Astronomaly Protege: Discovery Through Human-Machine Collaboration

TL;DR

Astronomaly Protege is a human-machine collaborative framework that efficiently identifies scientifically interesting and diverse anomalies in large astronomical datasets using active learning and self-supervised deep learning.

Contribution

It introduces Protege, an extension of Astronomaly, that leverages self-supervised learning and active anomaly detection to recommend diverse, scientifically relevant sources with minimal human labeling.

Findings

Protege effectively recommends diverse, interesting sources with minimal training.

It can identify sources that even the authors missed.

The framework enhances discovery potential in large astronomical datasets.

Abstract

Modern telescopes generate catalogs of millions of objects with the potential for new scientific discoveries, but this is beyond what can be examined visually. Here we introduce Astronomaly: Protege, an extension of the general purpose machine learning-based active anomaly detection framework Astronomaly. Protege is designed to provide well-selected recommendations for visual inspection, based on a small amount of optimized human labeling. The resulting sample contains rare or unusual sources which are simultaneously as diverse as the human trainer chooses, and of scientific interest to them. We train Protege on images from the MeerKAT Galaxy Cluster Legacy Survey, leveraging the self-supervised deep learning algorithm Bootstrap Your Own Latent to find a low-dimensional representation of the radio galaxy cutouts. By operating in this feature space, Protege is able to recommend interesting sources with completely different morphologies in image space to those it has been trained on. This provides important advantages over similarity searches, which can only find more examples of known sources, or blind anomaly detection, which selects unusual, but not necessarily scientifically interesting sources. Using an evaluation subset, we show that, with minimal training, Protege provides excellent recommendations and find that it is even able to recommend sources that the authors missed. We briefly highlight some of Protege's top recommendations, which include X- and circular-shaped sources, filamentary structures and one-sided structures. These results illustrate the power of an optimized human-machine collaboration such as Protege to make unexpected discoveries in samples beyond human-accessible scales.