Machine Learning Anomaly Detection with PyTorch in WFC3/IR Images

WFC3/IR data has shown a range of known anomalies that are consistently occurring and have known corrections using pipeline processing. The Quicklook project is a data management software for quick access to and inspection of Hubble Space Telescope Wide Field Camera 3 data. One of the features of the project is anomaly detection, which allows Quicklook team members to visually inspect new observations and flag them for anomalies. We introduce a method for creating a deep learning algorithm to complement the existing Quicklook software by automatically detecting known and unknown WFC3 image anomalies, thus improving detection accuracy and reducing time spent on manual image inspection. The results from our algorithm using a test set of Hubble Ultra Deep Field images show R2 values >= ~0.97, and a reconstructive loss consistently below 10^-3 .