Space

Neural networks focus in on old data to find new discoveries

View 2 Images
A neural network AI has shown itself to be better than the current best techniques at cleaning up distorted optical images from space
From left: The original image; the same image degraded; the GAN's attempt at cleaning it up; and the restoration using deconvolution
K. Schawinski/C. Zhang/ETH Zurich
A neural network AI has shown itself to be better than the current best techniques at cleaning up distorted optical images from space

Remastering of movies and video games allows us to see old classics with more detail and definition than the original. Taking a page out of that book, astronomers have developed a new way to "remaster" old optical images of space, using neural networks to potentially bring new discoveries into focus.

As forms of artificial intelligence designed to function like the human brain, neural networks can be fed a type of data, and trained to pick out patterns and apply them to new examples of that data. In this case, astronomers at ETH Zurich taught the system how to clean up blurry images by showing it both clear and fuzzy pictures of the same galaxy.

The study used a generative adversarial network (GAN), a relatively new neural network structure that pits two systems against each other, improving the effectiveness of both in the process. According to the researchers, training the GAN only took a matter of hours.

Then they set it loose on their data. By feeding it artificially-degraded images and tasking it with restoring them, the team found it could pick out features like star-forming regions, bars and dust lanes that would normally be lost. The GAN-improved images could then be compared with the sharp originals, and the scientists found the AI was more accurate than other clean-up systems like deconvolution, a technique often used to reduce distortion from older images.

From left: The original image; the same image degraded; the GAN's attempt at cleaning it up; and the restoration using deconvolution
K. Schawinski/C. Zhang/ETH Zurich

"We can start by going back to sky surveys made with telescopes over many years, see more detail than ever before, and for example learn more about the structure of galaxies," says Kevin Schawinski, lead researcher on the study. "There is no reason why we can't then apply this technique to the deepest images from Hubble, and the coming James Webb Space Telescope, to learn more about the earliest structures in the Universe."

That opens up the possibility that new discoveries could be found in old data, and the team says the technique could lead to a future where information about astronomical objects is automatically gathered and analyzed. Patterns could be spotted more easily, saving human astronomers the task of manually putting together physics models.

The benefits of the system are not just limited to what it can teach us about space, either. The more data that a neural network is trained on, the better it gets at its job, so feeding them a data diet this healthy can turn these AI systems into more interesting test subjects for computer scientists.

"The massive amount of astronomical data is always fascinating to computer scientists," says Ce Zhang, co-author of the study. "But, when techniques such as machine learning emerge, astrophysics also provides a great test bed for tackling a fundamental computational question – how do we integrate and take advantage of the knowledge that humans have accumulated over thousands of years, using a machine learning system? We hope our collaboration with Kevin can also shed light on this question."

The research paper was published in The Monthly Notices of the Royal Astronomical Society.

Source: Royal Astronomical Society

  • Facebook
  • Twitter
  • Flipboard
  • LinkedIn
0 comments
There are no comments. Be the first!