Representation Learning and Foundation Models for Electroencephalography Analyses: Current trends, fundamental insights, and future directions?

Since the first ground-breaking human electroencephalography (EEG) recordings in 1924 [1], the past century has witnessed a tremendous growth of EEG applications in cognitive neuroscience, clinical and engineering applications due to EEG's low operational cost and mobility [2]. On the one hand, advances in high-density non-invasive scalp EEG or invasive intracranial EEG (iEEG) have offered both excellent temporal resolution and increasingly improved spatial resolution to study brain functions and their link to emotions, memory, learning, and diseases. EEG-based brain-computer interface (BCI) can offer new dimensions for entertainment, virtual reality, neurofeedback and closed-loop therapy. On the other hand, recent advances in artificial intelligence (AI) and machine learning have opened new opportunities for analyses of EEG and other neural data [3]. This article aims at presenting an overview of the cutting-edge machine learning techniques for EEG analyses. By leveraging large-scale EEG data with state-of-the-art representation learning and transfer learning paradigms, we are empowered to discover latent EEG features that are proved useful for clinical care and BCI. We discuss some general principles of representation learning and show walk-through practical examples of EEG analysis. The article also aims at highlighting the effort of applying AI models to discover neuroscience insights and linking them to the fundamentals of EEG signal analyses from a signal-processing perspective. While our focus is on EEG and iEEG signals, most of the approaches discussed here are generally applicable to other brain signal modalities.