Advancing hvEEGNet as a general-purpose deep learning model for noisy time-series

Deep learning models applied to biological signals, particularly EEG, face significant challenges from noisy time-series data, which can impact their performance and generalizability. This thesis focuses on hvEEGNet, a hierarchical variational autoencoder (VAE) specifically designed for EEG signal re- construction, evaluating its ability to address these challenges. Through experimental evaluation on synthetic EEG datasets with varying noise conditions, this study examines hvEEGNet’s performance. While hvEEGNet demonstrates a strong theoretical foundation and innovative features, such as Soft Dynamic Time Warping (Soft-DTW) reconstruction loss, experimental results reveal limitations in handling high noise levels and complex noise structures. This thesis concludes with recommendations for enhancing hvEEGNet’s noise resilience and generalizability, offering valuable insights for future developments in robust deep learning solutions for noisy time-series data.