Slashdot Mirror

← Back to Stories (view on slashdot.org)

AI Trained on Images from Cosmological Simulations Surprisingly Successful at Classifying Real Galaxies in Hubble Images (ucsc.edu)

Posted by msmash on from the interesting-usecase dept.

A machine learning method which has been widely used in face recognition and other image- and speech-recognition applications, has shown promise in helping astronomers analyze images of galaxies and understand how they form and evolve. From a report: In a new study, accepted for publication in Astrophysical Journal and available online [PDF], researchers used computer simulations of galaxy formation to train a deep learning algorithm, which then proved surprisingly good at analyzing images of galaxies from the Hubble Space Telescope. The researchers used output from the simulations to generate mock images of simulated galaxies as they would look in observations by the Hubble Space Telescope. The mock images were used to train the deep learning system to recognize three key phases of galaxy evolution previously identified in the simulations. The researchers then gave the system a large set of actual Hubble images to classify.

The results showed a remarkable level of consistency in the neural network's classifications of simulated and real galaxies. "We were not expecting it to be all that successful. I'm amazed at how powerful this is," said coauthor Joel Primack, professor emeritus of physics and a member of the Santa Cruz Institute for Particle Physics (SCIPP) at UC Santa Cruz. "We know the simulations have limitations, so we don't want to make too strong a claim. But we don't think this is just a lucky fluke."

20 comments

  1. What about Galaxy Zoo? by Anonymous Coward · · Score: 0

    I'm sick and tired of artificial intelligence taking over the work of natural intelligence!

    1. Re:What about Galaxy Zoo? by kcelery · · Score: 1

      Then you should make use of artificial intelligence to sort out what AI cannot excel and work on it.

  2. Nerd-collar jobs killer by Tablizer · · Score: 1

    Now what are astrophysics interns gonna put on their resume?

    --
    Table-ized A.I.
    1. Re:Nerd-collar jobs killer by Anonymous Coward · · Score: 1

      "Worked with AI to more accurately identify galaxy formations from Hubble space data."

    2. Re:Nerd-collar jobs killer by Tablizer · · Score: 1

      "...and made my three competitor colleagues obsolete."

      --
      Table-ized A.I.
    3. Re:Nerd-collar jobs killer by Anonymous Coward · · Score: 2, Insightful

      Probably something a heck of a lot more valuable than analyzing images and avoiding bias using near impossible criteria.

      It isn't surprising a machine could beat a human at a totally mechanistic and inhuman task.

      Now those same students can better use their time to analyze large datasets produced by these algorithms and come up with results the algorithms never could... until we find a way to create such an algorithm. Which is something those students can also work on.

    4. Re: Nerd-collar jobs killer by houghi · · Score: 1

      The next step will be for the computers to come up with algorithms. They already do that in other fields

      --
      Don't fight for your country, if your country does not fight for you.
  3. There goes galaxy zoo. by jd · · Score: 1

    Still, they run a lot of other projects and some of those are almost as much fun.

    --
    It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
    1. Re:There goes galaxy zoo. by Anonymous Coward · · Score: 0

      Back in the day I toyed with the idea of using galaxy zoo as a testing ground for ML (SVMs, which dates me, but anyhow). Never did it in the end as extracting features from images was a bigger hurdle then than it is now, but I don't see why this would signal the end rather than the beginning of a surge of renewed interest.

    2. Re:There goes galaxy zoo. by jd · · Score: 1

      Well, you only need to crowdsource the galaxies that can't be solved by AI. (If four out of five slightly different AIs all reach the same conclusion about a galaxy, then they're probably right. You'd need that many to cover all the cases this one can't cope with.)

      However, the pool of unsolvable galaxies will be much smaller and much noisier, otherwise the AI would have solved them.

      --
      It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
    3. Re:There goes galaxy zoo. by ShanghaiBill · · Score: 2

      Well, you only need to crowdsource the galaxies that can't be solved by AI.

      Or you can use Boosting.

      Boosting: Train two or three neural nets to solve the same problem. Then train another NN only on inputs where the others disagree, and use it as a tie-breaker.

  4. Seeing is believing... by Anonymous Coward · · Score: 0

    Check out this video, which compares a simulation of colliding galaxies with actual observations:

    https://www.youtube.com/watch?v=D-0GaBQ494E

    1. Re:Seeing is believing... by PPH · · Score: 1

      That looks a lot like my screen saver.

      --
      Have gnu, will travel.
  5. Tautology? by Anonymous Coward · · Score: 0

    So they showed images of galaxies to a neural network trained to find galaxies in images and that neural network discovered (wait for it) galaxies in those images?

    So I’m still in that “AI ain’t gonna take anyone’s jobs” camp then.

    1. Re:Tautology? by Herve5 · · Score: 0

      No, that's much worse than that -or else, a tautology in a quite different meaning :
      they trained an AI onto artificial images of what they THINK should be what they will see, and then the AI is confirming their bias -of course.
      To me this is the apogee of biased training, to date.
      (of course if it has been someone training the same NN onto artificially simulated jewish greediness, then showing that actual jews presented to it are found greedy, it'd have been an entirely different thing, isn't it?)

      --
      Herve S.
    2. Re:Tautology? by Herve5 · · Score: 1

      Wow. My first 'moderated flamebait' in years, for this... I should have replaced 'jews' with 'muslims', I'd have been moderated 'fashion' maybe ;-)
      But, back to serious : I really believe what I said : trained with artificial simulations introduces a bias.

      --
      Herve S.
  6. Good way to train AI by yes-but-no · · Score: 1

    This should be used even for the usual image recognition. You don't need 1000s of dog pictures/photos to detect a dog. A 3 year old child knows what a dog is - not by seeing 1000s but even one dog is enough. A generator AI using laws of physics/rotation/scaling etc should generate images for what a dog will look like using a base 3D model of a dog (like a dog toy you find in a toystore) [say from a given point of view/positioning of the 2D camera]. The final image recognizer will work even if the dog is made tiny, rotated, turning its head etc (in the model, you have to program which part can turn to what degree - like head say 80 degrees..knee say 120 degrees). The point is your abstraction of the object/content should go deeper - and look at the content/data generation process. Just seeing final bit pattern is going to be very hard as it's done today with photos - you lose those 3D movement/laws information.