Stat Cafe - Dr. Yeseul Jeon
Uncertainty-Aware Neural Multivariate Geostatistics
- Time: Wednesday, 3/25/2026 from 11:15AM to 12:30PM
- Location: BLOC 457
Description
We propose Deep Neural Coregionalization (DNC), a scalable framework for uncertainty-aware multivariate geostatistics. DNC models multivariate spatial effects using spatially varying latent factors and loadings, assigning deep Gaussian process priors to both components. This structure learns shared latent spatial patterns together with response-specific, location-dependent mixing weights, enabling flexible nonlinear and spatially varying cross-variable associations. We develop a variational formulation linking deep Gaussian processes to deep neural networks, showing that maximizing the ELBO corresponds to training DNNs with weight decay and Monte Carlo dropout. The approach supports efficient mini-batch optimization without MCMC, producing calibrated credible surfaces for prediction and spatial effect estimation.
Our Speaker
Dr. Yeseul Jeon is a postdoctoral research associate working with Dr. Rajarshi Guhaniyogi and Dr. Aaron Sheffler at UCSF. She received her Ph.D. in Statistics from Yonsei University in South Korea. Her research focuses on developing statistical and deep learning methods for high-dimensional, complex, and structured data with dependence, with an emphasis on interpretability and uncertainty quantification.