Convolutional neural networks for decoding of covert attention focus and saliency maps for EEG feature visualization

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Convolutional neural networks (CNNs) are powerful function approximators used in electroencephalography (EEG) signal decoding for brain-computer interfaces (BCI).
• However, the "black box" nature of artificial neural networks hinders the interpretation of their internal decision-making processes.
• Understanding these processes is crucial for advancing BCI technology and cognitive neuroscience research.

Purpose of the Study:

• To systematically evaluate CNNs for EEG signal decoding and explore methods for visualizing their decision-making processes.
• To compare CNN performance against traditional methods like linear discriminant analysis (LDA) across varying data dimensionalities.
• To investigate the impact of specific CNN components on decoding accuracy and identify optimal model configurations.

Main Methods:

• Developed a CNN model for decoding covert attention from EEG event-related potentials during object selection.
• Compared CNN performance with LDA on datasets of differing dimensionality, assessing transfer learning capabilities.
• Systematically altered model components to validate their impact and employed saliency maps for feature visualization.

Main Results:

• CNNs achieved comparable accuracy to LDA on low-dimensional data but significantly outperformed LDA on high-dimensional data (90.7% accuracy vs. 8.3% chance level).
• Specific components like parallel convolutions, tanh/ELU activation, and dropout regularization enhanced CNN performance.
• Saliency maps effectively visualized spatial and temporal features, correlating with the expected P300 component.

Conclusions:

• CNNs offer a powerful and interpretable approach for EEG decoding, particularly with high-dimensional data.
• Saliency maps provide valuable insights into the neural correlates of cognitive tasks by highlighting relevant EEG features.
• Recommendations are provided for optimal CNN implementation in EEG-based BCI and cognitive research.