High-dimensional inference from the Cosmic Microwave Background on GPU

Marius Millea (University of California, Berkeley & Davis)

The problem of inferring cosmological parameters (like the Hubble constant, or the dark matter density) from observations of the Cosmic Microwave Background (CMB) is a difficult high-dimensional inference problem which is central to all future CMB observations. In this talk, I’ll describe how Julia made our solution, CMBLensing.jl, possible, and how this gives 10X better constraints that other existing methods. The code relies on 1) a sophisticated “array and metadata” custom AbstractArray, 2) a flurry of custom chain rules to allow reverse-mode automatic differentiation, and 3) works on GPU nearly for-free. I’ll describe lessons from each of these points which others can take away for their own code, as well as showing how we hook this inference problem into a new inference algorithm, MuseInference.jl, which is applicable to a wide range of problems in Astronomy & Astrophysics.