Transfer Learning on Electromyography (EMG) Tasks: Approaches and Beyond.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Machine learning on electromyography (EMG) data assumes consistent data distribution, which is often unmet in real-world scenarios.
• Model recalibration requires costly data re-collection and annotation, hindering practical EMG applications.
• Transfer learning (TL) offers a solution by leveraging knowledge from source domains to improve target task performance with less data.

Purpose of the Study:

• To survey and assess the applicability of over fifty transfer learning (TL) methods for electromyography (EMG) analysis.
• To provide biological insights into existing TL methods by examining muscle physiology and EMG generation.
• To categorize existing TL research for EMG into data-based, model-based, training scheme-based, and adversarial-based approaches.

Main Methods:

• Systematic review and categorization of peer-reviewed transfer learning literature relevant to EMG.
• Analysis of biological underpinnings of EMG signals, including muscle structure and generation mechanisms.
• Classification of transfer learning approaches based on their core strategies (data, model, training, adversarial).

Main Results:

• Identification and evaluation of more than fifty representative transfer learning approaches for EMG applications.
• Categorization of existing research into four main types: data-based, model-based, training scheme-based, and adversarial-based.
• Discussion of the biological foundations influencing the effectiveness of TL in EMG analysis.

Conclusions:

• Transfer learning is a promising paradigm for reducing calibration efforts in EMG machine learning.
• Understanding the biological basis of EMG signals is crucial for developing effective TL strategies.
• Further research is needed to enhance the practicality of EMG transfer learning algorithms for real-world deployment.