Gated Recurrent Neural Networks for EMG-Based Hand Gesture Classification. A Comparative Study.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Electromyographic (EMG) signals from upper limb muscles correlate with hand gestures.
• This correlation offers potential for advanced prosthetic and cybernetic control systems.

Purpose of the Study:

• To compare various recurrent neural network (RNN) architectures for classifying EMG-based hand gestures.
• To evaluate the impact of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) cells, attention mechanisms, and learning rates on classification accuracy.

Main Methods:

• Comparative analysis of different RNN configurations, including LSTM and GRU units.
• Implementation and evaluation of an attention mechanism within the RNN architecture.
• Testing the influence of varying learning rates, including a step-wise approach.

Main Results:

• A classifier utilizing a bidirectional recurrent layer with LSTM units demonstrated superior performance.
• The addition of an attention mechanism significantly improved classification accuracy.
• Training with a step-wise learning rate yielded better results compared to other tested rates.

Conclusions:

• The optimal RNN configuration for EMG-based hand gesture classification involves bidirectional LSTM layers with an attention mechanism.
• This optimized model, trained with a step-wise learning rate, provides a robust solution for prosthetic and cybernetic applications.