Neural likelihood for irregular spatial data via graph neural networks (GNNs)
Directly estimating a statistical model’s likelihood function can be computationally intensive or even intractable for large spatial datasets such as those commonly found in the environmental sciences. As a part of my Ph.D thesis, I am applying graph neural networks (GNNs) to learn the likelihood for our locally stationary Gaussian process model for ocean heat content. Please see my recent poster for more details.
(Figure - Example likelihood surfaces for an isotropic Gaussian process. The GNN used to produce the neural surface was trained on actual Argo float sampling patterns.)