Re:Astronomers explore uses for AI-generated images
by exit on Sunday February 05, 2017 @03:42PM
(#13574515)
maybe the complex distribution of "looks like a galaxy" images just isn't that overwhelmingly complex, relative to how cheap computation & storage is nowadays.
what surprises me is how these techniques can structure the space of compressed representations in interesting ways. look at this discussion of thought vectors: https://www.technologyreview.com/s/541356/king-man-woman-que...
However, it leaves me with the question why this process should converge to a desired network. I mean, why doesn't this process generate training artifacts, that get multiplied as the interaction between discriminator and generator progresses?