“A deep-learning search for technosignatures of 820 nearby stars,” led by Peter Xiangyuan Ma and published in Nature Astronomy in 2023, applied deep learning to the search for extraterrestrial intelligence (SETI). The team reanalyzed a Breakthrough Listen dataset of 820 unique target stars observed with the Robert C. Byrd Green Bank Telescope, totaling more than 480 hours of on-sky data.
The central challenge in SETI is separating a hypothetical alien transmission from the enormous amount of human-made radio interference that swamps the same frequencies. The authors trained a beta-Convolutional Variational Autoencoder - a neural network that learns to compress and reconstruct signals - to flag anomalies in a semi-unsupervised way, so it could surface unusual signals without being told in advance exactly what to look for. The method returned eight promising signals of interest worth re-observing that earlier classical pipelines had missed.
As with every SETI search to date, the flagged candidates did not turn out to be confirmed alien technology; the value of the work was methodological, showing that anomaly-detecting neural networks can mine archival radio data for signals that rule-based searches overlook. The approach is part of a broader shift toward machine learning in radio astronomy, where surveys generate far more data than humans can inspect by hand.