Based on materials from The Verge
In our attempts to understand the universe, we become increasingly greedy, and our observations often outpace our understanding. Satellites transmit hundreds of terabytes of information to us every year, and just one telescope, which is now being built in Chile, will allow us to receive 15 terabytes of pictures of the starry sky every night. Humanity is simply not able to process so much information. As astronomer Carlo Enrico Petrillo says, “Looking at pictures of galaxies is the most romantic part of our job. The problem is staying focused. ' That is why Petrillo trained artificial intelligence to do this job for him – observation.
Petrillo and his colleagues are looking for phenomena that are essentially 'space telescopes'. When a large object (a galaxy or black hole) is caught between a distant light source and an observer on earth, it distorts the space and light around it, forming a lens that allows astronomers to see up close to very old, distant parts of the universe that should not be visible. This phenomenon is called a gravitational lens, and such lenses are the key to understanding what the universe is made of. However, finding them is a long and tedious job.
And this is where artificial intelligence comes in. And the search for gravitational lenses is just the beginning. As Stanford University professor Andrew Ng once said, artificial intelligence is capable of automating everything “what the average person can do […] in less than a second of thought.” Less than a second is not a time that seems like enough time to think, but when it comes to using the huge amounts of data that modern astronomy creates, time is worth its weight in gold.
AI astronomers aren't just looking for ways to make technology sort data. They are developing a possible radically new approach to scientific discovery, with AI mapping parts of the universe that we've never seen.
Gravitational lens. A nearby red galaxy distorted light from a more distant blue galaxy into a horseshoe shape. Source: NASA / APOD
But in the beginning – all the same gravitational lenses. Einstein's general theory of relativity predicted this phenomenon back in the 1930s, but the first prototype was not found until 1979. Why? The cosmos is very, very large, and in order to examine it, a person takes a very long time. Especially if he doesn't have a modern telescope. And so the hunt for gravitational lenses was a thankless job.
“The lenses we now know have been found in a variety of ways,” says Lilia Williams, professor of astrophysics at the University of Minnesota. – Some were discovered by accident, by people who were looking for something completely different. There were those who were found by people specifically looking for them, in just a couple of observations. But the rest were found by fluke. '
Viewing images is something that artificial intelligence is very good at. So Petrillo and his colleagues in Bonn, Naples and Groningen turned to the AI tool that is so beloved in Silicon Valley. It is a kind of computer program made up of digital 'neurons', modeled after brain cells, that respond to an incoming signal. Give these programs (neural networks) a huge amount of data, and they will begin to identify patterns. They are especially good at processing visual information, and they are used in all digital surveillance systems – from cameras in self-driving cars to Facebook with its recognition and the ability to tag people in photos.
As described in a recently published report, the application of this technology to hunting for gravitational lenses has proven surprisingly simple. Initially, scientists used a dataset to train a neural network, which involved creating 6 million images that showed how gravitational lenses should and should not look. And then the neural networks fed the data, allowing it to slowly identify patterns. A little bit of final adjustments, and you're done – here it is, a program that can instantly calculate gravitational lenses.
“A very cool person can evaluate images at about a thousand per hour,” says Petrillo. According to him, in the case of the data that this team of scientists had at their disposal, one lens per 30,000 galaxies is obtained. And a person, working without sleep and rest for a week, will be able to find only five or six lenses. For comparison, the neural network processes a database of 21,789 images in just 20 minutes. And this, Petrillo says, is possible with just one processor from an ancient computer. So the time can be greatly reduced.
However, the neural network lacks computer accuracy. In order not to overlook some of the lenses, their parameters were set in a rather general form. Humans looked at 761 of the network's proposed 'candidates', and after dropping out, the total was reduced to 56. Further observations are required to confirm the veracity of these findings, and Petrillo predicts that about a third will be genuine. This makes approximately one lens per minute. Compare this to about a hundred lenses discovered by the entire scientific community over the past few decades. Incredible progress and a great example of the use of AI in astronomy.
Finding these lenses is critical to understanding one of the biggest mysteries of astronomy: what our universe is made of. It is assumed that objects familiar to us (planets, stars, asteroids, etc.) are only 5 percent of physical objects, and the remaining 95 are other, unusual forms of the existence of matter. These are hypothetical forms like dark matter, which we have never observed directly. Instead, we study the gravitational effect it has on the rest of the universe, with gravitational lenses serving as one of the key metrics.
AI is being used in astronomy in a variety of ways, including sorting data from radio telescopes, like this one in Australia. Photo: Ian Waldee / Getty Images
What else is artificial intelligence capable of? People are developing many new tools based on it. Some, like Petrillo, have focused on identification, for example when classifying galaxies. Others scour huge streams of data looking for interesting signals, such as a neural network that filters human-generated signals that clutter up information from radio telescopes, thus helping scientists focus on potentially interesting signals. More neural networks are being used to identify pulsars, locate unusual exoplanets, or sharpen low-resolution telescope images. The potential applications are countless.
This surge is driven in part by broader hardware trends that have given rise to the widespread use of AI, such as lowering the cost of computing performance. However, the very nature of astronomy has also changed. Scientists no longer spend sleepless hours on cloudless nights, contemplating the path of individual planets. Now it is a complex algorithm that sifts the starry sky piece by piece, operating with huge chunks of data, unthinkable for scientists of old times. Better telescopes, more storage options, so more data to analyze than ever before.
Analyzing huge amounts of data is the strength of AI. And we can teach him to identify patterns and then make him work as a diligent assistant with unblinking eyes and unflagging attention.
Do astronomers worry that they have so much confidence in a machine that might lack human intuition to detect something sensational? Petrillo says he is not worried: “In general, people are more biased, less efficient and more prone to making mistakes than machines.” Williams agrees with him: “The computer can miss something by accident, but it will never do it systematically.” As long as we know what they don’t know, we can use automated systems without much risk.
A snapshot of many galaxies from the Hubble telescope. Curved strokes of light are gravitational lenses. Source: ESA / NASA
For some astronomers, the potential application of AI goes beyond conventional sorting of data. They believe that artificial intelligence can be used to create information and fill in the gaps, blind spots, in our exploration of the universe.
Astronomer Kevin Schawinski and his team, which specializes in the study of galaxies and black holes, used AI to sharpen blurry telescope images. To do this, they used a type of neural network that is excellent at creating variants of the data that are loaded into it. It is like a skilled artisan who can imitate the manner of a famous painter. Such neural networks are called generative adversarial networks (or GAN, for generative adversarial networks), in particular, they were used to create fake photos of celebrities, fake audio dialogs with simulated voices and other similar data. This is one of the richest layers of modern AI research, and for Shawinski, this means that information can be obtained that did not exist before. A report that he and his team published earlier this year showed how GANs can be used to improve the quality of space photography. Scientists downgraded the array of galaxy images by adding noise and blur, and then used a network trained on the telescope images to increase their resolution, and then compared the result with the original. The results were surprisingly accurate, enough to convince Schawinski that AI has the potential to improve the quality of all types of data in astronomy.
The scientist says he has something to share, but cannot divulge details until the results are published. He also expresses suspicion about the project. After all, in the end, it contradicts the basic principles of science: the universe can only be studied through direct observation. “And therefore it is a dangerous tool,” he says. And it can only be used if we a) have enough accurate data to train b) can verify the results. You can train a GAN to generate data about black holes and freely project it onto a part of the sky that has not been studied in detail before. And then, if there is an assumption about the presence of a black hole, astronomers have to check it with their own hands, as is the case with gravitational lenses. Shawinski says that, as with any scientific instrument, this one needs to be carefully and patiently tested, making sure the results are not misleading.
If these methods prove to be useful, they could become a new word in research, which Schawinski puts on a par with classical computer simulation and good old observation. We are at the beginning of the journey, but the effect can be huge. According to the scientist, “if you have this tool, you can apply it to all the existing data that is gathering dust in the archives, perhaps slightly improve them and get more scientific use out of them.” Add value that was not there before. AI will do something like scientific alchemy, transforming old knowledge into new knowledge. And we will be able to explore space deeply like never before, without even leaving Earth.