Speaker
Description
We investigate a strategy for improving radio frequency interference (RFI) flagging using roughly 60 hours of observations from the 2016 EoR high-band observing run of the Murchison Widefield Array. We develop a novel machine learning-based \textit{weak supervision} method where traditional flagging methods such as AOFlagger, SSINS, $\chi^2$, X-RFI and EAVILS are used to generate stronger labels for training. Visibilities in observations are aggregated into statistical summaries and treated as tabular data for a transformer-based classifier. To evaluate our classifier we create a large, 10-hour, human-annotated ground-truth dataset enabled by novel annotation software. We find that our models outperform the traditional RFI flagging methods on F1 score and are able to flag a wide variety of RFI events.