Astrophysics abounds with precision measurement problems, for example when we search for exoplanets, map the dark matter in our Galaxy, or test the cosmological model. Because astronomy is observational, not experimental, no measurement is clean; we always have to model many auxiliary components (instrument calibration, foregrounds, and backgrounds) in addition to the component we most care about. I'll argue that getting data with good or rich causal structure is key to component separation. I will highlight the roles of classical statistics (especially the likelihood principle) and machine learning in making good measurements. When deployed correctly (but counterintuitively), machine-learning methods can be used to make precision measurements and discoveries more conservative and more believable. I will give examples from recent work in exoplanets, dark matter, and cosmology.
Speaker: David Hogg, New York University, Center for Computational Astrophysics
Host: Adam Miller