Deep Learning for Brain-Signal Decoding from Electroencephalography

Abstract

Brain-signal decoding using machine learning can process larger amounts of signals and extract different information than humans can, with potential uses in medical diagnosis or brain-computer interfaces. In particular, brain-signal decoding from electroencephalographic (EEG) recordings is a promising area for machine learning due to the relative ease of acquiring large amounts of EEG recordings and the difficulty of interpreting them manually. Deep neural networks are a natural choice to train to decode EEG signals as they have been successful at a variety of natural-signal decoding tasks like object recognition from images or speech recognition from audio. However, prior to the work in this thesis, it was still unclear how well deep neural networks perform on EEG decoding compared to hand-engineered, feature-based approaches, and more research was needed to determine the optimal approaches for using deep learning to decode EEG. This thesis describes constructing and training deep neural networks for EEG decoding that perform as well as feature-based approaches and developing visualizations that suggest they extract physiologically meaningful features.

Contents

Chapter	Summary
Introduction and Background
Introduction	Deep learning on EEG is a very promising approach for brain-signal-based medical applications
Prior Work	Prior to 2017, research did not clearly show how competitive deep learning is compared with well-optimized feature baselines
Filterbank Common Spatial Patterns and Filterbank Network	Filter Bank Common Spatial Patterns was an inspiration for initial network architectures
Methods
Neural Network Architectures for EEG-Decoding	Progressively more generic neural network architectures for EEG decoding were created
Cropped Training	A training strategy to use many sliding windows was implemented in a computationally efficient manner
Perturbation Visualization	A visualization of how frequency features affect the trained network was developed
Invertible Networks	Invertible networks as more interpretable EEG classifiers
Applications and Results
Movement-Related Decoding	Deep learning can be at least as good as feature-based baselines for movement-related decoding; deep networks also learn to extract known spectral features
Generalization to Other Tasks	Our deep networks generalize well to decoding other decoding tasks
Decoding Pathology	Deep networks designed for task-related decoding can also decode pathology well
Understanding Pathology Decoding with Invertible Networks	Better understanding of pathology decoding with invertible networks
Discussion	Cross-dataset decoding, long-time-scale architectures and in-context learning may improve performance and interpretability