Speaker
Description
Advances in experimental techniques allow brain activity to be measured at scale, for example using electroencephalography (EEG) and functional magnetic resonance imaging (fMRI). Alongside experimental advances, computational platforms for simulating brain dynamics have also shown meaningful progress. Together, they have created new opportunities to understand spatiotemporal patterns of brain activity underlying diverse physiological and cognitive processes, including sleep and working memory. However, optimally integrating computational models with experimental data remains an open challenge. Data assimilation is challenging because the brain is high-dimensional system governed by strongly nonlinear dynamics. To address these challenges, we propose a score-based Bayesian data assimilation method for high-dimensional Jansen–Rit models using EEG measurements. Based on the measurement equation of the EEG forward model, we construct a forward process from latent states to measurements. This formulation enables closed-form computation of the likelihood score for posterior sampling of latent brain states, and learning the prior score with a score model allows the method to scale to high-dimensional settings. Numerical experiments on EEG-based state estimation in the Jansen–Rit model support the effectiveness of the proposed method.
