For some astronomers, the potential for the development and evolution of artificial intelligence (AI) goes far beyond mere data classification. They believe that artificial intelligence could be used to create information, filling in the blind spots in our observations of the Universe .
A group of astronomers from the Californian University of Berkley, who specialize in astrophysics of galaxies and black holes, have used artificial intelligence to improve the resolution of the telescope’s blurred images. To do this, they deployed a type of neural network that is optimal at generating variations of the data it studies, much like a well-trained forger who can mimic the style of a famous painter. These networks, called generative adversarial networks, or GANs, have been used to create fake faces based on images of celebrities; fake audio dialogue that mimics people’s voices; and a wide variety of other data types. They are one of the most developed variants of contemporary artificial intelligence research and, for researchers, it means obtaining information that did not exist before.
Imitate colors like a famous painter
The paper published by this group of Berkley researchers in late 2021 showed how generative adversarial networks could be used to improve the quality of images from space . They lowered the image quality of a bunch of galaxy images, adding noise and blurring, then used a specialized generative matching network on telescope images to increase their resolution, compared to the originals. The results were surprisingly precise: good enough to convince the scientific community that studies the cosmos that artificial intelligence has enormous potential to improve all kinds of data sets in the field of astronomy. The scientists involved in the project say they have “many interesting results in the works”, but they cannot reveal anything before they are published in the usual specialized scientific publications.
Even so, the researchers are still very cautious about the project. After all, it seems to go against the fundamental principles of science : that you can only learn about the Universe by directly observing it. “This is a dangerous tool for precisely this reason,” stated the group of scientists from the university, and should only be used when we have large and accurate databases, and can verify the results.
Therefore, a generative adversarial network can be used to generate data on black holes and then drop it on a part of the sky that has not been observed in much detail before. Then, if the data suggests that there is a black hole there, astronomers will confirm this finding firsthand, just as with gravitational lensing. The scientists say that, as with all scientific tools, rigorous and patient testing is necessary to ensure that the results you get don’t “divert attention from the thing you’re looking at,” and seriously misunderstand.
