EMGFlow: Robust and Efficient Surface Electromyography Synthesis via Flow Matching

The Thesis

Surface electromyography (sEMG) — the technique of reading electrical signals from muscles through skin electrodes — powers prosthetic limb control, gesture interfaces, and rehabilitation devices, but deep-learning models for these applications starve for labeled training data. EMGFlow proposes using Flow Matching (FM), a generative modeling approach that learns a smooth, continuous path between random noise and real data distributions, to synthesize realistic sEMG signals and augment scarce datasets. The authors claim FM outperforms both GAN-based and diffusion-based baselines on three benchmark datasets under a 'train on synthetic, test on real' protocol — the hardest test of whether fake data can substitute for the real thing. The catch: this is a relatively small-scale academic study, and the downstream benchmark tasks are gesture classification on existing datasets rather than a live prosthetics deployment.

Catalyst

Flow Matching as a generative modeling framework only matured as a practical tool in 2022-2023, with key theoretical and implementation work by Lipman et al. enabling stable, fast training without the adversarial instability of GANs or the slow iterative sampling of diffusion models. Simultaneously, the sEMG research community has consolidated around a handful of public benchmark datasets (such as NinaPro and CSL-HDEMG), making rigorous cross-dataset evaluation feasible for the first time. Open-source code release accompanying this paper removes a common barrier to replication.

What's New

Prior generative approaches to sEMG synthesis fell into two camps: GANs (Generative Adversarial Networks), which pit a generator against a discriminator but are notoriously unstable to train and prone to mode collapse, and diffusion models, which produce high-quality samples but require hundreds of sequential denoising steps and are therefore slow at inference. EMGFlow replaces both with Flow Matching, which learns a direct continuous-time trajectory from noise to data using an ordinary differential equation solver. The authors claim this yields better synthetic data quality on their 'train synthetic, test real' benchmark while also being faster to sample from — a quality-efficiency trade-off that neither GAN nor diffusion baselines achieved simultaneously.

The Counter

The core problem here — small sEMG datasets limiting gesture recognition — is real, but it is not clear that synthetic data generation is the right fix. The dominant bottleneck in real-world myoelectric control is cross-subject and cross-session variability: a classifier trained on one person's signals degrades badly on another person's, or even on the same person after a short break. Generating more synthetic samples from the same underlying distribution does not solve that variability problem; it may just make classifiers more confident on the wrong features. The benchmarks in this paper are relatively clean academic datasets where the problem is sample count, not distribution shift — a much easier setting than real deployment. Flow Matching is also not obviously superior to the latest diffusion or score-based models for time-series; the diffusion baselines used here are not identified as state-of-the-art in the time-series synthesis literature broadly. Finally, the prosthetics and rehabilitation device market has been aware of data scarcity for over a decade and has largely adapted through transfer learning, domain adaptation, and calibration protocols rather than synthesis — meaning EMGFlow is entering a space with established alternative solutions.

The Paper