A Deep Generative Model for Resting-State EEG Synthesis and Transferable Representation Learning

The article presents REST-GAN, a generative adversarial network framework designed for synthesizing resting-state EEG signals and learning transferable representations from raw data. By employing adversarial training alongside a self-supervised reconstruction objective, REST-GAN generates time series that accurately reflect the temporal, spectral, and connectivity properties of real EEG, achieving high precision and recall in band-power feature space and demonstrating superior performance in demographic classification tasks compared to models trained on raw EEG. This approach offers a computationally efficient alternative for EEG analysis, reducing the need for manual feature engineering and extensive training data.