GANs’ Generated Virtual Reality

Computer generated imaging is hard meticulous human labor, like translating English into another language, but that might become a thing of the past. As Google translate gets better and better with deep learning, CGI will become more automated, and indistinguishable from real footage with AI.

 

Generative adversarial networks, aka, GANs, is redefining how fast artificial intelligence is able to learn, and what it is able to learn. The concept is very similar to how the human learning process works, trial and error, but at a pace beyond the human capacity. The concept of GANs can be simplified as asking “is this it?” over and over, each time getting closer to “this is it”, but with the speed of computation.

GANs is a neural network with two main components, a ‘generator’ that keeps asking “is this it?”, and a ‘discriminator’ that answers “not quite yet” until, finally, “yes, that’s it”. Imagine a child mimicking a sound until it gets a word right and gets the desired response, though GANs is not exactly like a baby’s language acquisition, but the result is similar, the generator comes up with its own pronunciation that makes sense in general. In the case of how GANs is currently being  applied, the language is imaging. Of course, GANs is not limited to imagery, impersonating voices was just so last nanosecond for it. Listen to these vocal impersonations of Trump posted by CNBC, and you might start to question who you’re talking to through the phone.

Like language acquisition, there is a standard or pattern that is the basis for learning. Take the human face, for example, there is a pattern of what a human face looks like. The generator creates output that at first just seems like random blurs, like a baby’s goo goo gaga. Then, the blurs take shape, dada, and voila, an actual face, daddy! And, like unique human pronunciations, the generator creates its own image of a human face, but just not a face of any actual human.

 

An example of GANs at work that has recently made waves was the “Deep fake” video of Donald Trump criticizing Belgium. The deep fake video was a publicity stunt by a Belgian political party to attract attention towards Belgian commitment to our global climate challenge, and attention was what they got. According to reports by TheGuardian, comments like:

“Trump shouldn’t blow so high from the tower because the Americans are themselves as dumb.”

So many people genuinely got upset of Trump, and the US, that the people behind the publicity stunt had to start hinting that the video was a fake.

But, what’s more terrific and terrifying at the same time is not GANs ability to mimic a voice, or the lips and facial expressions, it is the combination of everything. No one has seen GANs being applied as the sole creator of, let’s say, an artificial murder scene, yet. So far, GANs has just been playing around with voices, making faces, and tricking out some really cool popping moves which should be turned into an app soon. Hitherto, seeing was believing.