In trying to understand the universe, we become obsessed - we are attracted by the thirst for observation. Satellites transmit hundreds of terabytes of data information every year, and just one telescope in Chile will produce 15 terabytes of space images every night. No human can handle them manually. As astronomer Carlo Enrico Petrillo says, “Looking at pictures of galaxies is the most romantic part of our job. The problem is how to stay focused. Therefore, Petrillo is developing an AI that will help him.
Petrillo and his colleagues were looking for a phenomenon that is essentially a space telescope. When a massive object (a galaxy or black hole) gets caught between a distant light source and an observer on Earth, it bends space and light around it, creating a lens that allows astronomers to take a closer look at incredibly old and distant parts of the universe hidden from our view. This effect is called gravitational lensing, and these lenses are the key to understanding what the universe is made of. Until now, finding them has been slow and tedious.
This is where artificial intelligence is needed - and the search for gravitational lenses is the very beginning. As Stanford professor Andrew Ng put it, the ability of AI to automate everything that "a typical person can do in less than one second of thinking." Less than a second may not sound like a lot, but when it comes to sifting through large amounts of data, it's a godsend.
The new wave of astronomers is looking at AI for more than just a data sorter. They are exploring something that could be a completely new way of searching for scientific discoveries, where artificial intelligence will display parts of the universe that we have never seen.
But first: gravitational lenses. Einstein's general theory of relativity predicted this phenomenon as early as the 1930s, but the first examples did not appear until 1979. Why? Because space is very, very large, and people took a long time to observe it, especially without modern telescopes. The hunt for gravity lenses was challenging.
“The lenses we have now have been found in a variety of ways,” says Lilia Williams, professor of astrophysics at the University of Minnesota. “Some were discovered by accident, people were looking for something completely different. Some were found by people who were looking for them, the second or third time."
Promotional video:
The AI is very good at looking at pictures. So Petrillo and his colleagues turned to a beloved AI tool in Silicon Valley: a type of computer program composed of digital "neurons" modeled after real ones that fire in response to input. Feed these programs (neural networks) a lot of data and they will learn to recognize patterns and patterns. They work especially well with visual information and are used in a variety of machine vision systems - from cameras in self-driving cars to facial recognition in pictures on Facebook.
As it was written in an article published last month, applying this technology to hunting gravitational lenses was surprisingly simple. First, the scientists made a dataset to train the neural network - they generated 6 million fake images with and without gravitational lenses. Then we fed our data to the neural network and left it to figure out the patterns. A little tweaking and the result is a program that recognizes gravitational lenses in the blink of an eye.
“A great classifier in the human face parses images at a speed of a thousand per hour,” says Petrillo. One lens is found approximately once every 30,000 galaxies. Therefore, the classifier will have to work without sleep and rest for a week to find only five to six lenses. A neural network, by comparison, parses 21,789 images in just 20 minutes. And this is with one ancient processor.
The neural network was not as accurate as the computer. So that she did not overlook the lens, she was given broad parameters. She came up with 761 possible candidates, which humans studied and reduced to 56. Confirming that these are real lenses will have to be verified and confirmed, but Petrillo believes a third will be real. That's about one lens per minute, compared to a hundred lenses discovered by the entire scientific community over the past few decades. The speed is incredible, the prospects are huge.
Finding these lenses is essential to understanding one of the great mysteries of astronomy: what is the universe made of? The matter that we know (planets, stars, asteroids, etc.) represents only 5% of all physical matter, and another 95% is completely inaccessible to us. This 95% is represented by hypothetical matter - dark matter, which we have never observed directly. We just have to study the gravitational influence that it has on the rest of the universe, and gravitational lenses serve as one of the most important indicators.
What else can AI do? Scientists are working on a number of new tools. Some, like Petrillo, take on the task of identification: they classify galaxies, for example. Others scour data streams for interesting signals. Some neural networks remove artificial interference for a radio telescope by isolating only useful signals. Others have been used to identify pulsars, unusual exoplanets, or improve low-resolution telescopes. In short, there are many potential uses.
This explosion is due in part to general hardware trends that are expanding the field of AI, such as the availability of cheap computing power. Astronomers no longer need to sit out their pants on cloudless nights, observing the movement of individual planets; instead, they use a sophisticated technique that scans the sky one by one. Improved telescopes and data storage technologies mean there is even more room for analysis, Williams says.
Analyzing large datasets is what artificial intelligence is great at. We can teach him to recognize patterns and make him work tirelessly, and he will never blink or make mistakes.
Are astronomers worried that they trust a machine that may lack human understanding to detect something sensational? Petrillo says no. "In general, humans are more biased, less efficient, and more error prone than machines." Williams agrees. "Computers may miss certain things, but they will systematically miss them." But as long as we know what they don’t know, we can deploy automated systems without much risk.
For some astronomers, the potential of AI goes beyond simple data sorting. They believe that artificial intelligence can be used to create information that fills the blind spots in our observations of the universe.
Astronomer Kevin Schawinski and his team in astrophysics of galaxies and black holes are using AI to improve the resolution of blurry telescope images. To this end, they deployed a neural network that generates unrivaled variations in the data under study, as if a good forger imitates the style of a famous artist. These same networks were used to create fake images of star images; fake audio dialogs simulating real voices; and other types of data. According to Shavinsky, such neural networks create information that was previously inaccessible to us.
In a paper published by Shavinsky and his team earlier this year, they showed that these networks can improve the quality of space imagery. They lowered the quality of images of a number of galaxies, added noise and blur, and then passed them through neural networks along with the original images. The result was amazing. But scientists cannot share it yet.
Shawinski is wary of the project. After all, it goes against the basic tenets of science: you can only know the universe by observing it directly. “For this reason, this tool is dangerous,” he says. And it can only be used when we have accurate data and when we can verify the result. You can train a neural network to generate data about black holes and send it to work in a certain area of the sky that has been poorly explored until now. And if she finds a black hole, astronomers will have to confirm the find with their own hands - as is the case with gravitational lenses.
If these methods prove to be fruitful, they can become entirely new research methods, complementing classical computer simulations and good old observation. So far, everything is just beginning, but the prospects are very promising. "If you had this tool, you could take all the data from the archives, improve some of it, and extract more scientific value." A value that wasn't there before. AI will become a scientific alchemist, helping us transform old knowledge into new knowledge. And we could explore space like never before without even leaving Earth.
Ilya Khel