Self-supervised learning via VICReg enables training of EMG pattern recognition using continuous data with unclear...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Surface electromyography pattern recognition (sEMG-PR) models are crucial for human-computer interfaces.
• Training sEMG-PR models with dynamic data, including transitions between muscle activation patterns, is challenging due to the lack of ground-truth labels during these transitions.
• Self-supervised learning offers a potential solution to overcome labeling limitations in dynamic sEMG data.

Purpose of the Study:

• To investigate the efficacy of self-supervised learning using pre-trained Long Short-Term Memory (LSTM) networks for sEMG-PR model training.
• To compare the performance of LSTM models (fully-supervised vs. self-supervised) against a conventional Linear Discriminant Analysis (LDA) model.
• To evaluate model performance on both segmented ramp data and continuous dynamic data with class transitions.

Main Methods:

• Employed pre-trained Long Short-Term Memory (LSTM) networks utilizing self-supervised learning (VICReg) for sEMG-PR.
• Compared LSTM models (trained with self-supervised and fully-supervised loss) against a Linear Discriminant Analysis (LDA) model.
• Utilized two datasets: segmented ramp data (no transitions) and continuous dynamic data (with transitions).

Main Results:

• Temporal models (LSTMs) outperformed non-temporal models (LDA) when trained on continuous dynamic data.
• The VICReg pre-trained LSTM model using continuous dynamic data demonstrated superior performance compared to all other tested models.
• LSTMs trained solely on ramp data performed worse than LDA, indicating potential overfitting without sufficient dynamic information.

Conclusions:

• Representative dynamics in training data are critical for effective sEMG-PR model development.
• Self-supervised learning approaches, particularly with pre-trained LSTMs on dynamic data, can significantly enhance sEMG-PR model performance.
• The choice of model architecture and data type significantly influences sEMG-PR outcomes.