Title: Reconstructing the initial conditions of the Universe
Abstract: Over the last decade, cosmologists have been exploring a novel way of analyzing survey data by reconstructing the Gaussian initial conditions at the beginning of the Universe. This requires solving a high dimensional inverse problem with an expensive, non-linear forward model: a cosmological N-body simulation. I will talk about three advances made over the last couple of years to make this exercise tractable, and put the lessons in a broader context. 1) First, I will talk about the automatically differentiable N-body solvers that we have developed in ML frameworks. 2) Next, I will talk about how we combine these with recurrent networks to accelerate optimization by learning a path to the maximum-a-posteriori (MAP) estimate of the initial conditions. 3) Finally I will talk about how we combine Hamiltonian Monte Carlo and variational inference in a framework called variational self-boosted sampling to accelerate sampling in these high dimensions.