We become greedy, more observations than we are aware of what to do, in our efforts to understand the Universe. Every year, the satellites collect hundreds of terabytes of information, and in Chile, a telescope produces 15 terabytes each night of photos of space. People can not sift it all. It is impossible. "Looking to images of galaxies is the most romantic part of our job." As astronomer Carlo Enrico Petrillo told The Verge. That is why Petrillo has trained an AI program to look for him. The problem is to remain focused.

Petrillo and his colleagues were looking for a space telescope phenomenon. When a massive object (a galaxy or a dark hole) enters a distant light source and the Earth observer, it bends space and light around and creates a lens which gives astronomers a closer look into unbelievably old, far-off sections of the Universe which should be blocked from viewing. It is called a lens of a gravity, which is a key to understanding the composition of the universe. However, it was sluggish and boring to find them to date.

Artificial intelligence comes in here — it's only the beginning to find gravitational lenses. As Professor Andrew Ng of Stanford once put it, the ability of AI is able to automate anything that "a typical person can [...] do with less than a second of thought." Less than a second sounds like little room to think, but when it comes to screwing up the vast quantity of data generated by contemporary astronomy, it is a gift.

First of all: lenses of gravity. This was predicted in the 1930s by Einstein's theory of general relativity, but only in 1979 was the first example found. Why does this happen? Well, it takes a long time for humans to look at it, especially without the telescopes of today. So far it has become a fragmentary affair to hunt for gravity lenses.

 

 

Liliya Williams, a professor of astrophysics at the University of Minnesota says to The Verge, "the lenses we have found right now in all kinds of ways." "Accidentally, people searched for something entirely different discovered some.

 

Some people were found by two or three surveys to search for them. The rest, however, have been found with servitude.

 

Looking at pictures is the very thing that an AI does. Petrillo and fellow students at universities in Bonn, Naples and Groningen therefore turned to Silicon Valley 's loved AI tool, which is a type of digital "neuron" computer program modeling after the brain's input that fires. Feeds other data into these systems and patterns (called neural networks).

 

They handle visual information particularly well and are used for powering all types of machine viewing systems — from cameras in autos to Facebook image tagging.

 

 

The use of this technology for gravitational lens was surprisingly straightforward, as described in a paper published last month. First of all, the researchers made a dataset to train the neural network to generate 6 million fake images to show what are and are not gravitational lenses. They then went off the data and let the neural network identify patterns slowly.

 

Later they had some fine tuning with a programme, which recognized lenses of gravity at a glance of the eye.

"We could classify images at a speed of about one millennium an hour in an exceptionally good human classification," says Petrillo. He estimates that every 30 000 galaxies one would find a lens with the kind of information his team used. Thus, a human classifier who works without sleep or rest for one week only plans to find 5 or 6 lenses. In just 20 minutes, the neural network broke down through a set of 21,789 images. And that was with a single old computer processor, says Petrillo. "It's possible to shorten this time by a lot," said he.

 

The neural networks were not so precise as a computer. Its parameters were quite generous in order to avoid overlooking any lenses. 761 potential candidates were produced, who were examined by humans and reduced to 56. There is a need to make more observations to confirm that these are valid results, but Petrillo estimates that approximately a third is the real deal. It works on about one per minute lens compared to the 100 or so that has been found in the past few decades by all the scientist community. It is an amazing acceleration and a good example of how AI can contribute to astronomy.

 

The search for these lenses is essential to understanding one of the great mysteries of astronomy. We know only 5 percent of all the material (planets, stars, asteroids, etc.); other weirder forms of matter make up the remaining 95 percent. This includes a hypothetical substance that we never directly observed, known as dark matter. Instead, we study the gravitational effect on the rest of the universe, with one of the most important indicators being gravitational lenses.

 

 

And how could AI do otherwise? A variety of new methods are being developed by researchers. Some, like Petrillo's, take up identification work: for example, classification of galaxies. Signals such as a neural network that removes interference with human-made radio telescope to help scientists get in possibly exciting signals via data streams. In addition, pulsar stars were identified, peculiar exoplanets found or images of the low-resolution telescope were sharpened. In short, the potential applications are booming.

 

This explosion is due in part to larger hardware trends that have led to an abundance of low cost computing power in the larger AI field. But this is also due to astronomy's changing nature. Astronomers don't continue to keep watch over cloudless nights, tracking the motion of individual planets; they are instead using sophisticated equipment, which blinkers portions of the sky in data which early scientists can't imagine. Better telescopes and better storing of data mean analyzing more than ever, Williams says.

 

Exactly what artificial intelligence is great at analysing large sections of data.

We can teach it to identify patterns and then put them to function like a tireless assistant.

 

Does astronomers worry about placing confidence in a machine that may lack the human insight needed to discover something sensational? He's not bothered, Petrillo says. Williams agreed: "Computers can miss certain things, but will systemically miss them." As long as we know what it's they know not we can deploy automated systems without too much risk. "In general, people are more inclined to be biased, less efficient and less susceptible to errors than machines.

 

The scope for AI goes beyond data sorting for other astronomers. They think that artificial intelligence can be used to create information and to fill our universe observations with blind spots.

 

Astronomer Kevin Schawinski used AI to sharpen resolution of blurry telescope images with his team specialized in galaxies and black-hole astrophysics. They have used a neural network like a well-trained fodder, capable of imitating a popular painter 's style, to produce variations in the data he studies.

 

These networks were used for creating fake faces on the basis of pictures of celebrities, fake audio dialog imitating the voices of the person and a variety of other data forms, called generative adverse networks or GANs. They are one of the richest seams in contemporary AI research and they meant to get information that had not been there before for Schawinski.

 

The paper published earlier this year by Schawinski and his team showed how GANs can be used to enhance the quality of space pictures. They lower the image quality of a group of galaxy images, add noise and bubbles and then use a GAN trained to improve their resolution compared to the originals in telescope imagery.

 

The results were strikingly accurate: sufficiently good to persuade Schawinski that AI could improve all kinds of datasets in astronomy. He says he and his team have "a lot of interesting pipeline tests," but before they're released they can't disclose much.

 

The idea is careful for Schawinski. After all, it sounds like a contravention of core science principles: that only by observing it directly can you learn about the Universe. "For precisely that reason this is a dangerous device," he says, and only if we a) have adequate, accurate training data, and b) can verify the results.

 

You can train a GAN to build black holes data, then set it on a portion of the sky which has not been closely observed before. And, if a black hole were located here — much like the gravitational lenses, astronomers will test this discovery on their own. Schawinski said there need to be rigorous and patient testing, as with all scientific tools, to ensure that the results do not "distract you."

Such methods can become a totally new exploration method if they are successful, with Schawinski placing alongside classic computer simulations and good old-fashioned observation. Early on, however, the payout might be enormous

 

"You can go to all the existing data in the records, maybe slightly improving some of it, and extract more scientific value if you have this tool," says Schawinsky. "Value that was never there before. AI will conduct a kind of scientific alchemy to help us turn old knowledge into new knowledge. And without even leaving Earth, we could explore space as never before.