acceptodds
Under review as a conference paper at ICLR 2027

Neuromotor Hierarchy Network: Physiological Inductive Biases for Robust Generalization in sEMG Decoding

Abstract

Surface electromyography (sEMG) provides a wearable, noninvasive interface to neuromuscular activity for movement decoding and human-computer interaction. Population-scale decoding remains difficult because the relationship between sEMG and neuromuscular activity varies across users and sessions, while task-relevant dynamics span channels and multiple timescales. Learning waveform-to output mappings from task labels leaves the distinction between recording variability and coordinated motor activity implicit. We introduce the Neuromotor Hierarchy Network (NHN), which learns a compact latent neuromotor state from task supervision to represent task-relevant neuromuscular coordination. NHN constructs this latent state through a hierarchy inspired by neuromotor organization. It adapts recording statistics while preserving relative intensity. Its spatiotemporal encoder uses parameter-efficient channel interactions and modulates features with multi-timescale history. The resulting features yield candidate activations of learned motor primitives, which are temporally integrated and continuously weighted to form the state. Theoretical analysis characterizes the efficiency, temporal behavior, and optimization of NHN's core mechanisms. We evaluate the architecture for both continuous hand-pose estimation on emg2pose and touch-typing recognition on emg2qwerty. On emg2pose, NHN reduces user-averaged angular error by 0.52% to 2.84% across all three generalization splits in both Regression and Tracking relative to Hadidi et al.'s best task-specific variants, using 48.42% to 48.51% fewer parameters. On emg2qwerty, NHN reduces beam-search character error rate by 19.40% zero-shot and 30.42% after fine-tuning relative to SplashNet-Upscale, using 65.86% fewer parameters. Physiology-guided inference of a latent neuromotor state supports parameter-efficient sEMG decoding. Code: https://anonymous.4open.science/r/nhn-review-artifact-B7C1/README.md.

Primary area:
applications to neuroscience & cognitive science
Keywords:
surface electromyography, physiological inductive biases, representation learning, cross-user generalization, cross-session generalization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_pgErOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes NHN, a compact architecture for sEMG decoding inspired by neuromotor organization. It combines adaptation of recording statistics with preserved intensity cues, structured feature mixing, history-dependent feature modulation, and temporal integration and graded allocation of nonnegative candidate drives. The architecture is evaluated on emg2pose for hand-pose estimation and on emg2qwerty for typing recognition, with separately trained models for each task. NHN’s main results are averaged over five training seeds and show improved accuracy over several established baselines with fewer parameters. The paper also presents component ablations on emg2pose and analyses of the learned representations in support of the design.

Critical strengths
  1. NHN improves accuracy on two quite different tasks while using substantially fewer parameters than the main convolutional baselines. On emg2qwerty, zero-shot beam-search CER on the test split drops from 35.67% for SplashNet-Upscale to 28.75%, and fine-tuned beam-search CER drops from 5.51% for its shared variant to 3.83%, with 0.88 million parameters instead of 2.58 million. On emg2pose, NHN has the lowest reported AE and LD among the compared models across all three splits for both Regression and Tracking, with roughly half the parameters of vemg2pose and the Hadidi et al. variants. NHN’s results are consistent across five training seeds. Although the distilled Transformers achieve lower greedy CER on typing, the balance between accuracy and model size is a substantive strength of NHN.
  2. The main design choices are backed by controlled ablations. On emg2pose official-small, replacing structured mixing with conventional TDS mixing raises Tracking AE by 1.11° even though it adds parameters, and removing adaptive gain or FIR integration raises it by 0.82° and 0.79°, respectively. Disabling waveform conditioning and removing the relative-intensity channels increase AE by 0.61° and 0.69°, respectively. Each of these changes increases AE in all three seeds. In the Regression backbone comparisons, NHN also outperforms TDS in every seed at both matched LSTM hidden widths. These comparisons give me reasonable confidence that the proposed operations contribute to pose accuracy.
Critical weaknesses
  1. The full-data comparison with the strongest pose baselines leaves some uncertainty about the source of the improvement. NHN uses a 300-epoch schedule with time warping and amplitude scaling, while the results for Pos-ST and Vel-MT are taken from prior work. The authors' ablations show that training choices affect performance on official-small, but no full-data comparison with matched training conditions is reported. This matters because NHN's Tracking AE advantage over Pos-ST is 0.05° to 0.09° across the three splits. The controlled backbone comparisons on official-small are helpful, but do not fully resolve this uncertainty. I therefore place more weight on the parameter reduction than on the small accuracy advantage.
  2. The component ablations provide useful evidence on emg2pose, but no retrained component ablations are reported for typing. The allocation intervention shows that the trained typing decoder relies on the learned allocation pattern, but does not establish how a model trained without that operation would perform. The overall typing gains are convincing, while the contributions of the individual components to those gains remain less clear.
Questions for authors
  1. Could you clarify the main differences in training budget and augmentation between NHN and the strongest pose baselines? If any full-data comparisons under matched conditions are already available, they would help interpret the small accuracy differences in Table 1.
  2. Are any retrained typing ablations already available for temporal integration or graded allocation? The fixed-model allocation intervention is useful, but does not show how a model trained without graded allocation would perform.
4 comments
AuthorsOpenReview ↗

Response to pgEr Part 1 of 2

Thank you for your careful review and constructive feedback. We appreciate your recognition of the accuracy and parameter efficiency across both tasks and the existing Pose controls. We added full-data training comparisons and retrained Typing ablations to address both remaining concerns.

W1 and Q1. Training differences and controlled full-data comparisons

We agree that parameter efficiency is the clearest advantage in the original Pose comparison, where the accuracy gains are small (Table 1). Typing also shows a clearer CER advantage with fewer parameters (Table 2). To address the training concern, we added ten full-data runs in Appendix E.10 (Tables 18 and 19): five for Pos-ST Tracking and five for Vel-MT Regression. They use Meta's emg2pose implementation and retain each baseline's task formulation.

Table 18 lists the changes: 300 rather than 150 epochs, time warping from 0.85 to 1.18, channelwise amplitude scaling from 0.80 to 1.20, and a 256-wide prediction head. It also aligns the window, AdamW schedule, batch size, weight decay, and clipping. Amplitude scaling is applied once at the baseline input. The two heads have 0.882M and 0.887M parameters, close to NHN's 0.886M. The following results are from Appendix E.10, Table 19 and report means over five runs. Each entry is angular error (AE) in degrees / landmark distance (LD) in millimeters. Lower is better.

Task and modelUserStageUser × Stage
Tracking matched Pos-ST7.36 / 9.7010.49 / 14.1410.39 / 14.40
Tracking NHN7.30 / 9.5510.12 / 13.5910.16 / 14.03
Regression matched Vel-MT11.69 / 14.9914.30 / 19.0414.77 / 20.19
Regression NHN11.38 / 14.4213.78 / 18.2114.22 / 19.26

NHN retains lower mean errors in every condition after these changes. The participant-level comparisons in Appendix E.1.1, Table 4 have p < 0.01. These results provide additional support for the comparison after aligning the major training differences and closely matching head capacity.

Complementary NHN adaptation. We also added Appendix E.11, Table 20, adapting Hadidi's training ideas to NHN. The table below gives user-weighted AE/LD, with NHN's main setting from Appendix E.9, Table 17. NHN values are means over five runs.

TaskHadidi published bestNHN mainNHN adapted
TrackingPos-ST: 9.27 / 12.539.20 / 12.399.17 / 12.36
RegressionVel-MT: 13.37 / 17.6113.12 / 17.3013.23 / 17.49

NHN's errors change only slightly and remain below the best published task-specific baselines. E.10 therefore tests fairness under matched baseline training, while E.11 tests NHN's sensitivity to the reverse adaptation.

Broader performance and efficiency. We also added Appendix E.13 and Table 22 with seven gesture datasets, each using a separately trained NHN model. The table gives NHN accuracy and macro area under the receiver operating characteristic curve (AUROC), both in percent.

DatasetAccuracyMacro AUROC
MyoArmband76.2592.00
UCI EMG81.7595.05
Ninapro DB541.2676.28
Ninapro DB236.1273.61
Ninapro DB313.2052.94
MCS77.0693.24
FlexWear-HD85.0997.29

Table 22 shows higher accuracy than the best published baseline on five of seven datasets, with gaps of at most 0.84 percentage points on the other two. Macro AUROC exceeds the published best on two datasets. NHN averages 0.77M parameters and 0.066 network GFLOPs per 250 ms window.

Manuscript locations: Appendices E.10/Tables 18 and 19, E.11/Table 20, E.1.1/Table 4, and E.13/Table 22.

AuthorsOpenReview ↗

Response to pgEr Part 2 of 2

W2 and Q2. Retrained Typing integration and allocation ablations

Yes. We added six Typing controls trained from scratch with three seeds each (new Appendix E.12, Table 21). The complete table below includes removal of finite impulse response (FIR) integration, removal of allocation, and four additional controls. All follow Full NHN's protocol apart from the specified intervention. Entries are mean character error rate (CER, %) across three seeds. TDV and TDT denote test-domain Validation and Test, evaluated over complete sessions.

ConfigurationTDV greedyTDV beamTDT greedyTDT beam
No measurement adaptation42.9433.7542.6533.31
Conventional TDS encoder41.1431.6340.8631.73
Linear bottleneck39.2329.3539.2329.30
Independent gating39.6630.2739.3629.89
FIR identity39.8330.4039.3429.70
No allocation39.0329.6838.7729.17
Full NHN38.9529.2438.7028.80

Full has the lowest mean CER in all four columns. Following Meta's participant-level approach, we use Wilcoxon signed-rank tests after averaging matched seeds within each participant. Both FIR and allocation removals increase TDT beam CER at p < 0.05. All six TDT beam comparisons favor Full at this threshold. Table 21 also reports standard deviations across seeds and individual statistical annotations.

These retrained controls test whether each component helps decoding. NHN draws inspiration from physiological organization rather than simulating its biological mechanisms. The new Appendix A and Figures 4 to 6 test that link. Temporal integration retains sustained activity, while priorities concentrate weak drives and broaden participation for stronger input (A.2, Figure 4e/f). Figures 5 and 6 show repeatable movement relationships and association with classical coordination activity (A.3/A.4).

Manuscript location: Appendix E.12, Table 21 gives the three-seed results and participant-level tests. Appendix A, Figures 4 to 6 (pp. 21, 23, 24), provides the complementary physiological organization analyses.

Thank you for the clarifications and the extra experiments. That addresses my concerns.

Thank you for your thoughtful feedback. Your questions helped us strengthen the paper, and we’re glad the clarifications and additional experiments addressed your concerns

Reviewer_F1szOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the Neuromotor Hierarchy Network (NHN), a physiology-inspired architecture for sEMG decoding. NHN combines measurement adaptation, intensity preservation, efficient local-global feature interactions, multi-timescale contextualization, and a compact representation constructed through non-negative drives, temporal integration, and graded allocation. Experiments cover hand-pose estimation and typing recognition, demonstrating competitive accuracy with substantially fewer parameters and lower reported computational cost.

Critical strengths
  1. Strong accuracy and efficiency results across two tasks, improvements on both emg2pose and emg2qwerty support the usefulness of the architecture beyond one output formulation. The typing results, including subject-specific fine-tuning, are particularly encouraging.
  2. Substantial experimental support. The main experiments use five training seeds. The appendix includes component ablations, alternative prediction heads, capacity controls, training-objective comparisons, and tracking evaluations extending to 60 seconds and full valid recording intervals.
  3. Careful qualification of the physiological interpretation. The paper explicitly distinguishes learned task representations from genuine physiological sources. The mathematical analysis also makes the properties of individual operators more transparent.
Critical weaknesses
  • The strongest pose comparisons are not fully controlled. Improvements over the strongest published pose variants are relatively small, while NHN uses revised training windows, augmentation, and optimization settings. Matched backbone and head comparisons are useful, but are conducted primarily in the reduced-data setting. Full-data retraining under a common recipe would better isolate architectural gains.
  • The central representation mechanism needs broader validation. Comparable typing ablations would strengthen the claim that the same hierarchy explains both tasks’ improvements. Replacing allocation weights in a frozen model establishes reliance on those weights, but does not establish superiority over a generic learned bottleneck or gating mechanism.
  • Appendix D.1 reports sensitivity comparisons on official-small test partitions and describes these comparisons as guiding the adopted configuration. It is unclear whether test performance influenced selection of architectural hyperparameters and whether those evaluation data overlap the final benchmark.
  • Practical efficiency remains incompletely established. FLOPs accounting excludes operations such as whitening eigendecomposition and normalization, and measured end-to-end latency is not reported. The efficient SSM approach discussed in Related Work is also excluded from quantitative comparisons because of protocol differences.
Questions for authors
  1. Can the strongest pose baselines be retrained with NHN’s training recipe and matched prediction-head capacity on the full dataset?
  2. Can the principal integration and allocation ablations be repeated on emg2qwerty?
  3. Were all architectural choices selected using validation data only? Please clarify the relationship between official-small development experiments and the final test partitions.
  4. What are the complete preprocessing-plus-network latency and memory costs, and how does NHN compare with an efficient SSM under an aligned protocol?
6 comments
AuthorsOpenReview ↗

Response to F1sz Part 1 of 4

Thank you for your careful review and constructive feedback. We appreciate your recognition of the two-task results, extensive controls, and the distinction between task-learned representations and physiological sources. We have added full-data matched training, retrained Typing controls, a clearer selection protocol, and measured Pose inference costs.

W1 and Q1. Fair comparisons through full-data matched training

Yes. We added ten full-data runs with matched training conditions and head capacity: five for Hadidi Pos-ST in Tracking and five for Vel-MT in Regression (new Appendix E.10, Tables 18 and 19). The implementation starts from Meta's emg2pose code and data protocol. Each baseline retains its task formulation and model design. We align the main training and capacity differences identified.

The changes listed in Table 18 are 300 rather than 150 epochs, time warping, channel-wise amplitude scaling, NHN's window and optimization settings, and a two-layer LSTM head of width 256 rather than 512. The resulting baseline heads contain 0.882M and 0.887M parameters, compared with 0.886M for NHN. Amplitude scaling is applied once at the raw input because these baselines do not contain NHN's measurement adaptation module. Table 18 specifies the mapping and all aligned settings.

Baseline training aligned with NHN (E.10, Table 19). The first table reports means over five runs. Both tables give angular error (AE) in degrees / landmark distance (LD) in millimeters. Lower is better.

Task and modelUserStageUser × Stage
Tracking matched Pos-ST7.36 / 9.7010.49 / 14.1410.39 / 14.40
Tracking NHN7.30 / 9.5510.12 / 13.5910.16 / 14.03
Regression matched Vel-MT11.69 / 14.9914.30 / 19.0414.77 / 20.19
Regression NHN11.38 / 14.4213.78 / 18.2114.22 / 19.26

NHN retains lower mean errors in every condition after these changes. The participant-level comparisons in Appendix E.1.1, Table 4 have p < 0.01. These results provide additional support for the comparison after aligning the major training differences and closely matching head capacity.

NHN adapted to the baseline's training ideas (E.11, Table 20). We added Appendix E.11 and Table 20 for this complementary experiment. The second table compares user-weighted AE/LD with the best published baseline for each task and includes NHN's main setting from Appendix E.9, Table 17 as a reference. NHN values are means over five runs. The strongest published baselines use different task formulations: Pos-ST for Tracking and Vel-MT for Regression. NHN uses a unified training framework across both tasks.

TaskHadidi published bestNHN mainNHN adapted
TrackingPos-ST: 9.27 / 12.539.20 / 12.399.17 / 12.36
RegressionVel-MT: 13.37 / 17.6113.12 / 17.3013.23 / 17.49

NHN's errors change only slightly under the adapted setting and remain below the best published baseline for each task. E.10 tests the baselines under NHN's main training conditions. E.11 tests NHN under an adaptation of the baseline's training ideas. NHN retains its advantage in both comparisons.

Evidence beyond Pose and Typing. To complement these controlled comparisons, we added Appendix E.13 and Table 22 with seven gesture datasets. Each uses a separately trained NHN model. The table gives NHN accuracy and macro area under the receiver operating characteristic curve (AUROC), both in percent.

DatasetAccuracyMacro AUROC
MyoArmband76.2592.00
UCI EMG81.7595.05
Ninapro DB541.2676.28
Ninapro DB236.1273.61
Ninapro DB313.2052.94
MCS77.0693.24
FlexWear-HD85.0997.29

Table 22 shows higher accuracy than the best published baseline on five of seven datasets, with gaps of at most 0.84 percentage points on the other two. Macro AUROC exceeds the published best on two datasets. NHN averages 0.77M parameters and 0.066 network GFLOPs per 250 ms window.

Manuscript locations: Appendices E.10/Tables 18 and 19 for matched training, E.11/Table 20 for NHN adaptation, E.1.1/Table 4 for paired tests, and E.13/Table 22 for the seven-dataset extension.

AuthorsOpenReview ↗

Response to F1sz Part 2 of 4

W2 and Q2. Retrained Typing ablations and generic alternatives

Yes. We added six Typing controls, each trained with three seeds, in Appendix E.12, Table 21. They directly test component value and generic alternatives. A linear bottleneck replaces the entire latent neuromotor state construction with a 64-to-32 affine map. Independent gating multiplies this map by a separate sigmoid gate. Two other controls remove finite impulse response (FIR) integration or allocation. The remaining controls remove measurement adaptation or use a conventional time-depth separable (TDS) encoder.

We reuse the corresponding Full checkpoints and keep the training and evaluation protocol fixed apart from each intervention. The table reproduces all six controls and Full from Table 21. Entries are mean character error rate (CER, %) across three seeds, with complete-session greedy and beam decoding. TDV and TDT denote test-domain Validation and Test. Lower is better.

ConfigurationTDV greedyTDV beamTDT greedyTDT beam
No measurement adaptation42.9433.7542.6533.31
Conventional TDS encoder41.1431.6340.8631.73
Linear bottleneck39.2329.3539.2329.30
Independent gating39.6630.2739.3629.89
FIR identity39.8330.4039.3429.70
No allocation39.0329.6838.7729.17
Full NHN38.9529.2438.7028.80

Full has the lowest mean CER in all four columns. Following Meta's participant-level approach, we compare Full and each control with Wilcoxon signed-rank tests. We first average matched seeds within each of the eight participants. All six TDT beam comparisons favor Full at p < 0.05. Table 21 also gives standard deviations across seeds and the tests for the other columns.

NHN uses physiological organization as inspiration for a learned representation, not as a biological simulation. The new Appendix A and Figures 4 to 6 test this design link. Filters retain sustained drives, and learned priorities broaden participation as the same input becomes stronger (A.2, Figure 4e/f). The state preserves movement relationships across recordings (A.3, Figure 5) and remains associated with classical coordination activity (A.4, Figure 6). These analyses explain the organization, while the retrained controls test its value for decoding.

Manuscript locations: Appendix E.12, Table 21 for retrained controls and participant-level tests. Appendix A, Figures 4 to 6 (pp. 21, 23, 24), provides the complementary organization analyses. The three-seed Full reference here is 28.80% TDT beam CER.

AuthorsOpenReview ↗

Response to F1sz Part 3 of 4

W3 and Q3. Validation selection and official-small provenance

Yes. Architecture and hyperparameters were selected using official-small Validation data. We train candidates on small Train and select them using small Validation. We then fix the configuration for full-data training and evaluation. The revised E.1.2 states this selection path explicitly.

We added Table 6 in Appendix E.1.2 to show that each Train, Validation, or Test partition in metadata_small.csv is drawn from the corresponding partition in metadata.csv. The table counts recording files, with stages denoting movement categories.

ScopePartitionRecordingsParticipantsSessionsStages
FullTrain17,13615861623
FullValidation1,950155729
FullTest6,16717867829
SmallTrain85715745723
SmallValidation98154327
SmallTest30812921628

Table 6 shows that small Validation retains all 15 full Validation participants and has no participant or recording overlap with full Test. Small Test contains 308 of the 6,167 full Test recordings (4.99%), retaining the User, Stage, and User × Stage conditions. Small Test thus provides a subset evaluation within the full benchmark.

Small Train still covers 157 of 158 full Train participants and all 23 training movement categories, making repeated architecture comparisons feasible with broad participant and movement coverage. The ablation tables report small Test outcomes. These ablations establish component effects at the reduced training scale. The matched baseline experiments in Appendix E.10, Tables 18 and 19 provide complementary evidence at the full training scale.

Manuscript locations: Appendix E.1.2 and the new Table 6 give the selection process, Full/small coverage, and partition relationship. Appendix D.1 gives the main data protocol.

Reviewer_TBA1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a physiologically inspired deep learning model to decode finger movements from sEMG signals. The main contribution of the work stands in the definition of a backbone aimed at extracting a task specific neuromotor latent representation that can be used then to feed a prediction head. The method has been evaluated exclusively offline on two public datasets: emg2pose for the decoding of finger kinematics, and emg2qwerty for the classification of character typing. Validation includes comparison with baseline models (including the ones originally proposed with the datasets), ablation analysis, and a simulated application of text writing combining the model output with a GPT language model.

Critical strengths

The first critical strength is related to the rigor in model description for reproducibility. The paper details every component with complete mathematical rigor, backing key architectural properties with theoretical proofs (such as bounded history aggregation and linear priority convergence). Reproducibility is supported by exact specifications of all layer dimensions, kernel sizes, loss weightings, and learning schedules, alongside an open-source repository containing complete code, configurations, and over 50 trained checkpoints spanning multiple random seeds.

The second strength regards the results, not much on the performance but rather the reduction in model complexity with respect of baseline models. By incorporating domain-specific physiological inductive biases—such as compact motor primitive representations and low-rank local-global feature mixing—the proposed approach achieve comparable performance with a smaller footprint compared to the top baseline variants of emg2pose, resulting in an approximate 35.5% reduction of FLOPs.

Critical weaknesses

A first major concern regards the claim on the "physiology-inspired" network hierarchy. I would be cautious in this claim as for this to be true it would be fundamental to quantitively demonstrate that the learned representation is similar to the representation obtained with well-validated methods whose outcome are known to be physiologically meaningful, an analysis which however is totally missing. For example, saying that the reduction of the muscle dimension C to the reduced dimension of N non-negative drives is inspired by muscle coordination through synergies is quite strong. Any dimensionality reduction algorithm could be seen as "physiology-inspired" in this sense. To effectively demonstrate it, it would have been necessary to understand if the organization of the learned drives shares some similarities with the muscle synergies identified for example by a simple Non-Negative Matrix Normalization. Thus, while the methodology is well presented mathematically, the explanation of its bio-inspired motivations which should be the major contribution of the work as claimed by the author is weak.

The second weakness regards the soundness of the results. Despite the better results of the proposed approach (e.g. as shown in Table 1), the benefit in performance of the proposed approach with respect to the baselines is unclear, in particular because not test has been performed to evaluate if the improvement is statistically significant. I strongly doubt it, considering for example that the improved AE in Table is less than 0.3° with a standard deviation of more than 1° (similar considerations can be drawn also for the other metrics of the emg2pose, as well as the results of emg2qwerty). It is true that the proposed approach is better in efficiency (num of parameters, FLOPs) with respect to the best performing baselines. However, it would have been fundamental to verify if it is statistically better than emg2pose or NeuroPose which are more or equivalently efficient than the proposed approach. Also for the ablation study, it would be fundamental to understand if the improvements are statistically significant. Indeed, if the improvement is not significant an ablated version could be better than the full NHN as it would be more efficient (e.g., less parameters) while not significantly losing performance. Similar consideration about statistical analysis is valid also for the results of Table 5 and of Table 6. In the latter in particular, the difference in efficiency (num of parameters) between the models is quite limited, and also the difference in performance seems negligible in a statistical sense. This aspect linked to the absence of an effective real-world application (real-time performance, robustness to uncertainties, etc.) significantly limits the scientific contribution of the work.

Questions for authors
  • Title: Considering the very specific dataset used for the validation (muscles only from the wrist), I would specify this in the title to avoid unintended claim that the approach has been verified in general sEMG decoding task (e.g. upper limb or lower limb motion)
  • The related work literature on muscle synergy-based EMG control should be strengthened [a]. For example, some recent works on the use of deep learning models for cross-subject synergy-based decoding are missing, such as [b,c]. Additionally, other works faced the problem of task-independent synergy extraction and decoding, which are also not mentioned, such as [d,e,f].

[a] Ison, Mark, and Panagiotis Artemiadis. "The role of muscle synergies in myoelectric control: trends and challenges for simultaneous multifunction control." Journal of neural engineering 11.5 (2014): 051001. [b] Shi, Kecheng, et al. "MCSNet: channel synergy-based human-exoskeleton interface with surface electromyogram." Frontiers in Neuroscience 15 (2021): 704603. [c] Mu, Tong, et al. "Muscle Synergy-Informed Feature Representation with Semantic Alignment for Continuous Multi-Joint Torque Estimation." Biomedical Signal Processing and Control 128 (2026): 111307. [d] Buongiorno, Domenico, et al. "Task-oriented muscle synergy extraction using an autoencoder-based neural model." Information 11.4 (2020): 219. [e] Liu, Ruofan, et al. "From pose to muscle: Multimodal learning for piano hand muscle electromyography." Advances in Neural Information Processing Systems 38 (2026): 115558-115587. [f] Zhang, Hairong, et al. "GenJAPNet: A Generalizable Joint Angle Prediction Network with Non-Redundant Muscle Synergy Features for Lower-Limb Exoskeletons." 2026 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2026.

  • If I understood correctly, the model assumes that motion to be decoded for a specific task k (e.g., Angle regression or Typing classification) have always the same length of L_k frames. This is quite a strong limitation for an effective application of the proposed approach in a real-world application, as it is difficult to ensure that the number of frames in a motion of a given task are constant. Please, explain better or comment on this aspect.
  • The meaning of D(x) in eq. (3) is not explained.
  • What are the dimensions of U_in, u_mid, and u_out in lines 204-207?
  • What does function LN mean in eq (4) and (5)?
  • Lines 221-222: What does it mean "when need"? Which are the cases in which the final projection is need?
  • Provide more details on the learned pointwise affine map g_d, as currently its function is not clear nor it is clear what it is learned.
  • Since the implementation of the Non-negative candidate drives is inspired by the concept of muscle synergies, it would be better to provide a clearer theoretical explanation of muscle synergies and their link to the proposed approach.
  • What does M_f in eq (9) represent?
  • How are the non-negative kernels k_m defined?
  • Provide a clear theoretical explanation of the Henneman's principle and clearly explain the link with the proposed methodology.
  • The legend of figure 3b is missing the meaning of the empty triangle markers.
  • It is not clear how much the performance are dependent to the selection of the 5-s window. Indeed, assuming that a single motion is always of length 5-s is quite strong assumption. Is this the case?
9 comments
AuthorsOpenReview ↗

Response to TBA1 Part 1 of 5

Thank you for your careful review and constructive feedback. We appreciate your recognition of the mathematical specification, reproducibility, and reduction in model complexity. We address both weaknesses and each of your questions below.

W1. Physiological inspiration and classical coordination

NHN is inspired by physiological organization. It learns a compact latent neuromotor state for decoding rather than imitating biological mechanisms. We added Appendix A and three new figures (Figures 4 to 6) to test this link. The analyses show how the design constraints shape activity and whether the learned components retain relationships with movement and classical muscle coordination.

Coordinated muscle activity motivates structured drives. Hand grasping combines recurring activation patterns [1]. C counts sEMG channels, not muscles. NHN maps C=16 channels to D=64 encoded features, then N=32 drives (Sections 3 and 4). N therefore exceeds C. The inspiration concerns nonnegative activity, temporal integration, and learned participation priorities, rather than direct C-to-N compression (Section 3.3 and Appendix A.1).

Spatial muscle activity motivates local and shared processing. Forearm muscle activity produces spatially distinct sEMG patterns [2]. Motor neuron pools also receive common input [3]. Measurement adaptation reduces persistent channel scaling while preserving a cue for local activity changes (Appendix A.2, Figure 4a/b). The spatiotemporal encoder combines local and distributed responses and uses recent activity to modify its current response (Figure 4c/d). These tests show that the modules separate measurement scale from activity and combine information across channels and time.

Contraction dynamics motivate temporal integration. Human motor units contract over finite and varied timescales [4]. NHN reflects this persistence by integrating recent positive drives. Its learned filters preserve sustained activity and reduce fluctuations (Appendix A.2, Figure 4e). Temporal variation falls by 17.6% on recorded trajectories. This supports the intended accumulation of activity over time.

Recruitment motivates a changing range of participation. Motor units have different recruitment thresholds, so more units participate as voluntary force increases [4,5]. NHN uses learned priorities to give its components different participation preferences. For the same activity pattern, weak input concentrates weights on preferred components, while strong input spreads them more widely (Appendix A.2, Figure 4f, lower). Setting priorities to zero removes this response. The upper plot also shows that a weaker component can contribute more because of its learned priority. Thus, participation depends on both activity strength and preferences learned for decoding.

Movement relationships persist across recordings. Recurring coordination patterns [1] motivate this test. NHN preserves relationships with joint motion more consistently than the tested dimensionality reduction baselines (Appendix A.3). Figure 5a/b illustrates these relationships and summarizes participant results. Figure 5c shows all recording pairs, including one exception. Figure 5d extends the comparison across classical methods: NHN has more repeatable relationships, even when some baselines have stronger correlations within a recording.

The representation remains related to classical muscle coordination. We compare the drives and state with nonnegative matrix factorization (NMF) activities [1]. Both remain more closely associated with correctly timed NMF activity than with a time shift control after accounting for shared amplitude (Appendix A.4, Figure 6a). Both have stronger NMF association than four dimensionality reduction methods with 32 components (Figure 6b). These results support a link to coordination in a representation learned for decoding. The new Typing ablations also favor Full NHN over trained linear bottleneck and independent gating controls on Test beam decoding (Appendix E.12, Table 21). These complementary tests connect the physiological motivation to module behavior, learned organization, and decoding value.

References [1] Scano et al. Front. Neurorobot. 12:57, 2018. Muscle synergy analysis of hand grasps. [2] Gallina & Botter. Front. Physiol. 4:367, 2013. Spatial localization of forearm EMG. [3] Negro et al. J. Physiol. 594:5491, 2016. Common synaptic input. [4] Milner-Brown et al. J. Physiol. 230:359, 1973. Human motor unit recruitment. [5] Henneman. Science 126:1345, 1957. Neuron size and discharge.

Manuscript location: Appendix A.1 to A.5, Figures 4 to 6 (pp. 21, 23, 24), and E.12/Table 21. Section 5.2 introduces the new analyses.

AuthorsOpenReview ↗

Response to TBA1 Part 2 of 5

W2. Statistical reliability and practical evidence

Paired evidence. Following Meta's emg2pose study (Salter et al., 2024), we added participant-level Wilcoxon signed-rank tests to the baseline and ablation comparisons (Appendix E.1.1, Tables 4 and 5, Appendices E.1.3 to E.1.6, Tables 7 to 10, and Appendix E.12, Table 21). The test uses each participant's difference between models. NHN's standard deviations (SDs) in Table 1's split columns and Table 2 describe variation across participants, after averaging seeds within each participant. They do not measure variation across training runs. Seed SDs are reported separately in Appendix E.9, Table 17, and Tables 7 to 10 and 19 to 21. A mean gain smaller than the participant SD can still be significant in a paired test. We have clarified the SD definitions in the relevant captions.

The new Appendix E.1.1 reports the baseline results and statistical tests. Table 4 includes emg2pose, vemg2pose, NeuroPose for Regression, and our Hadidi reproductions. NHN has lower AE and LD with p < 0.01 in each of the three generalization conditions. Table 5 reports p < 0.05 for the evaluated Typing baseline comparisons on test-domain Validation and Test with both decoders. Tables 7 to 10 give the individual tests for Pose components and prediction heads.

Removing either finite impulse response (FIR) integration or allocation significantly increases test-domain Test beam CER in the new Typing ablation section (p < 0.05, Appendix E.12, Table 21).

Training fairness. The new Appendix E.10, Tables 18 and 19 report ten full-data runs of the strongest Pose baselines with matched training conditions and head capacity. NHN retains lower AE and LD in all three conditions. Appendix E.11, Table 20 tests the complementary direction by adapting the baseline's training ideas to NHN. The errors change only slightly. These experiments address training differences in both directions.

Practical cost and perturbations. The new Appendix E.2, Table 11 measures preprocessing, the network, transfers, and output alignment together. For a five-second scoring interval, NHN has a mean processing latency of 80.15 ms for Regression and 87.21 ms for Tracking on an RTX 5090. Peak allocated CUDA memory is 43.55 MiB. This is less than 1.8% of the scored signal duration. The controlled probes in Appendix A.2, Figure 4a/b test whether adaptation reduces measurement-scale changes while retaining local activity changes.

Manuscript location: Appendix E.1.1/Tables 4 and 5, E.1.3 to E.1.6/Tables 7 to 10, E.9, E.12/Table 21, E.2/Table 11, and E.10/E.11, Tables 18 to 20. Appendix A.2, Figure 4a/b/d (p. 21) shows the controlled scale and history responses. Question 14 addresses longer evaluation intervals.

AuthorsOpenReview ↗

Response to TBA1 Part 3 of 5

Q1. Title and evaluated scope

Our main tasks are hand pose and Typing from wrist sEMG (Section 4, Tables 1 and 2). We clarified this scope in the revised title: "Neuromotor Hierarchy Network: Physiological Inductive Biases for Generalizable Forearm sEMG Decoding".

Broader dataset evidence. We added Appendix E.13 and Table 22 with seven gesture datasets, training one NHN model per dataset. NHN accuracy / macro area under the receiver operating characteristic curve (AUROC), both in percent, are:

DatasetAccuracyMacro AUROC
MyoArmband76.2592.00
UCI EMG81.7595.05
Ninapro DB541.2676.28
Ninapro DB236.1273.61
Ninapro DB313.2052.94
MCS77.0693.24
FlexWear-HD85.0997.29

Table 22 shows higher accuracy than the best published baseline on five of seven datasets, with gaps of at most 0.84 percentage points on the other two. Macro AUROC exceeds the published best on two datasets. NHN averages 0.77M parameters and 0.066 network GFLOPs per 250 ms window.

Manuscript location: Revised title, Section 4, and the new Appendix E.13/Table 22.

Q2. Recommended related work

All six recommended references are now included in Section 2, under "Synergy-informed myoelectric representations." They cover muscle synergies (Ison and Artemiadis), MCSNet (Shi et al.), torque estimation (Mu et al.), task-oriented extraction (Buongiorno et al.), pose and muscle activity (Liu et al.), and joint-angle prediction (Zhang et al.). Appendix A.1 to A.4 also cites the relevant studies when explaining our design and analyses. We appreciate these suggestions, which helped clarify the link between classical coordination analysis and task-learned representations.

Manuscript location: Section 2, especially Synergy-informed myoelectric representations, and the relevant analyses in Appendix A.

Q4. Meaning of D(x)

D(x)\mathcal{D}(x) denotes channelwise standardization using cumulative channel means and standard deviations. This definition appears immediately before Eq. (3) on page 4 of the submitted manuscript. To make the operation clearer, we have added its explicit formula at this location in the revised manuscript. For channel cc and input sample nn after the initial warmup,

μc,n=1n∑i=1nxc,i,σc,n=max⁡(1n∑i=1nxc,i2−μc,n2,0),\mu_{c,n}=\frac{1}{n}\sum_{i=1}^{n}x_{c,i},\qquad \sigma_{c,n}=\sqrt{\max\left(\frac{1}{n}\sum_{i=1}^{n}x_{c,i}^{2}-\mu_{c,n}^{2},0\right)},

[D(x)]c,n=xc,n−μc,nσc,n+ε.\left[\mathcal{D}(x)\right]_{c,n}=\frac{x_{c,n}-\mu_{c,n}}{\sigma_{c,n}+\varepsilon}.

Each channel is centered and scaled using its own samples up to the current time, with ε=10−6\varepsilon=10^{-6} for numerical stability. It standardizes channels independently. The fixed reference scale s0s_0 is applied separately in Eq. (3). As specified in C.4, for the first 128 samples the implementation holds these statistics at their 128-sample estimates within the unscored left context.

Manuscript location: Section 3.1, immediately before Eq. (3), and Appendix C.4 for warmup settings. The new formula is marked in blue.

Q5. Encoder tensor dimensions

For encoder block r, u_in has D_r channels and L_in frames. Both u_mid and u_out have D_r channels and L_out frames. Section 3.2 now states these dimensions directly in the block definition.

Manuscript location: Section 3.2, encoder block definition.

Q6. Meaning of LN

Here, LN denotes layer normalization. Its expansion appears in Section 3 of the submitted manuscript, before Equations 4 and 5.

Manuscript location: Section 3, notation preceding Equations 4 and 5.

Q7. When the final projection is needed

The final projection is used when the last encoder block's width differs from the required output width D. Section 3.2 now states this condition in place of "when needed." Appendix C.4 lists the widths used for each task.

Manuscript location: Section 3.2. Appendix C.4 gives the task-specific widths.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Clarification of the paper title

We thank Reviewer TBA1 for suggesting that the title more clearly reflect the scope of the study. We revised the title to specify upper limb sEMG decoding and emphasize generalization. The revised title is: Neuromotor Hierarchy Network: Physiological Inductive Biases for Generalizable Forearm sEMG Decoding This title edit clarifies the scope while retaining the paper’s core contribution. The revised manuscript uses the title above.

AuthorsOpenReview ↗

Summary of Contributions and Revisions

We thank the reviewers for their constructive feedback and the Area Chair for their time. The Neuromotor Hierarchy Network (NHN) is a generalizable, low-parameter architecture for continuous hand pose estimation and typing recognition from forearm surface electromyography (sEMG).

  1. NHN generalizes with fewer parameters. Across all evaluated generalization settings, NHN achieves lower mean errors than the strongest task-specific Pose baselines with about 48% fewer parameters (Table 1). The lower errors persist after aligning training conditions and closely matching prediction head capacity (new Appendix E.10). For typing, NHN achieves lower mean character error rates than all tested SplashNet variants across all evaluated settings. SplashNet is a strong baseline designed for efficient sEMG typing, and NHN uses 66% fewer parameters than SplashNet-Upscale (Table 2).

  2. NHN’s learned behavior reflects its physiological inspiration. NHN connects the compact latent neuromotor state learned through task supervision to muscle contraction, recruitment, and coordination (Figure 3 and Figure 4). Temporal filters preserve sustained activity and damp fluctuations, consistent with the persistence of muscle contractions. Strengthening the same input pattern broadens component participation, resembling motor unit recruitment as neural drive increases. Components retain similar relationships with joint motion across recordings and show activity associated with classical muscle synergy activations, consistent with recurring muscle coordination (Figures 5 and 6).

  3. Competitive accuracy across recording systems with lower computational cost. Across seven additional gesture classification datasets, NHN achieves higher mean accuracy than the baselines reported in EMGBench on five datasets and comes within 0.84 percentage points of the highest reported baseline mean on the other two. Based on reported average model costs, NHN uses 64% to 93% fewer parameters and 67% to 98% less computation than these dataset-specific best baselines (new Appendix E.13, Table 22).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.