Machine Learning at the Flatiron Institute Seminar: Alan Heavens

Title: Hierarchical Bayesian Models and Simulation-based Inference in Cosmology

Abstract: The context for this talk is cosmology, but the methods described apply to a wide range of situations. I’ll cover our recent work in principled Bayesian methods for handling large datasets for which we have a good underlying data model (which can be complex). Cosmic shear is an example where the likelihood can be applied at the field level, without computing lossy summary statistics. In the process, a very large (~10 million) dimensional latent parameter space can be explored with Hamiltonian Monte Carlo techniques. As an alternative, we also explore likelihood-free (or simulation-based) inference for cases where we have a good data simulator. Using extreme (and in some sense lossless) data compression methods, similar precision to HBMs can be achieved in at least some use cases. I will also touch on new related work in cosmology involving graph neural networks.