Automatic discovery of resource-restricted Convolutional Neural Network topologies for myoelectric pattern...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) excel at information extraction but lack systematic topology selection methods.
• Current CNN topology design relies heavily on domain expertise and empirical refinement.
• This limitation hinders advancements in muscle-computer interfaces, which require effective CNNs.

Purpose of the Study:

• To develop a systematic approach for selecting Convolutional Neural Network (CNN) topologies.
• To address the lack of heuristics for CNN topology selection in muscle-computer interfaces.
• To create computationally efficient CNN topologies suitable for embedded systems.

Main Methods:

• A novel evolutionary algorithm was employed to search for optimal CNN topologies.
• The search space was constrained to exclude topologies with high inference-time computational complexity.
• The algorithm requires users to specify only a few intuitive hyperparameters.

Main Results:

• The approach generated computationally lightweight CNN topologies for myoelectric pattern recognition.
• Performance was evaluated using surface electromyography (sEMG) signals for movement decoding.
• The generated topologies achieved classification accuracies comparable to existing methods.

Conclusions:

• The proposed evolutionary algorithm offers a systematic and efficient method for CNN topology selection.
• This approach facilitates the development of practical muscle-computer interfaces by creating implementable CNNs.
• The method reduces reliance on domain expertise and iterative empirical evaluation.