AI can generate portaits indistinguishable from real human faces. what does this mean?

"

tsunamikun wrote:

By design this system only checks it against himself. The discriminator is not almighty and as with all programms that use a for of NNs behind it is actually vulnerable to bias. Especially if there is no target(you know, when you make a fake)

You always have a target when making a fake. Not sure what you mean.

"

tsunamikun wrote:

What i talk about is actual pixel computing, a far lower level. Also the poster above talked about faking others, not making realistic faces(which is the primary use of the software, and yes, it is a software no need for quoting marks). Making something new that people can't discrimitate as fake is completly different to take something exisiting and fake it.
Also it can only do portraits and it seems only with closed mouths. Still in its babyshoes :>

AI already made (not faked) voices and music which can easily trick people into believing it's real. The speed of development is also so fast that by calling something in its babyshoes is a bit naive. Five years ago we had basically nothing.

"

tsunamikun wrote:

Yeah, GANs is software, meaning an application. The mathematical algrothims and logical gating structure as well as the fundamentals (the actual meat behind a written software and all that shiny NN stuff)
exist since forever. We knew about it in the 40's. primitive structures were implemented in the 50's. What we use now has the mathematical foundations somewhere from the 80's.

That's like saying Einstein did nothing new because Newton laid out the ground work 300 years before him. And Newton sucked because Archimedes postulated some very early theories about the forces. It's a bad argument. GANs are entirely new stuff. Nothing even close to them existed just 5 years ago. There was some theoretical work about adversary networks in the 90s but that's it.

"

Xavderion wrote:

You always have a target when making a fake. Not sure what you mean.

No, you always NEED a target, and that's the problem. Especially in the use-case mentioned by the guy i quoted. So you define a target Which you can't reach perfectly without the software? So the bias and control system doesn't have good data the refer and check itself.

"

Xavderion wrote:

AI already made (not faked) voices and music which can easily trick people into believing it's real. The speed of development is also so fast that by calling something in its babyshoes is a bit naive. Five years ago we had basically nothing.

They were fake, as in not from a human, just to define what i mean with fake. But yes you are right, that exists. What exactly does it have to do with this system? It is a completely different technology with different use cases and is nowhere near performant or actually anything more than a tech-demo. (and a computer program could show easily that is was faked) And 5 years ago we had much. We already had the exact knowledge as well as the tools. But to build something like that takes time, to test it takes time, to improve etc...

And another thing: there is no AI. We have intelligent Algorithms, we have NNs but AI is a consumer term which doesn't have to do anything with actual artificial intelligence. Call it AI if the software can actually change it's own programming and doesn't need any targets or defined finish lines to compute something. Sorry about that rant, but i hate that term.

"

Xavderion wrote:

That's like saying Einstein did nothing new because Newton laid out the ground work 300 years before him. And Newton sucked because Archimedes postulated some very early theories about the forces. It's a bad argument. GANs are entirely new stuff. Nothing even close to them existed just 5 years ago. There was some theoretical work about adversary networks in the 90s but that's it.

That would be true if newton defined relativity. He didn't. I don't not talk about foundations but the actually used algorithms. and GANs are completely new? oh come on. It's just two neural networks competing. There is no new Technology used, but it is a new implementation of older technologies. That is fine, it's paper worthy. But let's not make it more then it is.

"

NemoJr wrote:

"

erdelyii wrote:

Be funny if the whales were singing ancestral songs about humans. Can't say we'd fare too well there.

Don't even joke about it. I hope we get a chance to explain ourselves before they initiate P.O.S. (Planet Obliteration Sequence), leaving the guy who came up with that acronym on the surface.

"Look, we were young and stupid back then..."

*whale screams in background*

"I TOLD YOU TO DO IT QUIETLY!.. Er, I mean... We have ice cream and pizza!"

- EARTH SAVED -

You can be the goodwill ambassador.

"

tsunamikun wrote:

No, you always NEED a target, and that's the problem. Especially in the use-case mentioned by the guy i quoted. So you define a target Which you can't reach perfectly without the software? So the bias and control system doesn't have good data the refer and check itself.

Yes, there's indeed a problem with poor sample diversity when it comes to targets for the generators. Some papers have been published on it and people are working on a solution. But I don't see the relevancy to the topic discussed in this thread, which is faking images or videos of existing people. There you (ideally) have plenty of target material which you explicitly want to use to make better fakes.

"

tsunamikun wrote:

They were fake, as in not from a human, just to define what i mean with fake. But yes you are right, that exists. What exactly does it have to do with this system? It is a completely different technology with different use cases and is nowhere near performant or actually anything more than a tech-demo. (and a computer program could show easily that is was faked) And 5 years ago we had much. We already had the exact knowledge as well as the tools. But to build something like that takes time, to test it takes time, to improve etc...

Google has already partly rolled out its virtual assistant thingy which can make calls and book appointments for you. It's beyond the tech demo phase. And it sounds real enough that people got angry when they found out they were talking to a bot, so Google had to include the bot saying that it's a bot before continuing the conversation.

"

tsunamikun wrote:

And another thing: there is no AI. We have intelligent Algorithms, we have NNs but AI is a consumer term which doesn't have to do anything with actual artificial intelligence. Call it AI if the software can actually change it's own programming and doesn't need any targets or defined finish lines to compute something. Sorry about that rant, but i hate that term.

I agree, AI has become a buzzword and has very little to do with what we consider to be (human) intelligence. We're still in the very early phases of actual AI, unsupervised learning has become a thing maybe 2 years ago and it needs a ridiculous amount of processing power so it's gonna be a while until we get some really good applications.

And before you mention it, I know that there are 30 year old papers on unsupervised learning, but the means to actually experiment with it are very new.

"AI can generate portaits indistinguishable from real human faces. what does this mean"

There is technology few years old where fake president is talking in TV. You could see no difference vs original face, mimicry and sound. Guy was sitting and talking - on screen was president face.

"

Rexeos wrote:

There is technology few years old where fake president is talking in TV. You could see no difference vs original face, mimicry and sound. Guy was sitting and talking - on screen was president face.

can you share a video of this?

"

cgheezey wrote:

"

Rexeos wrote:

There is technology few years old where fake president is talking in TV. You could see no difference vs original face, mimicry and sound. Guy was sitting and talking - on screen was president face.

can you share a video of this?

He probably means this video:

https://www.youtube.com/watch?v=dkoi7sZvWiU

It's not perfect by any means but it's pretty fucking good.

Combine this with Google's voice generator which could easily mimic Obama's voice pretty much perfectly and you got yourself a mess.

"

Xavderion wrote:

He probably means this video:

https://www.youtube.com/watch?v=dkoi7sZvWiU

yeah i find this pretty alarming

Hey, lets bring back Alec Guinness! I figure we could do some new bruce lee films with this.




Oh and we could render the adult industry obsolete.