Source-linked AI summary
MyoFlow: Anchor-Tied Rectified Flow for HD-sEMG Gesture Recognition Across Sessions and Subjects
Chenhao Wu, Dingjie Peng, Satoshi Funabashi, Satoshi Konishi, Wuqiang Yang, Hiroshi Onoda, Hironori Washizaki, Jiang Liu
TL;DR
HD-sEMG gesture recognition suffers from session- and subject-related distribution shifts, while prior generative approaches typically leave prediction to a separate classifier. MyoFlow uses domain-conditioned rectified flow toward shared gesture anchors that also define nearest-anchor decisions, achieving reported gains across Hyser and CEMHSEY evaluations.
Problem
HD-sEMG gesture recognition must handle distribution shifts across sessions and subjects, while prior generative models do not directly make the class decision.
Method
MyoFlow uses subject- and session-conditioned rectified-flow transport from encoded windows to shared gesture anchors serving as transport targets and class prototypes.
Results
MyoFlow improves mean Hyser cross-session and cross-subject accuracy over DiffHGR by 4.24 and 6.37 percentage points, respectively, and reaches 91.71% mean zero-shot accuracy on CEMHSEY.
Takeaways & Limitations
Anchor-tied transport couples representation dynamics to the decision rule and supports zero-shot transfer plus few-shot calibration across session and subject shifts.
Takeaways & Limitations
Training uses only labeled source data; no target-session window, statistic, or label enters training.
Abstract
from arXiv · showhide
High-density surface electromyography (HD-sEMG) gesture recognition supports prosthetic control, assistive robotics, and rehabilitation, but electrode re-donning and physiological variability cause distribution shifts that degrade accuracy across sessions and subjects. Generative HD-sEMG models primarily synthesize signals for augmentation; although diffusion models enhance representation learning, prediction still relies on a separate classifier. To tie learned dynamics to the decision rule, we propose MyoFlow, the first discriminative flow-matching framework for HD-sEMG recognition across sessions and subjects. It recasts classification as anchor-tied transport: a domain-conditioned rectified flow moves encoded windows toward gesture anchors that serve as transport targets and define the nearest-anchor decision geometry, enabling zero-shot prediction without an independent head. On the Hyser dataset, MyoFlow improves mean cross-session and cross-subject accuracy over the strongest diffusion-based baseline by 4.24\% and 6.37\%, respectively, and achieves 91.71\% mean zero-shot accuracy and 97.39\% mean few-shot accuracy across multiple days on the CEMHSEY dataset.
1 Waseda University, Japan 2 KDDI Research Inc., Japan 3 The University of Manchester, UK
The paper concerns HD-sEMG, gesture recognition, representation learning, flow matching, and domain generalization.
- HD-sEMG is the signal domain addressed by the paper.
- Gesture recognition is the paper’s application task.
- Representation learning, flow matching, and domain generalization are the paper’s identified technical themes.
1. INTRODUCTION
MyoFlow addresses distribution shifts in HD-sEMG recognition across sessions and subjects by coupling rectified transport with anchor-based classification. It reports improvements across datasets and zero- and few-shot settings.
- 1. INTRODUCTION: HD-sEMG recognition is challenged by electrode re-donning, skin-impedance changes, movement variability, and anatomical or neuromuscular differences.
- 1. INTRODUCTION: Earlier approaches often require labeled target recordings for recalibration, while generative and diffusion-based methods do not directly use learned dynamics as the classification rule.
- 1. INTRODUCTION: MyoFlow formulates cross-session and cross-subject recognition as rectified transport in representation space.
- 1. INTRODUCTION: Gesture anchors serve as both transport targets and class prototypes, coupling representation learning to nearest-anchor decisions.
- 1. INTRODUCTION: MyoFlow improves recognition across sessions and subjects in zero- and few-shot settings on multiple datasets.
2. METHOD
MyoFlow formulates recognition as domain-conditioned rectified transport from encoded HD-sEMG windows to shared gesture anchors, coupling representation learning with nearest-anchor decisions. Source-only training and support-based calibration adapt the encoder while preserving the flow and anchor geometry.
- 2.2. Conditional Rectified Flow: MyoFlow maps each encoded window toward a label-specific learnable gesture anchor, making the terminal endpoint the basis for recognition.The latent endpoint dimension is D = 256, and the target anchor is selected from the window label during training.
- 2.2. Conditional Rectified Flow: The velocity field is conditioned on subject and session context but does not receive the label, allowing inference when the target label is unknown.For unseen sessions, the source-session embedding is reused; for unseen subjects, the mean training-subject embedding is used without target-domain data.
- 2.3. Anchor-Tied Classification: Nearest-anchor classification ties the flow destinations to the decision rule, making the learned class geometry explicit rather than an artifact of a separate read-out.The anchor bank supplies both transport targets and classifier prototypes, while only global isometries preserve the objective.
- 2.4. Source-Only Learning: Training uses source data only, combining flow matching, anchor supervision, and auxiliary regularization before optional cross-session adaptation on the evaluated subject’s source session.Leave-one-subject-out cross-subject evaluation excludes the held-out subject from pretraining and contributes no gradient.
- 2.5. Few-Shot Calibration: For K > 0, calibration recomputes BatchNorm statistics and initializes a support-mean linear head on e0, while freezing the flow and anchor bank.Optimization uses augmented support windows, and query repetitions are never seen.
3. EXPERIMENT SETUP
MyoFlow is evaluated on Hyser and CEMHSEY under cross-session and cross-subject shifts using specified source-target splits and calibration settings.
- Evaluation covers Hyser cross-session and cross-subject recognition plus CEMHSEY cross-session recognition across multiple day gaps.Hyser uses Session 1 as source and Session 2 as target, while CEMHSEY uses Day 1 as source and Days 2, 4, 6, 8, and 11 as targets.
- Hyser contains 20 subjects, two sessions, 11 gestures, and 256 channels, whereas CEMHSEY contains 6 subjects, 11 days, 11 gestures, and 320 channels.Both datasets are sampled at 2048 Hz and segmented into nonoverlapping 50 ms windows.
- Cross-subject evaluation uses 20-fold leave-one-subject-out validation on Hyser Session 1 without adaptation to the held-out subject.
4. RESULTS AND ANALYSIS
MyoFlow outperforms listed baselines across Hyser settings, while temporal-window, ROC, and transport analyses characterize its deployment trade-offs, discrimination, and decision-aligned geometry.
- Overall comparison: 4.24 and 6.37 percentage points: MyoFlow improves cross-session and cross-subject performance over DiffHGR, respectively, and exceeds the strongest unseen-subject baselines.For unseen subjects, it surpasses DiffHGR and ViT-MDHGR by 8.54 and 9.37 points, respectively.
- Temporal sensitivity: 50 to 100 ms: increasing the temporal window improves accuracy across settings, while 150 ms provides no significant further gain.The 50 ms setting halves acquisition time relative to 100 ms while retaining strong recognition accuracy.
- Classifier discrimination: 0.963 and 0.992: macro-AUCs on Hyser and CEMHSEY, respectively, with every class-wise AUC exceeding 0.950.The ROC analysis uses per-window, zero-shot, anchor-derived probabilities and indicates reliable class-wise discrimination under cross-session and cross-day shifts.
- Interpretability of learned transport: Endpoint trajectories move toward and concentrate around class anchors, organizing latent representations into a decision-aligned geometry.Because the same anchors determine prediction, each trajectory connects representation dynamics to the final decision at the sample level.
5. CONCLUSION
MyoFlow introduces anchor-tied discriminative flow matching for cross-session and cross-subject HD-sEMG gesture recognition. Its transport targets also define the decision rule, supporting zero-shot transfer and few-shot calibration.
- 5. CONCLUSION: MyoFlow is a discriminative flow-matching framework whose domain-conditioned rectified flow moves encoded windows toward gesture anchors used for nearest-anchor prediction.The anchors serve as both transport endpoints and prototypes, eliminating the need for an independent classification head.