Cross-User Adaptation for sEMG Gesture Recognition via Continual Learning

TL;DR

This paper evaluates continual learning strategies for cross-user adaptation in sEMG-based gesture recognition, finding that experience replay outperforms EWC and naive fine-tuning in mitigating catastrophic forgetting under class- and domain-incremental drift. It demonstrates that training on shuffled, independently-windowed sEMG segments is critical for stable learning, and that hand-crafted features boost performance only when preserved via replay.

Why It Matters

Robust cross-user adaptation is essential for real-world deployment of sEMG prosthetics, where users and gestures vary. The findings provide actionable insights for building stable, adaptive neural models without retraining from scratch, improving usability and accessibility of assistive technologies.

Key Findings

• Experience replay significantly outperforms EWC and naive fine-tuning in preserving accuracy under class-incremental and domain-incremental drift.
• Training on shuffled, independently-windowed sEMG segments prevents network collapse into predicting a single dominant gesture, enabling stable continual learning.
• Hand-crafted time-domain features improve per-task accuracy by up to 40%, but only experience replay maintains this gain across tasks.