Channel Reduction Approaches in Deep Learning for Brain-Computer Interfaces

Deep learning with convolutional neural networks transformed the study of computer vision by incorporating raw image data to train neural networks in what is called end-to-end learning. These techniques have since been applied to study and improve modeling techniques for interpreting and decoding brain activity in brain-computer interfaces. A popular modality of brain-computer interfaces utilizes non-invasive recordings of brain activity called electroencephalograms (EEGs). While end-to-end solutions for brain-computer interfaces analyzing EEGs have shown significant improvement in decoding brain activity over previous methods, the trained models employ tools for processing raw data in ways that lack explainability and connection to neuroscience. In our proposed solution, we explore methods of preprocessing data through various forms of dimensionality reduction that elucidate the learning mechanics of deep learning models trained to perform motor imagery classification. Through our novel approach, we reveal key insights into what aspects of data are most important for classification, discover which electrodes are pertinent to high-quality accuracy, and connect these findings back to neuroscience.