Slashdot Mirror

← Back to Stories (view on slashdot.org)

How 'Assassin's Creed' Or 'Fallout 4' Might Help Make AI Smarter (technologyreview.com)

Posted by timothy on from the learning-ethics-from-gta dept.

moon_unit2 writes: Apparently, playing computer games might provide a shortcut to greater intelligence. MIT Technology Review has a story about researchers using virtual game environments to train deep neural networks to recognize real-world objects. It's an important idea because deep learning usually requires huge quantities of annotated data, which isn't always available. So researchers from Xerox Europe, led by Adrien Gaidon, showed that training a deep learning system on a photo-realistic street scene could enable it to identify cars on real roads. "The nice thing about virtual worlds is you can create any kind of scenario," Gaidon says. Perhaps video games could play a bigger role in the future of AI than anyone realized.

1 of 45 comments (clear)

  1. Some examples by Anonymous Coward · · Score: 2, Interesting

    This idea has been around for a very long time in some form, though not taken full advantage of.

    For instance, PolyWorld was a very early version of having a neural net trained via a 2D world (it controlled little bots which would live or die depending on how well they survived). Here's the source code at github, and a Youtube video of the project being ran:
    https://github.com/polyworld/polyworld
    https://www.youtube.com/watch?v=RvcwuzeoQR0

    There are also a long series projects having neural nets learn to walk an arbitrary figure (different number of legs/body type, etc) around a 3D world with gravity applied. Here are some examples:
    https://www.youtube.com/watch?v=kQ2bqz3HPJE (has a bit of narration)
    https://www.youtube.com/watch?v=JFJkpVWTQVM
    https://www.youtube.com/watch?v=jev4UA7EVkc (this one has good comments, but the neural nets did not converge on a solution in that video)
    https://www.youtube.com/watch?v=fEM7YDNonSE
    https://www.youtube.com/watch?v=LCRPcz1B8rk
    https://www.youtube.com/watch?v=05Hycx1NpyE
    https://www.youtube.com/watch?v=6-N9WDMjCbE

    I've been playing around with WebGL recently, and I think that using that in the browser with three.js (or other high level javascript support API) can reduce the amount of code to generate a 3D scene to a minimum. Then sticking on an additional set up functions in the javascript for the neural net would create a great feedback loop in an absolute minimized environment. That's the direction I'm moving in currently, in any case.