Through the motions of human vigilance: geometry-aware generation and forecasting of eeg spectral dynamics

This thesis unfolds across eight chapters, detailing the importance of EEG in safety-critical workplace settings, possible methods for forecasting fluctuations in human vigilance, and a novel geometrical approach to data augmentation aimed at addressing the scarcity of biological signals. The geometrical part, in particular, details how to train a Generative Adversarial Network (GAN) framework exploiting the properties of the augmented covariance matrix of EEG data. In particular, a novel regularization strategy that leverages the geodesic distance on the Symmetric Positive Definite manifold is defined and tested with success, opening the road to geometry-informed data augmentation for biological signals.