• howrar@lemmy.ca
    link
    fedilink
    English
    arrow-up
    4
    ·
    7 days ago

    The long-forgotten technique? Normalizing flows. It feels like it was just last year that they were all the rage. It’s insane how fast the field is moving.

    • howrar@lemmy.ca
      link
      fedilink
      English
      arrow-up
      2
      ·
      7 days ago

      It looks like Laurent Dinh (dude who originally came up with normalizing flows) is one of the authors of this work.

  • Pennomi@lemmy.world
    link
    fedilink
    English
    arrow-up
    1
    ·
    7 days ago

    There doesn’t seem to be much going on here. What’s the actual advantage of this technique? It seems they’re doing things more or less the same as everyone else.

    I mean, the article talks about decoders as if they were some brand new innovation.

    • howrar@lemmy.ca
      link
      fedilink
      English
      arrow-up
      2
      ·
      edit-2
      7 days ago

      Generation ability seems to be about the same as any other model. The advantage of normalizing flows is that operations are invertible. That allows you to not just generate samples, but also calculate the probability of a sample.