Search For Extraterrestrial Intelligence: New AI Technique Helps To Separate Noise From Sound, Throws Up Eight Exciting Signals Previously Missed

Eight new radio signals suggest signs of extraterrestrial life, although an early follow-up hasn’t offered confirmation.

What’s especially exciting is that what could potentially be transmissions from extraterrestrial civilisations, were discovered using machine learning, a subset of artificial intelligence (AI).

The new algorithm was used to sieve data from telescopes and thereby help to distinguish between real “alien” signals and Earthly interference. The search was performed in July 2021.

The latest technique, along with the results it threw up, is revealed in a new paper in Nature Astronomy, published on 30 January. 

Its first author, undergraduate student Peter Ma of the University of Toronto, is behind the powerful algorithm. He devised it in high school, but his teachers at the time didn’t appreciate the novel approach. Now it’s showing remarkable promise.

Astronomers have been training their radio telescopes at different regions in the sky for decades in search of signs of extraterrestrial intelligence. But not all signals are useful. The telltale signs would be “technosignatures,” short for technologically generated signals.

If any alien civilisation is able to reach out to Earth — intentionally or unintentionally — across the vast of expanse of space, it would be through — since it would be capable of emitting — advanced signals.

Since our civilisation is advanced as well (a reluctant humble brag), the signals emerging in and around the blue planet — such as from mobile phones, the global positioning system (GPS), and television stations — would interfere with those coming in from an advanced civilisation in a faraway solar system.

This necessitates the process of sieving out the real signals from “the fake,” in the context of search for extraterrestrial life. The classical search algorithm, which Ma describes as “now older than my parents,” wasn’t quite sufficient. Here is where the new AI technique has come in handy.

“The goal for this shiny new algorithm is to run faster and to produce better candidates by leveraging AI and modern computer vision techniques,” Ma writes in his account of the discovery.

The algorithm leverages, Ma explains, “unsupervised and supervised learning paired with a novel transfer learning method.”

The classical method was predominantly constrained by supervised models — for signals to be deemed noteworthy, they had to look like signals of a certain kind, just as they were simulated.

But what about the signals that didn’t comply with the expectations? After all, AI algorithms cannot think for themselves and grow in the traditional sense, as humans do. 

Where unsupervised learning models were involved, a whole bunch of “junk” signals would be detected as warranting attention.

Ma’s technique sought to strike a balance between supervised and unsupervised learning.

“Peter inserted simulated signals into real data, and then used this dataset to train an AI algorithm called an autoencoder. As the autoencoder processed the data, it “learned” to identify salient features in the data,” writes Danny C Price, a senior research fellow at Curtin University and one of the authors of the study.

Then, an algorithm called a “random forest classifier” came into play, which helps to tell “if a signal is noteworthy, or just radio interference – essentially separating the technosignature “needles” from the haystack,” Price adds.

The new machine learning algorithm was then put to work on 150 TB of data (480 observing hours) from 820 nearby stars. This dataset was combed through previously in 2017 with no signals of interest to show for it. Now it looks like scientists had missed something.

Initially, as many as 30,000 results demanded a closer look. A manual inspection by Ma brought that number down to “more than 10” and then, finally, to eight.

The new technique led to the discovery of eight new radio signals that went undetected in the earlier analysis, not involving machine learning. These signals came from five different stars, located 30 to 90 light years away from Earth. 

They are noteworthy for a few reasons. For one, they are all “narrowband” signals, rarely produced by natural sources. Signals caused by natural phenomena tend to be “broadband.” Narrow band refers to a narrow spectral width, on the order of just a few Hertz.

Additionally, these signals appeared when the telescopes were pointed towards the targets and disappeared when looking away. Local interference, on the other hand, is always switched on.

Plus, the detected signals changed in frequency over time in a way that made them appear far from the telescope. It was an indication of the source being far away — and, therefore, not local interference — as it registered some relative acceleration with respect to Earth.

Still, in a sample set extending to millions, it can be the case that a handful of signals, by sheer chance, exhibited these noteworthy characteristics, according to Dr Steve Croft, Project Scientist for Breakthrough Listen on the Green Bank Telescope.

Therefore, these signals aren’t yet locked down as signs of extraterrestrial intelligence. Follow-up observations using the 100-metre Robert C Byrd Green Bank Telescope, whose gathered data was analysed using the new AI technique, didn’t throw up a confirmation either.

It’s possible that these eight signals are “rare cases of radio interference,” according to Price.

Researchers are waiting for the same signals to emerge again. They will keep monitoring the eight candidates. Meanwhile, further observations and analyses are underway.

Going forward, teams working on the search for extraterrestrial intelligence (SETI) plan to expand the use case of this new machine learning algorithm with the goal of uncovering new and exciting revelations.

They want to scale up the search from 800-odd stars to a million with the MeerKAT array of 64 radio telescope in South Africa and beyond. It promises to speed up the SETI search.

“This machine-learning approach presents itself as a leading solution in accelerating SETI and other transient research into the age of data-driven astronomy,” the researchers say in the paper.

However, the classical algorithms are not likely to go anywhere, as they “remain excellent at picking up candidate signals,” Jean-Luc Margot, an astronomer at the University of California, Los Angeles, told Nature.

Ma’s project work at the undergraduate level was funded by the Laidlaw Foundation. 

The broader alien detective work is being carried out as part of the Breakthrough Listen project. Listen is part of the Breakthrough Initiatives, launched by Yuri Milner and Stephen Hawking in 2015.

Breakthrough Listen is a $100 million scientific programme aimed at finding evidence of civilisations beyond Earth.

It includes a survey of the 1,000,000 closest stars to Earth and scans the centre of our galaxy, the Milky Way, and the entire galactic plane. Beyond the Milky Way, it keeps its ears open for messages from the 100 galaxies closest to us.

Researchers are working on data taken from the Green Bank Telescope (the world’s largest steerable radio telescope), the Parkes radio telescope in Australia, and the Automated Planet Finder (a robotic optical telescope equipped with cutting-edge spectrograph technology).

SETI experiments began in 1960 with Frank Drake’s Project Ozma at the Greenbank Observatory. But there's been a proliferation of data in recent years. New computational methods have, thus, become necessary to identify anomalous data pointing to signs of extraterrestrial life.

Also Read: Frank Drake Lives On In The Spirit Of SETI