A General Introduction to Hybrid Statistical Deep Learning Methods for Spatial and Spatio-Temporal Data

Music City Center 

Understanding and predicting complex processes from spatial or spatio-temporal data is crucial for addressing global challenges such as climate change, biodiversity loss, and resource management. Yet, modeling these data is challenging: They are large, noisy, and often only an indirect observation of the quantity of interest. In this introductory overview lecture we draw on classical approaches to modeling spatio-temporal data that involve two dominant approaches: descriptive and dynamical. We then discuss some of the challenges faced by these classical approaches, and how emerging hybrid methods incorporating deep learning and AI offer unprecedented opportunities in modeling and computation. In the second part of the lecture we present new neural-network based likelihood-free methods for making point estimation or fully Bayesian rapid inference with spatial and spatio-temporal models. We showcase several applications of these new methods, and conclude with an outline of current directions and future challenges in this evolving field.