Speaker
Description
This is a story about how breakthroughs in machine learning over the past 15 years have borne a paradigm shift in how to do data analysis in Astronomy, enabling impactful science that was previously out-of-reach.
The dynamics of galaxy evolution leaves its mark on the distribution of galaxy properties in the Universe. Studying the joint distribution of galaxy characteristics (mass, metallciity, dust, star-formaiton history, etc) - how they relate to eachother and how that evolves over cosmic history - therefore remains our main observational tool for constraining galaxy evolution. However, robustly measuring the joint distribution of galaxy properties in the Universe has stood as a formidable Astronomical data analysis problem for decades, for four key reasons: (1) the stellar population synthesis (SPS) models relating galaxy charateristics to their observable SEDs are too expensive to run at the scale required to analyse large samples, (2) strong degeneracies (eg dust-age-metallicity-redshift) make sampling even individual galaxy posteriors from photometric data challenging and expensive, (3) survey selection effects complicate population-level inference, making robust hierarchical inference computationally intractable, and (4) comprehensive parametric models describing the whole galaxy population have been out-or-reach. In this talk I'll show how advances in machine-learning allow us to smash through all of these barriers and derive comprehensive, robust constraints on the galaxy population from large, deep galaxy surveys.
