Deep transfer learning compared to subject-specific models for sEMG decoders.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Accurate surface electromyography (sEMG) decoding is crucial for muscle-to-machine interfaces, but inter-subject variability poses challenges.
• Deep learning models for sEMG often require extensive training data and can overfit with limited samples.

Purpose of the Study:

• To investigate and compare methods for calibrating deep learning models to new users with limited training data.
• To evaluate the effectiveness of transfer learning against subject-specific modeling for sEMG decoding.

Main Methods:

• Investigated transfer learning using weight initialization to recalibrate two pre-trained deep learning models on new subject data.
• Compared transfer learning performance against subject-specific models using three public sEMG databases.
• Evaluated models based on accuracy, required training data, and calibration time.

Main Results:

• Transfer learning improved performance by 5% over pre-trained models and 12% over subject-specific models on average.
• Transfer learning required 22% fewer training epochs compared to other methods.
• Demonstrated that transfer learning facilitates faster learning with fewer training samples.

Conclusions:

• This study provides the first direct comparison of subject-specific modeling and transfer learning for sEMG decoding.
• Transfer learning offers a more efficient calibration strategy for deep learning models in sEMG applications.
• Results guide engineers in selecting appropriate calibration schemes for muscle-to-machine interfaces.