task-specific component shared across people performing the same movement, and a subject-specific component that identifies the individual. Using a two-encoder, one-decoder network trained with reconstruction, triplet, and cross-reconstruction losses, the authors separate these components. The task-specific component lifts inter-day gesture recognition from 74.42% to 91.47%, and the subject-specific component lifts inter-day user identification from 50.7% to 64.65%, both with a simple KNN classifier. The reconstructed activation maps show task-specific components clustering by gesture across subjects and subject-specific components clustering by person across gestures. If correct, this gives prosthetics and biometric systems a way to generalize across users without per-user calibration.

What carries the argument

The load-bearing mechanism is a two-encoder/single-decoder autoencoder trained on RMS spatial features from 128-channel sEMG arrays. A task encoder and a subject encoder map the input into two latent subspaces; the training loss combines reconstruction loss, triplet losses that pull same-gesture (or same-subject) latents together while pushing different ones apart, and a cross-reconstruction loss that swaps latents between samples. The decoder reconstructs the input from the concatenated latents. To visualize what each component encodes, the authors replace the other latent with an all-zero matrix and decode; the resulting RMS maps are compared across gestures and subjects.

What would settle it

Compute the decoder's reconstruction error when one latent is set to zero instead of a held-out real latent; if the zero-substitution reconstructions are far outside the training distribution (large error or nonsensical maps), the visual interpretation in Figures 6-9 is not supported. Alternatively, train the identical network with gesture labels shuffled; if the task-specific maps still cluster by gesture label, the clustering is a label artifact.

Watch

Extended reading notes

Core claim

The central claim is that a neural network can disentangle the RMS spatial pattern of high-density sEMG into two orthogonal latent subspaces: one capturing what is common to a gesture across all people, and one capturing what is constant for a person across gestures. The paper shows that the task-specific component alone supports 91.47% inter-day accuracy across 20 subjects and 11 single-degree-of-freedom gestures, while the subject-specific component supports 64.65% inter-day identity recognition. Visualizations of the decoded maps (obtained by feeding one latent component and a zero matrix through the decoder) show consistent muscle-activation layouts for the same gesture across subjects,

Load-bearing premise

The physiological interpretation hinges on the assumption that feeding an all-zero matrix in place of one latent component produces a meaningful task- or subject-specific reconstruction, even though the decoder was trained only on real paired latents and never saw zero vectors.

Editorial extensions

If this is right

  • Cross-day gesture recognition on this 20-subject, 11-gesture dataset rises from 74.42% to 91.47% using only KNN, showing the gain comes from feature separation, not a stronger classifier.
  • Cross-day user identification rises from 50.7% to 64.65%, reaching a range where combining multiple gestures can push accuracy toward practical biometric authentication.
  • Task-specific RMS maps have smaller inter-subject variance and more separated gesture centroids than original RMS, providing a spatial explanation for the recognition gain.
  • Subject-specific components are less similar across days than task-specific components, which explains why identification degrades more than recognition in inter-day conditions.
  • Because the extracted features already support a simple classifier, they can be plugged into more complex or temporal models without additional calibration.

Reading between the lines

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

  • A testable extension the paper does not pursue: comparing zero-substitution reconstructions against independent measurements such as motor-unit decomposition, ultrasound, or MRI would test whether the decoded maps truly reflect muscle anatomy rather than decoder artifacts.
  • The disentanglement objective is generic enough that the same two-encoder, mixed-loss recipe could separate task from subject factors in other biosignals, such as EEG or ECG, where inter-subject variability blocks cross-user generalization.
  • Since single-DoF gestures can be combined linearly, the task-specific component might let a user synthesize multi-DoF commands from single-DoF templates, reducing the need for exhaustive multi-DoF training data.
  • The paper's own cross-day gap in subject-specific components suggests practical identity systems would need either frequent recalibration or multi-gesture aggregation; the authors mention the latter in the discussion.
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

4 major / 5 minor

Summary. The paper proposes a two-encoder/single-decoder autoencoder that disentangles high-density sEMG features into a task-specific latent space and a subject-specific latent space, using reconstruction, triplet, and cross-reconstruction losses. The authors evaluate the extracted latents on gesture recognition and user identification under intra-day and inter-day protocols on an open 20-subject, 11-gesture dataset, reporting large inter-day gains (gesture accuracy 74.42% to 91.47%; identification accuracy 50.7% to 64.65%) over PCA and standard autoencoder baselines. They also reconstruct RMS maps from single latent components by zeroing the other component, and use these maps to argue that task-specific components reflect consistent across-subject muscle activation patterns while subject-specific components capture individual characteristics.

Significance. If the central claims hold, the paper makes a useful contribution to sEMG-based human-machine interfaces: a feature extractor that simultaneously improves cross-subject/cross-day gesture recognition and user identification, with an interpretability analysis linking learned latents to muscle activation patterns. Strengths include the use of a public dataset, a clearly described network and loss, comparison against two standard feature-extraction baselines, and a code release. The classification improvements are substantial and, if reproducible, would be of practical interest. However, the physiological-interpretation claim is currently supported by a reconstruction procedure that is not validated for the zero-latent inputs it employs, and the statistical reporting contains inconsistencies that need correction before the quantitative claims can be fully assessed.

major comments (4)
  1. [§5.3, Figs. 6–9] The component RMS maps are generated by D(E_p, 0) and D(0, E_s), but the decoder was trained only on real encoded latent pairs and on cross-combinations of real latents via Eq. (5). An all-zero latent vector is never seen during training, so these reconstructions are out-of-distribution extrapolations. The resulting maps and all downstream centroid statistics (Figs. 8–9) therefore do not currently validate the physiological interpretation. The authors should either train with zero-replacement as an auxiliary reconstruction target, provide an in-distribution control (e.g., replacing the zeroed component with a random or mean latent and showing the component-specific maps change appropriately), or explicitly temper the physiological claims. Visual plausibility is not independent evidence, because a decoder trained to output real RMS maps can generate plausible-looking outputs for inputs ou
  2. [§4.3, Eqs. (3)–(5)] The triplet loss L_trip_p and L_trip_s explicitly encourage same-gesture clustering in E_p and same-subject clustering in E_s, while L_cross further enforces that swapping latents reconstructs the corresponding attribute. The t-SNE clusters in Fig. 4 and the high SIL scores in Fig. 5 are therefore at least partly a consequence of the training objective, not independent evidence that the model 'discovered' physiological components. The classification results in Tables 2–3 are not affected by this circularity, but the interpretability narrative should be reframed: these losses define the desired disentanglement, and the scientific claim should rest on held-out generalization and physiological validation rather than on clustering quality in the training objective's latent space.
  3. [§4.4.1, §5.1, Tables 2–3] The experimental protocol is described inconsistently. Section 4.4.1 first says 'Ak-fold cross-validation was implemented' but then states that two-thirds of samples from each gesture were used for training and the rest for testing, which describes a single split, not k-fold. If k-fold was used, the number of folds and the variance across folds must be reported; if a single two-thirds split, the paper should not call it k-fold. Additionally, §4.5 states that a 'non-parametric chi-square test' was used, but §5.1 reports Friedman tests with χ²(3) and post-hoc comparisons. These are different procedures; the statistical section must be corrected to match the analyses actually performed. Tables 2–3 report point accuracies without confidence intervals or repeated-split variability, so the reader cannot assess the stability of the reported large improvements.
  4. [§6, Discussion] The Discussion acknowledges that 'this physiological validation is relatively superficial' and calls for imaging or anatomical methods to verify interpretability. This admission is consistent with Major Comment 1, but it conflicts with the abstract's claim that the study provides 'deeper insights into the physiological mechanisms underlying sEMG signals.' The authors should either strengthen the physiological validation or soften the abstract and conclusion claims to match the actual evidence. Given that the abstract-level interpretability claim is a key selling point, this needs to be addressed before the paper can be accepted.
minor comments (5)
  1. [General] The paper uses 'LRLU' in Table 1 but 'LeakyReLU' in the text; unify the terminology. Also, Eq. (3) writes 'where l,i,k,j' without defining the sampling procedure for l and k; clarify that l indexes a different subject and k indexes a different gesture in the triplet formulation.
  2. [§4.4.1] The description of the inter-day protocol says Day 1 is used for training and Day 2 for testing, but the data augmentation description says each original sample undergoes one translation and one rotation, 'resulting in a final sample size three times that of the original dataset.' Clarify whether augmentation is applied only to the training split and whether test samples are unaugmented.
  3. [Fig. 6] The caption mentions that gray squares indicate unavailable data, but the figure is not visible in the text; ensure the final version clearly shows this. Also, the interpolation to scale 5 with cubic interpolation should be described in the caption so readers understand the maps are not raw electrode readings.
  4. [§5.3–5.4] The centroid distance is measured from the top-left corner of the RMS map, but Fig. 8 and Fig. 9 do not label the axes with units (mm or electrode indices). Adding units and a schematic of the electrode layout would improve interpretability.
  5. [References] Reference [38] is cited as an open dataset, but the URL is given only as a DOI in a footnote. The main text should include the full citation and a direct link to the dataset repository.

Circularity Check

1 steps flagged · score 6.0 of 10

Interpretability claims partly restate the training objective; held-out accuracy gains remain independent.

  1. self definitional [Eq. 3 (§4.3); interpreted as discovery in §5.3 Fig. 6 and §6]
    "in the reconstructed RMS array of task-specific components (bottom left in Figure 6), the heatmaps in the same column exhibit different patterns, whereas those in the same row are very similar. This indicates that the RMS of the task-specific components preserves the differences between gestures while showing high consistency for the same gesture across different subjects."

    The task-specific latent space is defined in §4.3 as one where same-gesture samples across subjects are tightly clustered, and Eq. 3 (L_trip_p) explicitly trains it to 'ensure that the same gestures are clustered together while different gestures are separated.' The RMS map used as evidence is D(E_p(x),0), a deterministic function of E_p. Once E_p is trained with this triplet objective, same-gesture rows are similar by construction (given decoder continuity), so the consistency is a restatement of the loss, not an emergent physiological property. The subject-specific same-column consistency is the same restatement of L_trip_s. The abstract's 'consistent activation patterns' claim therefore adds no independent information beyond the training objective.

full rationale

The paper's classification improvements (Tables 2 and 3) are genuine held-out predictions: KNN is applied to latent features from models trained on Day 1 / training folds and tested on Day 2 / held-out samples. Those numbers do not reduce to the losses by construction, so the accuracy claims are not circular. However, the interpretability narrative is partly self-confirming: the 'discovery' that task-specific components are consistent for the same gesture across subjects (and subject-specific components consistent within a subject) is exactly the objective of the triplet losses (Eqs. 3-4). The RMS maps in §5.3 are additionally produced by replacing one latent with an all-zero vector; the decoder was trained only on real latent pairs (Eq. 2) and cross-combinations of real latents (Eq. 5), so these maps are out-of-distribution extrapolations rather than validated component decompositions. The authors themselves state that 'this physiological validation is relatively superficial' (Discussion), which weighs against treating the component maps as independent evidence. On balance, the central interpretability claim reduces in part to the training definition, while the accuracy contribution remains independent, giving partial circularity rather than a complete collapse.

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

The paper's central empirical claims rest on a handful of hand-tuned hyperparameters and several modeling assumptions. The most consequential is that the loss functions produce physiologically meaningful latent components; because those losses directly enforce clustering, the interpretability claims inherit a large circularity burden. The zero-latent decoding assumption is an additional unverified premise.

free parameters (6)
  • Loss balance weights lambda_1, lambda_2 = lambda_1=1, lambda_2=0.5
    Chosen in preliminary experiments to balance cross-reconstruction and triplet losses; authors state performance is insensitive (<1%).
  • Triplet margin alpha = not reported
    Margin in Eq. (3) controls how much farther apart different-gesture/different-subject pairs must be; it is not numerically specified, yet it defines the degree of imposed clustering.
  • LeakyReLU slope = 0.2
    Selected because performance degrades when slope exceeds 0.2; a hand-tuned hyperparameter.
  • Dropout probability = 0.2 or 0.8 (contradictory)
    Table 1 and Section 4.3 state 0.2, while Section 4.4 states 0.8 for decoder layers; the value is ambiguous.
  • Data augmentation translation/rotation ranges = translation +/-15 mm, rotation +/-15 degrees
    Chosen to simulate electrode shift; not derived from measured electrode displacement.
  • Training hyperparameters = Adam lr=0.002, batch size=2000, 500 iterations
    Standard choices, not independently justified or tied to the dataset size.
assumptions (4)
  • domain assumption sEMG features decompose into orthogonal task-specific and subject-specific subspaces
    Section 4.3 states this as the fundamental concept; the network is built to enforce it, and no alternative decomposition model is tested.
  • ad hoc to paper Triplet and cross-reconstruction losses yield components that correspond to physiological muscle activation and individual characteristics
    The paper equates learned latent clusters with neuromuscular meaning; this is the target of optimization, not an independent finding.
  • ad hoc to paper Reconstructing with a zeroed latent vector produces a valid component-specific RMS map
    Section 5.3 replaces one component with an all-zero matrix and decodes; the decoder was not trained on zero latent inputs, so this out-of-distribution assumption is unvalidated.
  • domain assumption Data augmentation by translation and rotation approximates electrode displacement across days
    Used to justify inter-day generalization; the displacement distribution is chosen, not measured per subject.
invented entities (2)
  • Task-specific latent component (E_p output)
    purpose: Subject-invariant gesture representation used for gesture recognition and physiological mapping
    It is a learned construct. The paper claims it corresponds to shared muscle-activation patterns, but evidence is visual and partly imposed by the triplet loss; no independent physiological validation is provided.
  • Subject-specific latent component (E_s output)
    purpose: Task-invariant identity representation used for user identification
    Same issue as above: a learned construct whose claimed correspondence to individual characteristics is qualitatively inspected, not independently validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding of Task-specific and Subject-specific Components in Surface EMG." pith.science (2026). https://pith.science/paper/Y6EH36HK

@misc{pith2026260714744,
  author       = {Pith},
  title        = {Pith review of: Understanding of Task-specific and Subject-specific Components in Surface EMG},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y6EH36HK}},
  note         = {Machine review of arXiv:2607.14744}
}
read the original abstract

Surface electromyogram (sEMG) signals are widely used in human-machine interfaces for gesture recognition and user identification, but existing models often struggle to generalize across individuals due to subject-specific neuromuscular characteristics. This study introduces a disentanglement model that separates task-specific and subject-specific components from sEMG signals, thereby improving the generalization and interpretability of gesture recognition and user identification systems. Experimental results demonstrate that the disentangled components significantly improve the accuracy of both gesture classification and user identification across subjects and days, outperforming conventional methods under the same experimental conditions. Further analysis reveals that the task-specific components capture consistent activation patterns associated with the same gestures across individuals. In contrast, the subject-specific components reflect unique neuromuscular characteristics that can be used for user identification. Notably, the subject-specific components show lower similarity across days than the task-specific components, contributing to a greater decrease in user identification accuracy than in gesture recognition accuracy. These findings suggest that the disentanglement approach not only improves classification performance but also provides deeper insights into the physiological mechanisms underlying sEMG signals. The model's ability to isolate and interpret different neuromuscular components holds promise for enhancing the robustness of sEMG-based applications in real-world settings, including rehabilitation and user authentication. Our code is available at https://github.com/Open-EXG/HandDisentanglement.

Figures

Figures reproduced from arXiv: 2607.14744 by the authors.

Figure 1
Figure 1. Electrode setup in the experiment. as an indicator to screen out reliable instantaneous samples for updating the feature extractor, thus improving the alignment of feature representations of myoelectric patterns across users. To address the sEMG distribution variations between disparate training users, Zhang et. al. proposed a multi-source synchronize domain adaptation framework to align each source user and the new… view at source ↗
Figure 2
Figure 2. Gestures involved in the validation. 3. Materials We validated the proposed method on the open source dataset from our previous work [38], which can be accessed at the website 1 . In this study, we only used the pattern recognition subset of the dataset for analysis. In the following, we provided a brief introduction about the subject information and data acquisition. 3.1. Subjects The experiment invited 20 subjects… view at source ↗
Figure 3
Figure 3. The framework of the mode, composed of two encoders and one decoder. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The visualization of original features (top row, with two figures representing features labeled by identities and gestures, respectively) and [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: The SIL scores of the task- and subject-specific latent representations in intra-day and inter-day validation. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: The visualization of the original RMS (top) and the RMS reconstructed by the proposed approach (bottom left denotes the task-specific [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: The averaged ’task’ RMS and ’identity’ RMS. Ges. and Subj. denote Gesture and Subject respectively. (A) the overlap of RMS maps for [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: The comparison between the centroid distribution of each gesture across all subjects, as depicted by the original RMS and the RMS [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: The comparison between the centroid distribution of each subject across all gestures, as depicted by the original RMS and the RMS [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 2 linked inside Pith

  1. [1]

    Zheng, M

    M. Zheng, M. S. Crouch, M. S. Eggleston, Surface electromyography as a natural human–machine interface: a review, IEEE Sensors Journal 22 (10) (2022) 9198–9214

    2022

  2. [2]

    Xiong, D

    D. Xiong, D. Zhang, X. Zhao, Y . Zhao, Deep learning for EMG-based human-machine interaction: A review, IEEE/CAA Journal of Automatica Sinica 8 (3) (2021) 512–533

    2021

  3. [3]

    Y . Hu, Y . Wong, W. Wei, Y . Du, M. Kankanhalli, W. Geng, A novel attention-based hybrid CNN-RNN architec- ture for sEMG-based gesture recognition, PloS one 13 (10) (2018) e0206049

    2018

  4. [4]

    Z. Su, H. Liu, J. Qian, Z. Zhang, L. Zhang, Hand gesture recognition based on sEMG signal and convolutional neural network, International Journal of Pattern Recognition and Artificial Intelligence 35 (11) (2021) 2151012

    2021

  5. [5]

    S. Guo, H. Cai, J. Guo, A method of evaluating rehabilitation stage by sEMG signals for the upper limb rehabil- itation robot, in: 2019 IEEE International Conference on Mechatronics and Automation (ICMA), IEEE, 2019, pp. 1338–1343

    2019

  6. [6]

    Z. Yang, S. Guo, Y . Liu, M. Kawanishi, H. Hirata, A task performance-based sEMG-driven variable stiffness control strategy for upper limb bilateral rehabilitation system, IEEE/ASME Transactions on Mechatronics 28 (2) (2022) 792–803

    2022

  7. [7]

    S. H. Yeon, T. Shu, H. Song, T.-H. Hsieh, J. Qiao, E. A. Rogers, S. Gutierrez-Arango, E. Israel, L. E. Freed, H. M. Herr, Acquisition of surface EMG using flexible and low-profile electrodes for lower extremity neuroprosthetic control, IEEE transactions on medical robotics and bionics 3 (3) (2021) 563–572

    2021

  8. [8]

    Disselhorst-Klug, S

    C. Disselhorst-Klug, S. Williams, Surface electromyography meets biomechanics: correct interpretation of sEMG-signals in neuro-rehabilitation needs biomechanical input, Frontiers in Neurology 11 (2020) 603550

    2020

Show all 46 references
  1. [9]

    Khezri, M

    M. Khezri, M. Jahed, Real-time intelligent pattern recognition algorithm for surface EMG signals, Biomedical engineering online 6 (1) (2007) 1–12

    2007

  2. [10]

    Zhuojun, T

    X. Zhuojun, T. Yantao, L. Yang, sEMG pattern recognition of muscle force of upper arm for intelligent bionic limb control, Journal of Bionic Engineering 12 (2) (2015) 316–323

    2015

  3. [11]

    T. M. Bittibssi, M. A. Genedy, S. A. Maged, et al., sEMG pattern recognition based on recurrent neural network, Biomedical Signal Processing and Control 70 (2021) 103048

    2021

  4. [12]

    C. Dai, X. Hu, Extracting and classifying spatial muscle activation patterns in forearm flexor muscles using high-density electromyogram recordings, International journal of neural systems 29 (01) (2019) 1850025. 18

    2019

  5. [13]

    L. Meng, X. Jiang, X. Liu, J. Fan, H. Ren, Y . Guo, H. Diao, Z. Wang, C. Chen, C. Dai, et al., User-tailored hand gesture recognition system for wearable prosthesis and armband based on surface electromyogram, IEEE Transactions on Instrumentation and Measurement 71 (2022) 1–16

    2022

  6. [14]

    X. Zhai, B. Jelfs, R. H. Chan, C. Tin, Self-recalibrating surface EMG pattern recognition for neuroprosthesis control based on convolutional neural network, Frontiers in neuroscience 11 (2017) 379

    2017

  7. [15]

    Nasri, S

    N. Nasri, S. Orts-Escolano, M. Cazorla, An sEMG-controlled 3d game for rehabilitation therapies: Real-time time hand gesture recognition using deep learning techniques, Sensors 20 (22) (2020) 6451

    2020

  8. [16]

    Y . Guo, J. Liu, Y . Wu, X. Jiang, Y . Wang, L. Meng, X. Liu, F. Shu, C. Dai, W. Chen, sEMG-based inter-session hand gesture recognition via domain adaptation with locality preserving and maximum margin., International Journal of Neural Systems 34 (3) (2024) 2450010–2450010

    2024

  9. [17]

    Kanoga, T

    S. Kanoga, T. Hoshino, H. Asoh, Subject transfer framework based on source selection and semi-supervised style transfer mapping for sEMG pattern recognition, in: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2020, pp....

    2020

  10. [18]

    S. Suh, V . F. Rey, P. Lukowicz, Tasked: Transformer-based adversarial learning for human activity recognition using wearable sensors via self-knowledge distillation, Knowledge-Based Systems 260 (2023) 110143

    2023

  11. [19]

    Y . Lin, R. Palaniappan, P. De Wilde, L. Li, A normalisation approach improves the performance of inter-subject sEMG-based hand gesture recognition with a convnet, in: 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 2020,...

    2020

  12. [20]

    X. Chen, Y . Li, R. Hu, X. Zhang, X. Chen, Hand gesture recognition based on surface electromyography using convolutional neural network with transfer learning method, IEEE Journal of Biomedical and Health Informatics 25 (4) (2021) 1292–1304. doi:10.1109/JBHI.2020.3009383

    2021

  13. [21]

    Jiang, B

    S. Jiang, B. Lv, X. Sheng, C. Zhang, H. Wang, P. B. Shull, Development of a real-time hand gesture recog- nition wristband based on sEMG and IMU sensing, in: 2016 IEEE International Conference on Robotics and Biomimetics (ROBIO), IEEE, 2016, pp. 1256–1261

    2016

  14. [22]

    Jiang, K

    X. Jiang, K. Xu, X. Liu, C. Dai, D. A. Clifton, E. A. Clancy, M. Akay, W. Chen, Neuromuscular password-based user authentication, IEEE Transactions on Industrial Informatics 17 (4) (2020) 2641–2652

    2020

  15. [23]

    Gupta, R

    V . Gupta, R. Sadana, S. Moudgil, Image style transfer using convolutional neural networks based on transfer learning, International journal of computational systems engineering 5 (1) (2019) 53–60

    2019

  16. [24]

    Y . Deng, F. Tang, W. Dong, C. Ma, X. Pan, L. Wang, C. Xu, Stytr2: Image style transfer with transformers, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11326–11336

    2022

  17. [25]

    G. Kwon, J. C. Ye, Clipstyler: Image style transfer with a single text condition, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 18062–18071

    2022

  18. [26]

    Aberman, R

    K. Aberman, R. Wu, D. Lischinski, B. Chen, D. Cohen-Or, Learning character-agnostic motion for motion retargeting in 2d, arXiv preprint arXiv:1905.01680 (2019)

    1905 arXiv

  19. [27]

    X. Gu, Y . Guo, F. Deligianni, B. Lo, G.-Z. Yang, Cross-subject and cross-modal transfer for generalized ab- normal gait pattern recognition, IEEE Transactions on Neural Networks and Learning Systems 32 (2) (2020) 546–560

    2020

  20. [28]

    J. Fan, X. Jiang, X. Liu, L. Meng, F. Jia, C. Dai, Surface EMG feature disentanglement for robust pattern recognition, Expert Systems with Applications 237 (2024) 121224. 19

    2024

  21. [29]

    K.-T. Kim, C. Guan, S.-W. Lee, A subject-transfer framework based on single-trial EMG analysis using convo- lutional neural networks, IEEE Transactions on Neural Systems and Rehabilitation Engineering 28 (1) (2019) 94–103

    2019

  22. [30]

    Y . Wang, P. Zhao, Z. Zhang, A deep learning approach using attention mechanism and transfer learning for electromyographic hand gesture estimation, Expert Systems with Applications 234 (2023) 121055

    2023

  23. [31]

    Jiang, X

    X. Jiang, X. Liu, J. Fan, X. Ye, C. Dai, E. A. Clancy, D. Farina, W. Chen, Enhancing iot security via cancelable HD-sEMG-based biometric authentication password, encoded by gesture, IEEE Internet of Things Journal 8 (22) (2021) 16535–16547

    2021

  24. [32]

    Jiang, X

    X. Jiang, X. Liu, J. Fan, X. Ye, C. Dai, E. A. Clancy, W. Chen, Measuring neuromuscular electrophysiological activities to decode HD-sEMG biometrics for cross-application discrepant personal identification with unknown identities, IEEE Transactions on Instrumentation and Measu...

    2022

  25. [33]

    Zhang, X

    X. Zhang, X. Zhang, L. Wu, C. Li, X. Chen, X. Chen, Domain adaptation with self-guided adaptive sampling strategy: Feature alignment for cross-user myoelectric pattern recognition, IEEE Transactions on Neural Systems and Rehabilitation Engineering 30 (2022) 1374–1383

    2022

  26. [34]

    Zhang, L

    X. Zhang, L. Wu, X. Zhang, X. Chen, C. Li, X. Chen, Multi-source domain generalization and adaptation toward cross-subject myoelectric pattern recognition, Journal of Neural Engineering 20 (1) (2023) 016050

    2023

  27. [35]

    Holobar, M

    A. Holobar, M. A. Minetto, D. Farina, Accurate identification of motor unit discharge patterns from high-density surface EMG and validation with a novel signal-based performance metric, Journal of neural engineering 11 (1) (2014) 016008

    2014

  28. [36]

    M. Han, O. Özdenizci, Y . Wang, T. Koike-Akino, D. Erdo ˘gmu¸ s, Disentangled adversarial autoencoder for subject-invariant physiological feature extraction, IEEE signal processing letters 27 (2020) 1565–1569

    2020

  29. [37]

    M. Han, O. Özdenizci, T. Koike-Akino, Y . Wang, D. Erdo˘gmu¸ s, Universal physiological representation learning with soft-disentangled rateless autoencoders, IEEE journal of biomedical and health informatics 25 (8) (2021) 2928–2937

    2021

  30. [38]

    Jiang, X

    X. Jiang, X. Liu, J. Fan, X. Ye, C. Dai, E. A. Clancy, M. Akay, W. Chen, Open access dataset, toolbox and benchmark processing results of high-density surface electromyogram recordings, IEEE Transactions on Neural Systems and Rehabilitation Engineering 29 (2021) 1035–1046

    2021

  31. [39]

    Y . Yuan, C. Dai, J. Fan, C. Chou, J. Liu, X. Jiang, Training explainable and effective multi-DoF EMG decoder using additive 1-dof EMG, IEEE Transactions on Medical Robotics and Bionics 6 (3) (2024) 1212–1219

    2024

  32. [40]

    C. Dai, B. Bardizbanian, E. A. Clancy, Comparison of constant-posture force-varying EMG-force dynamic models about the elbow, IEEE Transactions on Neural Systems and Rehabilitation Engineering 25 (9) (2016) 1529–1538

    2016

  33. [41]

    Y . Yuan, J. Liu, C. Dai, X. Liu, B. Hu, J. Fan, Exploring pattern-specific components associated with hand gestures through different sEMG measures, Journal of NeuroEngineering and Rehabilitation 21 (1) (2024) 1–13

    2024

  34. [42]

    Jiang, X

    X. Jiang, X. Liu, J. Fan, X. Ye, C. Dai, E. A. Clancy, D. Farina, W. Chen, Optimization of HD-sEMG-based cross-day hand gesture classification by optimal feature extraction and data augmentation, IEEE Transactions on Human-Machine Systems 52 (6) (2022) 1281–1291

    2022

  35. [43]

    D. P. Kingma, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  36. [44]

    P. J. Rousseeuw, Silhouettes: a graphical aid to the interpretation and validation of cluster analysis, Journal of computational and applied mathematics 20 (1987) 53–65. 20

    1987

  37. [45]

    J. Li, X. Jiang, J. Fan, Y . Geng, F. Jia, C. Dai, Deep end-to-end transfer learning for robust inter-subject and inter-day hand gesture recognition using surface EMG, Biomedical Signal Processing and Control 100 (2025) 106892

    2025

  38. [46]

    J. Li, P. Kang, T. Tan, P. B. Shull, Transfer learning improves accelerometer-based child activity recognition via subject-independent adult-domain adaption, IEEE Journal of Biomedical and Health Informatics 26 (5) (2021) 2086–2095. 21

    2021

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.