Nvidia's A.I. Generates Perfect Headshots of Fake Celebrities

Nov 1 2017 -- 7:30 AM

Here's a fun little game for you to play if you're into celebrity trivia: Take a moment and try to identify the actor in the image below. You know you've probably seen him in something, right? Maybe he had a guest role on that one USA show, or did a stint as a Marvel TV villain. Feel free to take your time as you gaze into his gray-blue eyes.

If you ventured a guess as to who this celebrity is, we guarantee you that you're wrong. Why? Because this face was created by Nvidia's totally not evil sounding "generative adversarial network," or GAN. It's a type of dual neural network that was introduced in 2014 by Ian Goodfellow, a researcher who's worked at OpenAI and Google Brain. It is now, though, getting frighteningly good at inventing new, unique celebrity faces.

Nvidia's latest iteration of its GAN tech was highlighted in a recently published paper, and demonstrates unsupervised machine learning at its finest. The GAN works by giving two neural networks a single goal, which in this case is to create fake celebrity faces. This goal is achieved by having one neural network, the generative network, generate attempts at random faces, while having the second neural network, the discriminative network, decide whether or not the generated images are anything close to resembling real celebrity faces. Success is achieved when the generative network is able to fool the discriminative network into thinking that it has created (again, in this case) a real celebrity's face.

The discriminative network is trained to pick out real faces based on huge datasets, which in this case are tons and tons of real celebrity faces, while the generative network is seeded with a few faces, but then told to start generating more fake ones until it can produce some that the discriminator accepts as real. Below, Nvidia's GAN is shown doing exactly that. Over the course of about 20 days, the generator goes from producing nothing more than a few skin-color pixels to producing fake celebrity faces that could fool people.

It's easy to imagine this kind of technology being used for some seriously dystopian schemes, especially if it's somehow combined with this 3D face-making neural net to make full-on fake celebrity models. But at least in the case of Nvidia, it's likely to be used for gaming.

What do you think about this fake-celebrity-face-making GAN? Are you frightened of what's to come, or could this lead to some great gaming and entertainment? Give us your thoughts in the comments below!

Images: YouTube / Tero Karras Fl

We For One Welcome Our Robot Overlords...