Low-Rank Adaptation for Transformers in Dynamic Hand Gesture Recognition Using High-Density Surface Electromyography.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Dynamic hand gesture recognition (HGR) is crucial for advanced prosthetics and robotics.
• High-density surface electromyography (HD-sEMG) offers detailed muscle activity data for HGR.
• Adapting HGR models to individual users efficiently remains a significant challenge.

Purpose of the Study:

• To investigate the efficacy of Low-Rank Adaptation (LoRA) for transformer-based dynamic HGR.
• To evaluate LoRA's performance in adapting models to new subjects using HD-sEMG data.
• To assess the impact of varying LoRA rank values and signal window sizes on HGR accuracy and efficiency.

Main Methods:

• Implemented LoRA within a pre-trained generalized transformer model.
• Tested LoRA with different rank values (r = 32, 64, 96) and HD-sEMG window sizes (100 ms, 200 ms).
• Focused on adapting the model to new subjects with minimal retraining.

Main Results:

• LoRA achieved high accuracy in recognizing dynamic hand gestures.
• The method demonstrated improved computational efficiency by reducing the number of trainable parameters.
• Effective adaptation to new subjects was observed with LoRA.

Conclusions:

• LoRA is a promising technique for enhancing transformer-based dynamic HGR systems.
• The approach offers scalability and robustness for real-world applications.
• Rapid adaptation to new users with minimal retraining makes LoRA valuable for practical HGR solutions.