Deep Learning Facilitates Spatial Inference

While deep learning has revolutionized science as a flexible, black-box predictive framework, spatial statistical analyses predominantly rely on parametric models with interpretable parameters. In this talk, we present several applications that leverage deep learning, not as direct models for data, but as a computational tool to alleviate the heavy burden that notoriously hinders fitting spatial processes such as Gaussian and max-stable process models. These methods rely on Simulation-Based Inference (SBI), replacing intractable likelihood evaluations with model simulations and using deep neural networks to approximate the mapping from data to the parameters of interest. We demonstrate this framework across diverse statistical paradigms (frequentist, Bayesian, and possibilistic inference), approximation targets (likelihoods, posteriors, and p-values), and application domains such as extreme streamflow and disease spread.

Host: Joe Guinness