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

1 INTRODUCTION

• sEMG supports wearable movement and typing interfaces, but population-scale decoding must handle variation across users and sessions while remaining parameter-efficient.Population datasets pair wrist sEMG with hand kinematics or natural typing.Source: §1 · p. 1
• Prior work improves decoding through larger populations, temporal models, signal-statistics adaptation, shared weights, and cross-modal learning, but compact coordination representations remain open.The introduction identifies physiology as a guide for organizing such representations.Source: §1 · p. 1
• sEMG and hand kinematics share a neuromotor origin, while anatomy, recording conditions, and individual motor control vary their relationship.Muscle activation patterns vary in strength and timing; electrode and tissue effects shape the surface waveform.Source: §1 · p. 2
• NHN adapts recording statistics while preserving relative intensity, then encodes spatiotemporal activity and integrates non-negative candidate drives into a graded latent state.The encoder uses parameter-efficient channel interactions and multi-timescale activity histories.Source: §1 · p. 2
• NHN infers a compact, task-aligned latent neuromotor state intended to support generalization across users and sessions.The architecture is evaluated for continuous hand-pose estimation and discrete character-sequence recognition.Source: §1 · p. 2

2 RELATED WORK

• sEMG generalization work spans population-scale training, adaptation, disentanglement, rehearsal, pretraining, and architectures for gesture or typing decoding.The cited work also distinguishes intersubject transfer from adaptation and broadens demographic coverage.Source: §2 · pp. 2–3
• Physiological-signal models structure representations through sensor geometry, spatiotemporal or time-frequency features, pretraining, and local-data personalization.For task-supervised sEMG, the stated question is how physiological priors jointly guide measurement processing and activity representation.Source: §2 · p. 3
• Motor-state research spans motor-unit decomposition, aligned population recordings, dynamical embeddings, behavior-supervised representations, and synergy-based control.For sEMG, the identified challenge is inferring a compact task-relevant state from mixed surface measurements; NHN uses temporal integration and graded allocation under task supervision.Source: §2 · p. 3

3 METHOD: NEUROMOTOR HIERARCHY NETWORK

• NHN maps adapted sEMG through spatiotemporal encoding and latent-state construction to task-specific predictions for pose or typing.The state is formed by integrating and allocating candidate drives.Source: §3 · p. 3, §3.4 · p. 6

3.1 MEASUREMENT ADAPTATION WITH INTENSITY PRESERVATION

• The adapter adjusts recording-dependent waveform statistics while retaining relative intensity cues.Partial whitening blends channel-wise normalization and joint whitening; for pose, the waveform path also blends in cumulative channel-wise normalization.Source: §3.1 · p. 4

3.2 SPATIOTEMPORAL ENCODING OF ADAPTED SEMG

• Temporal convolutions and local-global mixing encode adapted sEMG, while multi-timescale causal histories reweight features to capture brief and sustained activity.The local-global path combines shared local interactions with a low-rank path connecting distant groups.Source: §3.2 · pp. 4–5

3.3 LATENT NEUROMOTOR STATE CONSTRUCTION FROM ENCODED FEATURES

• The state constructor projects features to non-negative candidate drives, integrates them over time, then uses drive-dependent graded allocation to form a compact state.Non-negativity prevents cancellation during integration; learned recruitment priorities and continuous weights shape allocation.Source: §3.3 · p. 5

3.4 TASK-SPECIFIC PREDICTION HEADS

• Task-specific heads map the latent state to joint-angle estimates with an autoregressive LSTM or to character probabilities with TDS and CTC.Source: §3.4 · p. 6

• Both tasks use independently trained NHN weights, a 64-wide encoder, 32 primitives, and causal processing of 16 sEMG channels sampled at 2 kHz.Pose uses one wrist; typing processes each wrist separately.Source: §4 · p. 6
• Pose evaluation uses User, Stage, and User×Stage held-out splits, with Regression and Tracking differing in whether the model receives an initial ground-truth pose.Evaluation uses 5-second windows and angular error and landmark distance metrics.Source: §4 · p. 6
• Typing evaluation tests pooled-data pretraining on held-out users both zero-shot and after subject-specific fine-tuning, reporting character error rate with greedy or beam-search decoding.Beam search uses a 6-gram language model; the state sequence is sampled at 100 Hz.Source: §4 · p. 6
• Pose training and typing pretraining use five random seeds, with participant scores averaged over seeds before reporting mean and standard deviation.Source: §4 · p. 6
• Pose baselines use published results under matching protocols, while typing compares against TDS ConvNet, SplashNet, and distilled Transformers; efficiency reports parameters and GFLOPs.Inference counts include prediction heads but exclude language-model and text-decoding costs.Source: §4 · p. 6

5.1 MAIN RESULTS

• Against Hadidi et al.’s best task-specific variants, NHN lowers user-level AE and LD with 48.42% to 48.51% fewer parameters and 35.55% to 35.64% less computation.The emg2pose comparison covers Regression and Tracking; Table 1 reports the split and session aggregation.Source: Table 1 · p. 7, §5.1 · p. 7
• Regression AE decreases by 1.44° on Stage versus 0.86° on User, and NHN’s advantage over vemg2pose is more pronounced when evaluation includes unseen movement stages.Session-equal aggregation preserves the advantage: AE decreases by 8.48% in Regression and 7.73% in Tracking.Source: §5.1 · pp. 6–7
• 19.40% zero-shot and 30.42% after fine-tuning are NHN’s relative beam-search CER reductions against SplashNet-Upscale and its Shared variant, respectively.NHN uses 65.86% fewer parameters and 74.99% less neural-network computation than SplashNet-Upscale.Source: §5.1 · p. 7, Table 2 · p. 8

5.2 LEARNED ORGANIZATION OF THE LATENT NEUROMOTOR STATE

• For pose, relative primitive activations overlap in their associations with joint angles across fingers, and the 12 largest allocation weights account for about half the total in the illustrated clip.Training increases all three adaptive-gain timescales but decreases mean integration lags.Source: §5.2 · p. 8, Figure 3 · p. 9
• For typing, same-key primitive-activation patterns are more similar than different-key patterns across all nine analyzed keys, while training shifts gain timescales differently than in pose.Typing’s short and long adaptive-gain timescales decrease, its intermediate timescale increases, and mean integration lags decrease; allocation remains unequal and spread across primitives.Source: §5.2 · p. 8, Figure 3 · p. 9
• In sampled validation data, uniform allocation raises pose AE by 3.94° and typing greedy CER by 42.12 percentage points versus learned allocation, despite unchanged total state activation.The result indicates that prediction heads rely on the distribution of activation across primitives, not only its overall strength.Source: §5.2 · p. 8

6 LIMITATIONS

• NHN’s learned components are task representations rather than genuine physiological sources, and evaluation uses wrist sEMG; device deployment and transfer to other sensor layouts remain untested.At longer fixed tracking horizons, errors are higher, though still below vemg2pose, and the gap between them is larger.Source: §6 · p. 9

7 CONCLUSION

• NHN infers a compact latent neuromotor state from wrist sEMG for hand-pose estimation and touch-typing recognition.The hierarchy is physiology-inspired, while task supervision shapes the state’s content.Source: §7 · p. 9
• Pose representations show distributed joint-angle associations, whereas typing representations show recurring key-specific primitive combinations and distinct temporal preferences.Allocation interventions indicate that decoding uses relative activation patterns beyond total intensity.Source: §7 · p. 9
• Theoretical analysis characterizes NHN’s computational mechanisms, and the reported generalization and efficiency results support physiology-guided representation structure with task-shaped content.Source: §7 · p. 9

AI USE STATEMENT

• Language models assisted with iterative manuscript revision and code writing and debugging.The authors reviewed AI-assisted text for technical accuracy and reviewed and tested AI-assisted code.Source: AI USE STATEMENT · pp. 9–10
• A generative image tool produced an initial draft of the Figure 1 anatomical schematic, which the authors later modified and refined manually.Source: AI USE STATEMENT · pp. 9–10
• The authors take responsibility for the final manuscript, figures, code, and reported results, including AI-assisted content.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

• The study uses the publicly available emg2pose and emg2qwerty datasets and collected no additional human-participant data.Source: ETHICS STATEMENT · p. 10
• The original studies report ethics approval and participant informed consent (Salter et al., 2024; Sivakumar et al., 2024).Source: ETHICS STATEMENT · p. 10
• The datasets are used in accordance with their respective licenses.Source: ETHICS STATEMENT · p. 10