• Todd Bonzalez@lemm.ee
    link
    fedilink
    arrow-up
    2
    ·
    2 months ago

    Did you actually read the article? The designers of this vision model used a software trick (inspired by the concept of quantum tunneling that has nothing to do with quantum computing) to allow inputs to bypass hidden layers at random, resulting in results that were able to see certain optical illusions in a way that other vision models cannot.

    This can be done by just adding some noise to the image. Sometimes it gets recognized as one thing and sometimes like another, just like humans would.

    Word soup by someone who knows way less than these researchers.

    • Umbrias@beehaw.org
      link
      fedilink
      arrow-up
      1
      ·
      2 months ago

      see certain optical illusions in a way that other vision models cannot.

      eh… but not in a way that is really like what humans see. which is the articles claim, but it makes a clasically cs approach to nuerology: zero effort to prove the quite substantial claim.

      Word soup

      that is most certainly not word soup. it’s also an accurate statement, though uncharitable to the authors claims.

      Also, the detail in description of their “quantum” inspiration (an effect not unique to quantum mechanics in fact, at that level of description) reads like they skimmed wikipedia’s intro to xyz topic, whether or not the author understands the topics more deeply.