New Study Shows AI-Generated Faces Are Becoming Inseparable From Reality

Monday, 17 December 2018 - 12:39PM
Technology
Artificial Intelligence
Dystopias
Monday, 17 December 2018 - 12:39PM
New Study Shows AI-Generated Faces Are Becoming Inseparable From Reality
< >
Screenshot: Tero Karras FI/YouTube
We previously reported on Google's DeepMind AI being able to generate new, photorealistic images of scenes that never existed, as well as the niche world of "deepfakes," an AI technique currently being used to insert Nicolas Cage into various Hollywood blockbusters, though Congress has recognized its (slightly more concerning) potential to undermine world peace. Now, a group of researchers at Nvidia has published a new study that shows just how close neural nets are coming to simulating reality.

Before we go into detail about the study and how it works, check out this video published by the researchers demonstrating the neural net in action. It blew our minds a little bit.



The technology behind this kind of AI is called a GAN, or "generative adversarial network." A GAN takes a different approach to learning that other types of neural networks: a GAN uses two separate elements, called a Generator and a Discriminator, which "compete" against one another to create the desired result. The Generator's job is to create realistic-looking fake images, while the Discriminator's job is to distinguish between real images and fake images. If both are functioning at high levels, the result is images that are seemingly identical real-life photos.


GIF from GIFS.com via YouTube/Tero Karras FI

As the video explains, the neural network also classifies major features of different faces, like hair or skin color, as "styles," which can be applied to other faces to create a new image. In the words of the study: "The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis."

To sum it up, this neural network is able to create realistic images of faces in addition to providing its users with a sliding scale that allows them to tailor those faces to the right specifications. Cool? Yes. Deeply troubling? Also yes.
Science
Science News
Technology
Artificial Intelligence
Dystopias
No