Visual AI-Spaces Auto-Pilot

Ich habe schon ein paar mal über Mario Klingemanns Arbeiten hier gebloggt, derzeit jagt er Neural Networks durch Feedback-Loops und erforscht so Face-Spaces und solche Dinge. Video oben: „Two pix2pix GANs, one that generates face markers from faces, the other that generates faces from face markers are set up in a feedback loop and generate an infinite stream of uncanny faces.“

Oder hier sucht ein NN Gesichter im Brides of Dracula-Trailer, während das nächste aus den Ergebnissen Gesichter generiert:

This is a little test that shows the limits of my trained face generation model. Left is the original trailer, on the right is the result of a combination of two pix2pix models, the first tries to find face markers, the second one tries to generate faces from the found markers. As you can see there is no threshold - the model will try to find and generate faces even if there are none in the source.

Bei diesen Arbeiten fallen eine Menge Crap-Ergebnisse in Form von Noise und Wiederholungen an und um die Kuration jetzt noch weiter zu vereinfachen hat er sich deshalb jetzt noch einen Auto-Piloten für Neues gebaut: „I've added a novelty-driven autopilot that is attracted to views it hasn't seen yet. 'To boldly go where no GAN has gone before.'“

Auf Wired erzählt Klingemann mehr über seine Arbeit als AI-Kurator: A ‘NEUROGRAPHER’ PUTS THE ART IN ARTIFICIAL INTELLIGENCE

He’s built himself a Tinder-style interface to quickly work through piles of newly generated “neurographs” and find the few that strike him as any good. “I produce a thousand images and maybe two or three are great, 50 are promising, and the rest are just ugly or repetitive,” he says.

As you may have noticed, the images he does select typically come with more than a little of the uncanny about them; Klingemann has lost count of the times he’s been told the faces and figures his code generates are reminiscent of Francis Bacon’s famously grotesque and disturbing work.

The comparison is apt. It’s also evidence of how far artificial neural networks are from really understanding images or art—not that computers have warped minds. “I’m doing creepy right now because I can’t do non-creepy, I wish I could,” Klingemann says. “In two or three years, the creepiness will go away, which might make it more creepy because we won’t be able to distinguish from a photo or painted artwork.”

Q: "Where do the little faces come from?"
A: "Well, when two not very well trained GANs meet and enter a feedback loop it just happens"

Generated by neural networks, chaining 3 GANs in a feedback loop.

Okay, this looks to me like an 80's glam metal boy group meets in a wind tunnel.