Generative adversarial network (GAN) technology

[Nydex]

Greetings,

Since there's no "Technology" subforum, and this particular tech is based on machine learning, which is technically a science, I thought this subforum was the place to post such a topic.

I may be way behind on modern day tech, but today I stumbled across this technology. It's basically two artificial neural networks competing against each other. One of them creates images of subjects that look like real human beings and feeds them to the second (along with pictures of actual human beings), and the second one tries to recognize which of the images is of a real human, and which is artificially created by its opponent.

They both learn in the process and get better. You can read more about the technology here.

The official website presents you with some of the creations of this technology, and you can see for yourselves that many of those look almost 100% identical to real people (sometimes the pics take a long time to load for some reason, but you can click the link to them in the bottom right and it loads it immediately). It's honestly a bit creepy to look at what seems to be a perfectly normal human being and knowing it actually doesn't exist. Some of them are quite obviously artificial, but those will probably decrease in frequency as time goes by and the system improves.

Thinking about this stuff, and sourcing from my rich sci-fi imagination, I feel like there are so many possible applications to this technology that it's kind of scary. It can be used for scams too, among other things.

What do you think about it?

 

[downwardsfromzero]

After looking at several examples, I found the examples of nonexistant persons to be a little 'off' in some way. Assymmetries seemed forced and there were often visible glitches in one or other of the facial features. There were several strange backgrounds and some rather weird clothing. Often the teeth looked weird. The faces were somehow slightly odd to look at - perhaps they are still on the very threshold of the "uncanny valley".

None of the pictures would look out-of-place on, for example, an ID card, however.

I found the pictures seemed to induce an effect of switching between left-eye and right-eye dominance, something I first noticed happening quite a few years ago when I had begun experimenting with psilocybin mushrooms (minus the GAN generated faces, of course). How this might be significant, I am yet to fathom. It might be relevant to add that I've been creating dazzle pattern optical illusions today, so that could have done strange things to my visual cortex.

I'm wondering whether they'll ever get to the point of looking utterly naturalistic, or whether the GAN will always introduce some kind of tell-tale artefact, which IMO at the moment they still quite certainly do, from the limited number of faces I've looked at.