Debating the potential of machine learning in astronomical surveys

A major revolution is now underway in astrophysics with the constant arrival of ever-richer and more complete datasets. The next generation of surveys soon starting will generate orders of magnitude more data than previously. However, it is becoming increasingly clear that traditional techniques are not up to the challenge of fully exploiting these data. At the same time, in the computer industry, large-scale application of “machine learning” methods on large quantities of data have been able to solve problems that until now have been intractable. These new techniques are now being adopted enthusiastically by astronomers who see them as a way to extract the maximum amount of science from new surveys. However, some caution is required; the concerns of industrial players developing machine learning techniques are not the same as astrophysicists who seek to explain observations in the framework of physical models. It is therefore very timely to survey the landscape of machine-learning techniques in astronomy and to critically evaluate their usefulness for solving astrophysical problems. At IAP, many advanced analysis techniques have been pioneered to analyse data from the Planck satellite and from ground-based surveys; the avalanche of rich datasets coming from future missions like Euclid and ARIEL makes the institute a natural location to hold this conference. The conference will explore the potential and applicability of machine learning techniques for future surveys such as DESI, SKA, Euclid, Rubin Observatory, Ariel and Gaia. In particular the impact of systematic errors on the reliability of inferred parameters (cosmological or otherwise) derived using these methods will be explored. The ability of machine-learning models to lead to scientific discoveries will be critically discussed.