Deep Generative Replay-based Class-incremental Continual Learning in sEMG-based Pattern Recognition.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Advancements in neural networks and sensing technologies drive interest in surface electromyogram (sEMG)-based pattern recognition.
• Incremental updating of pre-trained networks enhances user-centered interfaces but risks catastrophic forgetting.
• Traditional methods to mitigate forgetting, like replaying historical data, require substantial memory, limiting practical application.

Purpose of the Study:

• To propose a novel deep generative replay-based continual learning (CL) framework.
• To enable incremental addition of new classes to pre-trained networks without significant memory overhead.
• To address the challenge of catastrophic forgetting in sEMG pattern recognition.

Main Methods:

• Development of a deep generative replay-based continual learning framework.
• Evaluation using a public sEMG dataset.
• Testing under a two-class incremental learning scenario across four tasks.

Main Results:

• The proposed framework demonstrated superior performance compared to conventional CL methods.
• Performance was competitive with experience replay, which directly reuses historical data.
• The method effectively mitigates catastrophic forgetting in incremental learning scenarios.

Conclusions:

• The deep generative replay framework offers an effective solution for continual learning in sEMG pattern recognition.
• This approach overcomes the memory limitations associated with traditional experience replay.
• The framework facilitates flexible and efficient updates of pre-trained models for user-centered applications.