Bayesian Learning Spatiotemporal Relationships

Upcoming Seminar:  Bayesian Learning Spatiotemporal Relationships

Wednesday, October 6, 3:30pm – 4:30pm

Guest Speaker: Shiwei Lan, Arizona State University

Abstract: Tackling the emerging challenges imposed by data-intensive studies requires more well-designed statistical models and more efficient computational techniques to implement them. Motivated the functional brain connectivity study in neuroscience, we propose a novel Bayesian framework that integrates flexible spatiotemporal models and efficient uncertainty quantification (UQ) methods. To capture complex spatiotemporal correlations, we propose flexible models based on Gaussian processes with carefully designed kernels that overcome the limits of classical methods. To develop efficient UQ, we leverage deep learning techniques and scale up the inference by adopting a computational scheme called calibration-emulation-sampling (CES). The proposed methods not only show power in statistical learning of spatiotemporal structures, but could also lend their strength to data assimilation in biological and engineering applications to incorporate biochemical principles and physical laws (e.g., in mechanistic models). Our proposed methodology comes with theoretic guarantee of learning spatiotemporal correlations with sufficient data.