Abstract

Contributed Talk - Splinter EScience

Thursday, 14 September 2023, 15:10   (H 3005)

Predictive uncertainty and probability integral transform (PIT) histogram in astronomy

Ondřej Podsztavek
Czech Technical University in Prague

Automated methods are used widely in astronomy; however, they might produce incorrect values; therefore, we have to associate the values with uncertainties. Probabilistic machine learning methods are automated methods which associate predictions with predictive uncertainties; these methods fit astronomical data because of their increasing size, which they need to perform well. However, given predictive uncertainties by a method, how do we know they are correct? Predictive uncertainties are often derived from predictive distributions (e.g. their variances), and correct predictive distributions have to be sharp subject to calibration. The dispersion of a predictive distribution evaluates its sharpness, while the probability integral transform (PIT) histogram evaluates its calibration. We exemplify the necessity of using PIT histograms on spectroscopic redshift determination, where we have to produce predictive distribution and not only point estimates because of ambiguous spectral line patterns.