“GraphGP: Scalable Gaussian Processes With Vecchia’s Approximation and Application to the SPHEREx Survey”

Abstract:

Gaussian processes are a powerful tool for modeling continuous fields in astronomy, but their naive computational cost and memory requirements often limit their practical use. Vecchia’s approximation is a sparse precision matrix approximation for stationary, decaying kernels that conditions each point only on its nearest neighbors. I will present GraphGP, a new variant of Vecchia’s approximation that scales to a billion parameters on a single GPU with linear time and memory requirements. I will then discuss an ongoing application of GraphGP to all-sky modeling of diffuse emission with SPHEREx data, illustrating the utility of Gaussian processes in astronomy and the importance of scaling them to large domains.