Federated Learning in Offline and Online EMG Decoding: A Privacy and Performance Perspective.

Area of Science:

• Neuroscience
• Machine Learning
• Human-Computer Interaction

Background:

• Neural interfaces enable intuitive interaction but face privacy challenges with sensitive neural data.
• Federated learning (FL) is a privacy-preserving technique, but its application in real-time neural interfaces is underexplored.

Purpose of the Study:

• To propose a framework for applying FL to neural interfaces.
• To evaluate FL-based neural decoding in both offline and real-time settings using surface electromyography.

Main Methods:

• Developed a conceptual framework for FL in neural interfaces.
• Conducted offline simulations and real-time online user studies using high-dimensional surface electromyography (sEMG).
• Assessed performance and privacy trade-offs of FL in neural decoding.

Main Results:

• Offline FL simulations indicated potential for simultaneous performance and privacy enhancement.
• Real-time online experiments showed standard FL assumptions falter with user-decoder co-adaptation.
• FL maintained privacy benefits but introduced performance trade-offs not seen in offline tests.

Conclusions:

• Current FL methods are insufficient for real-time neural decoding due to co-adaptive dynamics.
• Specialized FL algorithms are needed to address the unique challenges of online neural interfaces.
• Further research is required to bridge the gap between offline FL potential and online application realities.