Plenary talk: Machine learning for SETI

A large corpus of interesting signal events from the Allen Telescope Array (ATA) is investigated to help find signs of extraterrestrial intelligence. The problem of identifying and categorizing different types of radio frequency interference (RFI) in unsupervised and semi-supervised settings are approached with representation learning using various types of autoencoding and self-supervised deep learning. The excision of RFI allows the remaining signals to be explored to find anomalies worthy of follow-up measurements. This collaborative work with the SETI Institute delves into topics of FFT-based feature engineering, automated representation learning, density equalization for enabling unsupervised learning and dimensionality reduction and clustering for RFI categorization. We find that there are dozens of categories of RFI captured by the ATA, but that a semi-supervised approach is required to allow for careful grouping.