Princeton University Special Astrophysics Seminar

In astrophysics and cosmology we are experiencing a period of profound changes caused by two major developments: 1) the availability of vast quantities of data, from large surveys and from simulations, and 2) the rapid progress in machine learning, in particular probabilistic modeling. But these developments lead to a conundrum: How do we bridge the gaps between data-driven and theoretical models? How do we actually learn profound aspects of physical systems from data?

I will discuss projects on spectroscopic, image, and time-domain analysis that combine physical with data-driven approaches. I will show that we can extract more information from observations, find truly rare objects, reveal unseen features, and make unexpected connections between different observational domains. And I will argue that, to fully exploit the information content of future observations and experiments, we have to design custom machine learning architectures for the physical sciences.