Prospecting MeerKAT Continuum Data for Enigmatic Radio Sources with Unsupervised Vector-Quantised Variational Autoencoders

TL;DR

This work tackles the challenge of locating anomalous radio sources in large MeerKAT MGCLS 1.28 GHz continuum images by applying Vector Quantised Variational Autoencoders ($VQ$-$VAE$) to learn a discrete latent distribution of typical morphologies. It compares semi-supervised (A) and fully unsupervised (B, C) training regimes, using the reconstruction score $NCC$ and binary metrics such as $AUC$, $F_{1}$, and $F_{2}$ to flag anomalies, while visualising latent spaces with $t$-SNE and UMAP. Key findings show that training on unlabelled data (B) yields improved anomaly detection over training on labelled typicals (A); however, performance under real-world class imbalance (C) remains modest, underscoring the need for expert review and context-aware thresholding. The study provides a publicly available labelled MGCLS test set and demonstrates the practicality of VQ-VAEs as a scalable tool to triage vast radio surveys ahead of SKA-era data, while acknowledging subjectivity in defining anomalies and the persistence of false positives.

Abstract

We present a novel application of Vector quantised variational autoencoders (VQ-VAEs) to deep 1.28 GHz radio continuum images taken from the MeerKAT Galaxy Cluster Legacy Survey (MGCLS).VQ-VAEs are deep learning models widely used in modern computer vision applications and pipelines. Designed for image generation, VQ-VAEs are trained to reconstruct the input dataset via a low-dimensional discrete embedding. VQ-VAEs effectively learn the distribution of training data, where samples that do not fit the distribution well yield the highest reconstruction errors. This property makes VQ-VAEs a good candidate for the task of anomaly detection. In this work, we examine the effectiveness of VQ-VAEs in identifying radio continuum sources with anomalous structures in the image-plane domain. We find VQ-VAEs to be useful as part of a solution for searching such large datasets. We observe that they are able to remove a majority of the typical sources in such data, even when trained in an unsupervised manner on unlabelled data. We also provide our testing set of a large sample of manually labelled radio sources, in particular radio galaxies, taken from the MGCLS at 1.28 GHz. Automated approaches to searching through high volumes of data are key in extracting the full scientific potential of the Square Kilometre Array and its pathfinders.

Figures (14)

• Figure 1: Basic structure of a VQ-VAE. The encoder reduces the input to the dimensions of the latent space. A code form is produced from a learned codebook of discrete vectors. The decoder attempts to reconstruct the original input from the code form.
• Figure 2: Comparison of the distribution of $F_1$ and $F_2$ scores for models in configurations A and B. Models in configuration A are trained on labelled data while those in configuration B are trained on the unlabelled data. Both are tested with testing sets having an equal number of exotic and typical sources. The five number summary for each distribution shows the median in the centre of the box, with half of all values falling within the box. The whiskers show the minimum and maximum, excluding the outliers which are marked by circles.
• Figure 3: (Left) Box and whisker diagram for AUC in configuration A (Centre) Box and whisker diagram for AUC in configuration B (Right) Box and whisker diagram for AUC in configuration C. The five number summary for each distribution shows the median in the centre of the box, with half of all values falling within the box. The whiskers show the minimum and maximum, excluding the outliers which are marked by circles.
• Figure 4: Averaged confusion matrix of the model in configuration A, trained on labelled typical images with an equal number of typical and exotic images in the test set (left). Averaged confusion matrix of the model in configuration B, trained on all unlabelled images with an equal number of typical and exotic images in the test set (centre). Averaged confusion matrix of the model in configuration C trained on all unlabelled images with all labelled images being used for the test set (right).
• Figure 5: Average ROC with AUC of the VQ-VAE in configuration B with labelled training data and tested on equally sized test sets. The shaded region represents the standard deviation in the curve.