A Generalizable and Discriminative Learning Method for Deep EEG-Based Motor Imagery Classification

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) are state-of-the-art for motor imagery (MI) classification.
• Current CNNs in MI primarily use cross-entropy loss, focusing only on feature separability.
• Objective functions for MI classification require further optimization.

Purpose of the Study:

• To propose a novel loss function for CNN-based MI classification.
• To enhance the discriminative capacity of deep features and prevent overconfident predictions.
• To improve the overall performance and robustness of MI recognition models.

Main Methods:

• A combined loss function integrating smoothed cross-entropy and center loss was developed.
• Smoothed cross-entropy regularizes predictions using uniform distribution noise.
• Center loss minimizes the distance between deep features and their class centers.

Main Results:

• The proposed loss function significantly improved MI classification accuracy on benchmark datasets (BCI competition IV-2a and IV-2b).
• The method outperformed existing state-of-the-art models.
• The approach demonstrated increased deep feature discriminative capacity and reduced overfitting.

Conclusions:

• The novel loss function provides a more robust optimization strategy for CNNs in MI tasks.
• This method enhances both inter-class separability and intra-class invariance of features.
• The proposed approach offers a promising direction for advancing MI recognition technology.