Weakly-Supervised Anomaly Detection in the Milky Way

Large-scale astrophysics datasets, such as the 1 billion Milky Way stars from the Gaia satellite, present an opportunity for novel machine learning techniques to identify regions of interest that might otherwise be overlooked by traditional techniques. We use a weakly-supervised anomaly detection technique, Classification Without Labels (CWoLa), to identify cold stellar streams within the Gaia dataset. CWoLa operates without the use of labeled anomalies or knowledge of astrophysical principles. Instead, we train a classifier to distinguish between mixed samples in which class proportions need not be known. This technique, originally designed for collider physics, has broad applicability within astrophysics as well as other domains interested in identifying anomalous localized features.