Parallel sessions

Machine learning has become fundamental to every astrophysics domain, promising high quality estimation of everything from gravitational wave parameters to weak lensing maps. At the same time, our current ML methods are often biased, overconfident, uninterpretable, brittle, and otherwise challenging to do useful science with. This session focuses on overcoming these challenges by sharing our ideas and our experiences experimenting with new methods. Accepted RAS ML conference abstracts have historically focused on direct performance, perhaps as a consequence of the competition for slots (the ML track was the most over-submitted category at NAM 2021). Unlike scientific domain tracks, where results have a direct bearing on how other science results are interpreted, highlighting excellent results is less helpful to the audience than how and why they were achieved. We feel a new format is needed to allow more researchers to share their ideas. The standard 12+3 minute talks restrict us to approximately 14 speakers, making it difficult to produce a program covering all the relevant domains (transients, galaxy images, spectra, exoplanets, etc.) and leading to early career researchers being 'outcompeted' for slots. Conversely, we feel the 1 minute 'lighting' talks introduced last year are too short to meaningfully convey both the scientific context and the machine learning approach applied. We therefore propose a program mixing invited speakers (15+5), spotlight (i.e. standard) talks (12+3) and ideas talks (4+1). 'Ideas talks' are short talks tightly focused on lessons useful to the community. These need not cover typical paper-style projects. They might cover interesting experiments like uncovering bias in a major currently-used ML catalog or projects where failure is informative; like finding that practical performance of a fancy new model on astronomical data fell far short of what CS literature claimed it would be (and how that might inform new algorithms). We also invite talks on best practices for reproducibility and software development in AI.