Scientists have developed a machine learning method they think could help filter out interference and more efficiently spot unusual radio signals from space, contributing to the ongoing search for extra-terrestrial intelligence.
Search for extraterrestrial intelligence (SETI) programmes have used radio telescopes for decades to detect unambiguous artificial signals coming from the firmament. However, this search is complicated by interference from human tech, which can generate false positive identifications that are time-consuming to filter out from large data sets.
“We simulated a host of signals, injected them into real observations and trained the random forest component to classify those simulations,” the SETI Institute told The Register.
“The autoencoder part is trained on both real observations and simulations in recreating the original inputs, thus helping us extract salient features about the input image. Together this helps build an effective anomaly detection algorithm.”
Research led by Peter Ma, third year physics and mathematics undergraduate at the University of Toronto, used observations from 820 stars, in the form of 115 million snippets of data. The deep learning models the team developed using ML library TensorFlow and Python library Keras, identified around 3 million signals of interest. The group was whittled down to 20,515 interesting signals, which is more than 100 times less than previous analyses of the same dataset, the authors claimed.
They went on to identify eight previously undetected signals of interest, although follow-up observations have not succeeded in redetecting these targets, according to a paper published in Nature Astronomy.
The authors suggest their method could be applied to other big datasets to accelerate SETI and similar data-driven surveys.
“SETI aims to answer this question by looking for evidence of intelligent life elsewhere in the galaxy via the ‘technosignatures’ created by their technology. The majority of technosignature searches so far have been conducted at radiofrequencies, given the ease of propagation of radio signals through interstellar space, as well as the relative efficiency of the construction of powerful radio transmitters and receivers,” the authors said.
“The detection of an unambiguous technosignature would demonstrate the existence of extraterrestrial intelligence (ETI) and is thus of acute interest to both scientists and the general public,” they argued.
Other applications of ML in the SETI, include a generic signal classifier for observations obtained at the Allen Telescope Array and at the Five-hundred-meter Aperture Spherical Radio Telescope, convolutional neural network-based radio frequency interference identifiers, and anomaly detection algorithms, the authors said.
One of the most famous projects in the field was SETI@home, which sent radio telescope readings to volunteers’ home computers to sift for potential signs of extraterrestrial life for more than 20 years, but stopped sending data in 2020.
The project was overseen since 1999 by the Berkeley SETI Research Center, which manages several related initiatives, and has used about 1.5 million days of computer time. Although it did not achieve its goal of pin-pointing intelligent extra terrestrial life, it successfully demonstrated volunteer computing projects could use Internet-connected computers as a viable analysis tool, out-scaling the world’s largest super-computers. ®