Machine learning helps hunt for ET

The global community dedicated to searching space for signs of aliens, SETI, the Search for ExtraTerrestrial Life, works under the assumption that we’re looking for signs of technology - a proxy for intelligence. The kinds of technology they search for include radio transmissions, which a sophisticated civilization similar to us would probably have made use of as a fast means of communication. The search for these signals is hampered, however, by false positive results returned from man-made radio signals which blind telescopes pointing at far away stars. Now, a team at the University of Toronto has developed a machine learning algorithm capable of combing through the thousands of signals of interest SETI telescopes return to reduce these pesky false positives, and can look for patterns in radio signals other algorithms miss. I spoke with Cherry Ng and Peter Ma to hear how they’ve done it…

Peter - Machine learning approaches problems in a unique way. It looks for patterns in the data without being instructed what the actual patterns are. So traditional algorithms are like baking a cake: you have a set of instructions which you follow to execute this process. That's if you know what you want to look for. In this case, we want to look for specific kinds of signals but, for ET, we don't actually know what exactly they're going to send us. And so what we want to do is we actually want to get the computer to decide different kinds of anomalies that are detected in our data by using machine learning. We're casting a wider net than we have ever done before with traditional algorithms. You can think of traditional algorithms as like preset conditions of things that I exactly want to look for and have to meet exactly to register as a hit or register as a potentially positive event. But for machine learning, it makes none of those assumptions and it just makes these decisions based on the data it has seen. And we've trained on hundreds and hundreds of thousands of examples such that when it does look at this data, it can make decisions based on what's actually out there.

James - So in terms of how much laborious work, how many hours you're able to save as a result of this new method, are we talking order of magnitude more efficient?

Peter - For my algorithm, it runs almost twice as fast as it takes to observe. In other words, if it takes 30 minutes to observe, it takes like 15 minutes to process or like 18 minutes-ish to process the actual data. The classic algorithm or what we call turboSETI, at least by the time when I was writing this paper, took almost an hour to run through the same piece of data. So it is significantly faster.

Cherry - Peter's machine learning algorithm is a lot faster. We were able to process the entire 800 star data set in like two weeks and with very modest computing resources. So indeed, one huge advantage of the machine learning algorithm is that we can search a much larger parameter of space than what a traditional algorithm might be able to do just because it's so much more efficient.

James - And I know this is not your guys' area necessarily, but indulge me a bit. What's the protocol for if we find some radio signal from space? If all this work bears some fruit and we find something that isn't man-made, what then?

Cherry - Yeah, I think what people say is an extraordinary claim requires extraordinary evidence. If we do find something interesting, the first most important thing is to confirm it, to be really sure that it is a genuine signal. And the only way to find out is by re-observing and re-detecting it also with different telescopes by involving other scientists to see if we can independently confirm these signals and then try to extract as much information from the data we have as possible. Is it from an exo planet? Can we find out about the radio velocity or any other information? And then if we are really sure, then I think it requires international collaboration to try and come up with a coherent plan of how to establish any communication.