Abstract:
With the advent of modern radio interferometers, a significant influx in data is expected. This influx will render the manual inspection of samples infeasible and thus necessitates the development of automated approaches to find radio sources with anomalous morphologies. In this paper, we investigate the use of autoencoders for anomalous source detection, based on the assumption that autoencoders will reconstruct anomalies poorly. Specifically, we compare an autoencoder architecture from the literature to two other autoencoder architectures, as well as to four conventional machine learning models. Our results showed that the reconstruction errors of these autoencoders were generally more informative with respect to identifying anomalies than machine learning models were when trained on PCA components. Furthermore, we found that the use of a memory unit in our autoencoders resulted in the best performance, as it further restricted the ability of autoencoders to generalize to anomalous sources. Whilst investigating the use of different reconstruction error metrics as anomaly scores, we determined that they were more informative when combined than they were in isolation. Thus, applying the machine learning models to the combined anomaly scores from the autoencoders resulted in the best overall performance. Particularly, random forests and XGBoost models were the most effective, with isolation forests also being competitive when using a small number of labelled anomalies to tune their hyperparameters. Such isolation forests are also more likely to generalize to unseen classes of anomalies than supervised models such as random forests and XGBoost.