Combining generative adversarial networks and multi-output CNN for motor imagery classification

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Motor imagery (MI) classification is crucial for brain-computer interfaces (BCIs).
• MI data are dynamic and challenging to acquire, impacting classification model performance.
• Traditional Convolutional Neural Networks (CNNs) may discard feature information during end-to-end processing.

Purpose of the Study:

• To propose a novel algorithm for enhanced motor imagery classification.
• To address the challenge of limited and dynamic MI data.
• To improve the performance of BCI systems through advanced deep learning techniques.

Main Methods:

• A novel algorithm combining Long Short-Term Memory Generative Adversarial Networks (LGANs) and Multi-Output Convolutional Neural Networks (MoCNN) was developed.
• A data augmentation method using LGANs was employed to generate additional training data.
• An attention network was incorporated to further enhance model performance.

Main Results:

• The proposed LGAN-based generative model produced more realistic data, validated on BCI competition IV datasets 2a and 2b.
• The expanded training dataset significantly improved the classification model's performance.
• The combined LGAN-MoCNN approach demonstrated superior MI classification accuracy.

Conclusions:

• The proposed method effectively improves motor imagery classification.
• This advancement facilitates better control and interaction in brain-computer interface applications.
• The integration of generative models and attention mechanisms offers a promising direction for BCI research.