Vanilla 1.1.9 is a product of Lussumo. More Information: Documentation, Community Support.

    Definitely a racist joke.
    • CommentTimeMay 28th 2019
    Not ageist?
    • CommentTimeMay 28th 2019
    It would be pretty easy around here.
    • CommentTimeMay 28th 2019
    Andrews Chess obsession leads him to see everything as Black or White.

    Chess is racist - rules say white goes first, white is given the advantage.

    We are all prejudice - we can't escape our evolved or cultural biases anymore than an AI (ML) can escape the biases of the datasets that are fed to it in training. Being aware of bias doesn't cure you of it.
    Ah but there's a reason for the "Zero" in Leela's name.
    • CommentTimeMay 28th 2019
    'How neural networks learn' - Part III: The learning dynamics behind generalization and overfitting

    No handwaving here :)
    • CommentAuthorAsterix
    • CommentTimeJul 4th 2019
    James lovelock still at it

    ...and very persuasive. I look forward to the book.
    And the book looks forward to you...
    That's a radicalising clip about Numenta. Jeff Hawkins would be tough to work with, I reckon. He might even be insane in the Silicon Valley sense (there are so many here). But the way he subsumes Hinton's capsule work and the way he seems to generalise current NN paradigms definitely gets my attention. He's definitely right about the unsung literature - I have linear metres here of NN journals from 1986-2001, each one packed with articles full of meaty mathematical analyses.

    During this same period I was architecting a scaleable 3D graphics chip whose crown jewel was a speedy multiply-accumulator - the essential component which makes GPUs today so attractive to NN researchers. That was back in 1987-89. Perhaps I blew the chance. But back then there was no developed internet and no publicly available big data, so perhaps not. Oh, and 33 MHz 386 PCs!
    Dammit, now I'm going to have to watch that clip. Maybe if I go out and mow the lawn the urge will pass.
    The AV production sucks totally.

    Sorry, Musk haters - it's him again.
    • CommentTimeJul 17th 2019
    You're incorrigible.
    • CommentAuthorBigOilRep
    • CommentTimeJul 17th 2019
    That's what love does.
    • CommentTimeAug 15th 2019 edited

    Adversarial triggering seems to me to be a bit of a problem. Given the apparent ease with which you can modify a few pixels in an image and fool a classifier into misclassifying with high certainty, how robust are these trained networks really? Would you trust one to drive you around when someone wearing the wrong dress might have you shunting into a wall?
    Top of the heap hardware (this week at least)

    0.5 POPS ain't bad, especially at <350 Watt
    • CommentTimeAug 28th 2019 edited
    But Huawei... who knows what is built into that AI? Or what might _not_ be built into it?

    "We decided it was wasteful to warm it up and cool it down all the time, so we did not built in a hard power switch. It is always on and warmed up, even when it appears to be sleeping. Actually it is refining its neural networks and processing new inputs and will have even better performance when it is awakened and put on task."

    If rocketry has a steep learning curve... I shudder to find out just how steep the AI learning curve might be.