Machine learning (ML) techniques are well-known to excel at pattern matching, but using a combination of semi-supervised and active learning, David O’Ryan and Pablo Gómez have demonstrated that an ML method can find exceptions to the rule too. With a tool called AnomalyMatch, 100 million image cutouts from the Hubble Legacy Archive were scrutinized over 2–3 days and the morphologically irregular — rather than the regular — astronomical objects were flagged. The rare objects unearthed in this study include 86 new candidate gravitational lenses, 18 uncommon jellyfish galaxies, 2 collisional ring galaxies and 417 unusual merging or otherwise interacting galaxies.
In the big data era of astronomy, automated methods are needed to identify objects of interest in large datasets because they are too vast to be scanned by humans. Automating these searches also reduces issues connected to human subjectivity (which may be encountered, for instance, in crowdsourcing efforts such as Galaxy Zoo). These techniques are even more powerful if the objects of interest are rare, or have never been seen before, because finding them improves our understanding of the cosmos. However, the lack of training data in these cases makes supervised methods unsuitable. AnomalyMatch treats anomaly detection as a binary classification problem (nominal or abnormal?) applied to extended source cutouts, relying on 5–10 labelled anomalies and the human verification and correction of positive results (active learning). Anomalies were classified into source types by humans, using SIMBAD and ESASky. Overall, ~65% of flagged anomalies were previously unknown objects, and ~10% of classified anomalies were found to be false positives.
This is a preview of subscription content, access via your institution
