Pitfalls of AI classification of rare objects: Galaxy Mergers

TL;DR

This paper discusses the challenges and limitations of using AI for classifying rare galaxy mergers, highlighting how current methods may miss morphological details and lead to contaminated samples, impacting future large-scale surveys.

Contribution

It proposes enhancing galaxy merger classification by incorporating pre-extracted morphological data to improve neural network performance and reduce contamination.

Findings

Adding morphological data improves classification accuracy.

Current neural networks do not fully utilize image information.

Sample contamination can significantly affect survey results.

Abstract

Galaxy mergers are hugely important in our current dark matter cosmology. These powerful events cause the disruption of the merging galaxies, pushing the gas, stars and dust of the galaxies resulting in changes to morphologies. This disruption can also cause more extreme events inside the galaxies: periods of extreme star formation rates and the rapid increase in active galactic nuclei activity. Hence, to better understand what goes on in these rare events, we need to be able to identify statistically large samples. In the last few years, the growth of artificial intelligence techniques has seen application to identifying galaxy mergers. These techniques have shown to be highly accurate and their application has grown beyond academic studies of ``can we?'' to deeper scientific use. However, these classifications are not without their problems. In this proceedings, we will explore how galaxy merger classification can be improved by adding pre-extracted galaxy morphologies alongside the traditional imaging data. This demonstrates that current neural networks are not extracting all the information from the images they are given. It will also explore how the resulting samples of rare objects could be highly contaminated. This has a knock on impact on the upcoming large scale surveys like Euclid and Rubin-LSST.