back to article Boffins turn to AI to zip through piles of gravitational lenses

A group of physicists has trained an artificial neural network to analyze gravitational lensing images ten million times faster than normal computational methods. Gravitational lensing is "the formation of multiple images of distant sources due to the deflection of their light by the gravity of intervening structures," …

  1. TDog

    Trained to see what we expect

    You’ve got to fake it to make it

    Real data is sparse, so the researchers simulated half a million gravitational lens images to train the CNN. The fake images have to be as realistic as possible for the CNN to be useful in dealing with real data, so blurring and noise effects were added.

    Wow - once we've taught it to find what theory predicts - guess what it finds?

    This is spot the sausage in a Wurst factory; where we have trained it to find the spotted dick. Even worse (wurst?) this will then lead to a 'done that; tick the box' mentality which will preclude the other, unfound information, since our tests have been defined by our expected results. A tad of a perversion of Rev (if he was) Beyes' there.

    1. Mephistro

      Re: Trained to see what we expect

      My guess is that before creating the test data, they extrapolated the real data they had from known gravitational lenses.

      The worst thing that could happen here is that the software, sometimes, wouldn't be able to identify a GL of a formerly unknown type.

      1. TDog

        Re: Trained to see what we expect

        The most exciting phrase to hear in science, the one that heralds new discoveries, is not “Eureka!” (I found it!) but “That’s funny …”

        — Isaac Asimov

        And that is what is missed with an AI trained on expected rather than real data - even with real data there is a serious problem.

      2. Michael H.F. Wilkinson Silver badge

        Re: Trained to see what we expect

        Fair points about seeing what it expects, but then essentially all pattern recognition methods will need training data. CNNs need far more than most, which is why the authors need to resort to simulated data. However, suppose we have a method that needs fewer training data to reach the same performance, we would still have a method that realistically only finds objects similar to those examples fed to it. One way out of this problem is to devise methods that will flag strange outliers in the distribution of objects as "weird object, please let experts have a closer look". The data volumes are so vast that we will need methods to separate the masses of ordinary objects (for a given value of ordinary), and peculiar but well understood types, from the really weird objects that do not obviously fall into a known category. There are various efforts under way, including the EU SUNDIAL project. Exciting times.

    2. Francis Boyle Silver badge

      This is just how science works

      We know how gravitational lensing works since GR is a well confirmed theory. We can use that theoretical knowledge to build instruments to give us data about the world we can't acquire directly. In principle it's no different to the way geologist use sophisticated models of the Earth to find mineral deposits except that astronomers have been doing it a lot longer. Look up Cepheid variables for a simple example of how this works in principle. The news here is just the use of ML.

  2. F111F

    Have I Got This Right?

    <quote>Yashar Hezaveh, co-author of the paper and researcher at Stanford, told The Register: “It’s hard to say” what features the CNN learned to extract to arrive at its output answers.

    “In fact we don’t really know. As we show the examples to neural networks and ask them to make the correct predictions, they may find very complex features in the data that they can use for their predictions. We can sometimes look at the features, but they will be highly non-intuitive.

    “I usually think about it like opening the brain of someone and looking inside it: it doesn’t tell us much about what the person is actually thinking about or how they see the world.”</quote>

    They're using results of a computer analysis but don't understand how the "neural network" got the results? How do you replicate and/or validate the conclusions/results independently?

    1. defiler

      Re: Have I Got This Right?

      Yep. That's pretty-much neural nets.

      There are no hard and fast rules. Just a bunch of weightings. Imagine, if you will, a flawed analogy:

      You have a machine with a video feed going in at one end, and an 8x8 grid of knobs to turn. None of the have any labels, and there is no map of how each one is wired together. At the other end is a screen that gives you stats on what is detected in the image.

      That's pretty much it.

      By testing the machine's output and saying "colder" or "warmer", you instruct a marvellous mechanical golem to tweak the knobs more-or-less randomly until it converges on some settings that *appear* to give the correct result on the training data.

      Then the real challenge is to repeat that success with fresh test data, so that the net detects tanks instead of cloudy days...

  3. Richocet

    This won't invalidate the findings

    The scientists will use this AI to speed up a process.

    Surely they will review the examples found by AI which will allow them to learn from and get insights from them, and they will eliminate any false positives.

    I see no reason to think that if the AI fails to find examples in sections of the data, the astronomers will conclude that there cannot be any in that data.

    When AI finds examples and those are reviewed, the refined data would make a better training data set which can be used to improve the effectiveness of the AI.

    They will be better off than searching for these manually.

    So I don't see any major accuracy flaw in the AI approach.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like