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
cs.LG
Quick verdict
Claims first discriminative flow-matching for HD-sEMG with anchors as transport targets; cross-session/cross-subject gains strong, but ablations, hyperparameters absent.
Summary
The paper proposes MyoFlow, which the authors call the first discriminative flow-matching framework for HD-sEMG gesture recognition across sessions and subjects. An encoder maps each 50 ms window to a 256-D endpoint; a velocity field conditioned on subject and session embeddings transports it toward learnable gesture anchors that serve simultaneously as flow targets and as prototypes, so prediction is nearest-anchor matching with no separate classifier head. Training is source-only, combining an LFM loss with anchor cross-entropy plus a pull term and three regularizers; few-shot calibration fits a nearest-prototype head on untransported endpoints from support repetitions. On Hyser, MyoFlow averages 87.36% cross-session and 80.26% cross-subject accuracy, beating DiffHGR by 4.24 and 6.37 points, with few-shot gains to 93.08% and 90.11% at two repetitions. On CEMHSEY it leads every target day zero-shot (91.71% vs 87.52%), but with calibration the margin nearly vanishes (97.83% vs 97.36% at 2 reps). No ablations justify the three regularizers or calibration design; hyperparameters are unspecified and statistical testing is limited to window-length comparisons.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
3 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- MyoFlow is the first discriminative flow-matching framework for HD-sEMG gesture recognition across sessions and subjects, enabling zero-shot prediction without an independent classifier head.
- Gesture anchors simultaneously serve as rectified-flow transport targets and nearest-neighbor class prototypes, making the learned class geometry explicit and identifiable.
- MyoFlow improves mean cross-session and cross-subject accuracy over DiffHGR by 4.24 and 6.37 percentage points on Hyser.
- On CEMHSEY, MyoFlow achieves 91.71% mean zero-shot accuracy across five target days, and the shared anchor geometry supports zero-shot transfer without target labels.
Strengths
Decision rule and transport share one geometry
Sec. 2.2–2.3 make anchors serve as both flow endpoints and prototypes; the squared-distance logits are expanded to show a tied linear read-out whose weights are the anchors themselves, and the identifiability argument (only global isometries preserve the objective, anchor pairwise distances invariant) is a clean touch beyond typical prototype baselines.
Consistent, sizable wins in the hardest regimes
Tables 1–2: cross-subject on Hyser 64.12% zero-shot and 80.26% average vs DiffHGR's 73.89% (6.37 points), with 86.55%/90.11% at 1/2 reps vs baselines stalling near 80%; the LOSO protocol (Sec. 2.4) explicitly excludes the held-out subject from pre-training, removing a common leakage route.
Both datasets are public benchmarks
Weaknesses
No ablation studies anywhere
Reproducibility information is minimal
The few-shot advantage nearly vanishes on CEMHSEY
Narrow evaluation scope
Conditioning description is internally inconsistent
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MyoFlow (arXiv 2609.17194) reformulates HD-sEMG gesture recognition as classification-by-transport: an encoder maps a 50 ms window to a 256-D endpoint, and a domain-conditioned rectified flow moves it toward one of C learnable gesture anchors via few-step Euler integration. The anchors double as nearest-neighbor prototypes, so logits are squared-distance readouts (Sec. 2.3) — equivalent to a tied linear head whose weights are the flow's own destinations, which the authors use to argue the class geometry is identifiable. Claims: (1) first discriminative flow matching for this task, (2) anchor-tied geometry enables zero-shot transfer without an independent head, (3) +4.24/+6.37 points over DiffHGR on Hyser cross-session/cross-subject, (4) 91.71% mean zero-shot accuracy on CEMHSEY.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 3 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MyoFlow and how does it classify gestures without a separate classifier?
MyoFlow treats HD-sEMG gesture classification as transport in a learned latent space: an encoder embeds each 50 ms signal window as a 256-dimensional point, and a small velocity field — conditioned on subject and session context — moves that point toward one of several learnable 'gesture anchors' via a few Euler steps of a rectified flow. Because the anchors serve as both flow destinations and class prototypes, prediction is simply nearest-anchor matching; there is no separate classifier head. A nearest-prototype head fitted on support repetitions is added only in the few-shot setting.
How accurate is MyoFlow compared to prior methods?
On the public Hyser dataset it averages 87.36% cross-session and 80.26% cross-subject accuracy, exceeding the strongest prior baseline, DiffHGR, by 4.24 and 6.37 percentage points; with one or two labeled target repetitions it reaches 89.80% and 93.08% (cross-session). On the multi-day CEMHSEY dataset it averages 91.71% zero-shot accuracy across target days 2–11 versus DiffHGR's 87.52%, though with two-repetition calibration the gap narrows to about half a point (97.83% vs 97.36%).
Is MyoFlow really the first discriminative flow-matching method for EMG?
The authors claim it is the first discriminative (classification-oriented) flow-matching framework for HD-sEMG recognition across sessions and subjects. Prior EMG flow work they cite, EMGFlow, uses flow matching only for signal generation with the label as an input, and the diffusion-based DiffHGR still requires a separate classifier. The claim is plausible for this specific task area but is not examined against discriminative or prototype-based flow methods from other domains.
Is the code or data available?
The paper evaluates on two public datasets — Hyser PR Dynamic (20 subjects, two sessions, 256 channels) and CEMHSEY (6 subjects, 11 consecutive days, 320 channels), both sampled at 2048 Hz — but the extracted text does not mention a code release, and full hyperparameters, loss weights, and the definitions of three auxiliary regularizers are not provided, so independent reproduction currently requires re-implementing the entire pipeline.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure content (window-sensitivity significance markers, ROC curves, Fig. 4 transport visualization) is only available as flattened text; plot-level claims could not be independently verified.
- The paper appears to be in short conference format; the three regularizers (geo/rad/con) may be defined in an appendix or supplementary not included in the extraction.
- Baseline accuracies appear only as cited values (presumably from prior papers); whether baselines were re-run under identical protocols cannot be determined from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.