Abstract

This talk sketches a plan for constructing highly-scalable approximate inference schemes in large spatiotemporal models, such as weather or molecular dynamics simulations. Specifically, we'll show how existing heuristics for scaling physical models such as coarse grids or mutli-scale temporal models could be learned automatically by adding auxiliary variables to an approximate variational posterior. We'll also demonstrate a new contribution to parallelizing adaptive SPDE solvers, allowing stateless sampling of entire Brownian sheets of any dimension.

Notes

• References:

  • Scalable Gradients for Stochastic Differential Equations

  • A Complete Recipe for Stochastic Gradient MCMC

• Personal website can be found here.