Just take a bloody picture with a camera you bunch of lazy buggers!
I like BigGANs but their pics do lie, you other AIs can't deny
Images generated by AI have always been pretty easy to spot since they are always slightly odd to the human eye, but it’s getting harder to differentiate what’s real and fake. Researchers from DeepMind and Heriot-Watt University in the UK have managed to significantly boost the quality of images simulated by a generative …
COMMENTS
-
Tuesday 2nd October 2018 00:53 GMT ThatOne
What are we actually talking about?
> Images generated by AI
Images generated how? Are they painted out of nothing ("paint me a brown dog"), are they modifications of existing images?
What exactly are we talking about? Can somebody please be so kind to explain what I'm looking at?
If they are "painted" out of nothing following a request like "paint a butterfly on some flowers" they are astonishingly good and definitely plausible: The dog could have an eye problem, and the butterfly an old scar (or there was some dirt on the lens).
As for the "dogball", I definitely don't see what kind of attempt gave birth to that creature: It's not a dog face on a tennis ball, it's not a tennis ball made to look like a dog, I have the nagging feeling I'm seeing the answer to a question I don't understand.
-
Tuesday 2nd October 2018 08:38 GMT Anonymous Coward
Re: What are we actually talking about?
Basically: you have two “AIs”. One generates a random image and the other one says “doesn’t look like anything to me”. The first one tweaks it and tries again. You repeat this process billions of times until the second “AI” says “yes, I think that’s a thing”. That is the “adversarial” A in GAN.
-
Tuesday 2nd October 2018 12:24 GMT Anonymous Coward
Re: What are we actually talking about?
> One generates a random image
In what way? Is it just a completely random collection of pixels, like snow on an old TV
It seems implausible that you'd get anything out of that, especially not with only "yes/no" for the entire picture as feedback.
A friend once tried creating fonts by generating random patterns of 8x8 pixels and stopping when he found something that looked like a character. It wasn't very successful.
-
Tuesday 2nd October 2018 19:01 GMT Michael Wojcik
Re: What are we actually talking about?
Is it just a completely random collection of pixels, like snow on an old TV
In a classic GAN, that's how it starts. Per the original GAN paper, you set up two untrained feedforward perceptron networks, G and D (generator and discriminator). D is trained using a corpus of the sorts of images you want to mock, but G is trained only on the output of D.
Obviously the "images" generated by G will simply be noise for many iterations. Gradually it converges on outputs that D is unable to distinguish from the training set.
There are various complexities and subtleties involved (see the paper referenced above or any of the many subsequent ones on the subject; Colyer has discussed some in the Morning Paper), but that's the basic idea.
It's really not hard to get an ANN to go from "output indistinguishable from random" to "output closely follows the desired distribution". People who are surprised by this just haven't been following the field. The key here is a lot of computational brute force (the GAN minmax game runs for a lot of iterations), some fairly complicated statistical massaging because things like the likelihood estimation are often not directly tractable, and care to avoid common statistical problems like overfitting. Just skimming the paper I cited conveys the flavor of the thing, I think.
-
-
-
Tuesday 2nd October 2018 14:00 GMT juice
Re: What are we actually talking about?
#metoo.
If we have an AI which is capable of generating a photo-realistic view of any scene we care to imagine, then I for one welcome our new and scarily capable overlords.
If on the other hand, the AI is just tweaking some existing photo and submitting that, then it's not quite yet time to call time on the human race.
-
Tuesday 2nd October 2018 19:12 GMT Michael Wojcik
Re: What are we actually talking about?
If we have an AI which is capable of generating a photo-realistic view of any scene we care to imagine, then I for one welcome our new and scarily capable overlords.
If on the other hand, the AI is just tweaking some existing photo and submitting that, then it's not quite yet time to call time on the human race.
You need another hand. It's neither.
The generator of a GAN does not "tweak[] some existing photo". It never sees any input from the corpus of existing images. All it sees is the output of the discriminator.
That said, a GAN can't "generat[e] a photo-realistic view of any scene we care to imagine", either. For one thing, it has no way of knowing what we imagine. More importantly, the discriminator part of a GAN has to have some corpus to be trained on.
In general you need an objective function to maximize. Going from a large collection of images that share some characteristic to an objective function is a fairly well understood problem with some decent approximate solutions. Going from, say, a text description to an objective function corresponding to the concept of an image representing that description ... well, that would be harder.
Deep Fakes, for example, work (to the extent they do) because the forgery is composed of one component that doesn't need to be forged at all (the majority of each frame of the generated video) and another component that can be directly modeled (because we have many images of faces, including some of the target face).
That's not to say that there aren't systems that can synthesize images from textual descriptions, and vice versa. But it's a different and far more complex problem, particularly if the text description is complicated.
We'll have it eventually, though. ML has been fooling judges in areas such as music composition for decades, and it's only going to get better. Those improvements do nothing to address the philosophical questions around machine intelligence and creativity, but in practice our ability to discriminate between real and synthetic images will continue to diminish.
-
-
-
-
-
Wednesday 3rd October 2018 14:04 GMT Eddy Ito
Re: Fake reality created by fake intelligence
I believe Skynet hacked into NORAD's WOPR to play a game and ultimately decided it would be easier to use VR headsets to herd people into the matrix rather than going all terminator on the masses.
Make no mistake the machines are rising; they're just being very crafty about it.
-
-
-
Tuesday 2nd October 2018 13:03 GMT Ragarath
Re: Is it just regurgitating the training dataset?
That was my thought. Is it pulling data as in cut and past from different images it has seen or is it generating the image pixel by pixel.
Perhaps it is painting it like you would in an art application. Who knows. I shall endeavor to find out.
-
Tuesday 2nd October 2018 19:13 GMT Michael Wojcik
Re: Is it just regurgitating the training dataset?
Can it generate original images, or just tweaks of the training dataset?
That's not a meaningful question. What's the discriminating attribute between the two? If a generated image doesn't appear in the training corpus, is it a "tweak" or "original"?
-
-
Tuesday 2nd October 2018 15:14 GMT Chris Evans
Only a matter or time.
"... he is also worried about how GANs can be used maliciously. “It's part of why I chose to focus on more general image modelling rather than faces - it's a lot harder to use images of Dogball for political or unethical purposes than it is to use an image of another person.”
It won't take long for the unethical to start using this sort of technique and whilst experts may be able to detect unreal[1] images it will become harder and harder to do so as the technique improves.
[1] I now try and avoid using the word 'fake' as its use by certain people means it is to me a discredited term.
-
Tuesday 2nd October 2018 19:17 GMT Michael Wojcik
Re: Only a matter or time.
Start? Forged images have been around for a long time, and I don't see why those making practical use of them would have restrained themselves from employing GANs and other ML techniques once they became readily available.
And that was a while ago. Take this handy article on Medium - not exactly a technical journal - which explains GANs in terms of an episode of Spongebob Squarepants and a handy Tensorflow implementation. It was published two years ago.
You can run this stuff on a smartphone.
-
-
Wednesday 3rd October 2018 13:33 GMT Michael Wojcik
Another example
Here's a Hackaday post from yesterday about Helena Sarin's use of GANs to create illustrations.
Sarin's original piece has many more examples, as well as another high-level description of how a GAN works and how she uses it to generate images.