Generative Adversarial Networks (GANs)

Creating 3D Models from Flat 2D Images
March 10, 2020
Ambient Computing Expands
March 10, 2020

Generative Adversarial Networks (GANs)

GANs are unsupervised deep learning systems comprised of two competing neural networks trained on the same data.

Facebook AI research director Yann LeCun has called generative adversarial networks (GANs) the most interesting idea of the decade—and in the past few years, there have been tremendous advancements in GANs.

Think of a GAN as the Turing test but without any humans involved. GANs are unsupervised deep learning systems comprised of two competing neural networks trained on the same data—such as images of people.

For example, the first A.I. creates photos of a woman that seem realistic, while the second A.I. compares the generated photos with photos of real women. Based on the judgment of the second A.I., the first one goes back and makes tweaks to its process. This happens again and again, until the first A.I. is automatically generating all kinds of images of a woman that look entirely realistic.

Last year alone saw a number of interesting experiments involving GANs. Researchers from Nvidia, the MGH and BWH Center for Clinical Data Science and the Mayo Clinic collaborated on a GAN that generates synthetic MRIs showing cancerous tumors.

Researchers at the Skolkovo Institute of Science and Technology and the Samsung AI Center made living deepfakes—the Mona Lisa moved her head and Rasputin sang “Halo” by Beyonce. (Naturally, it’s this same GAN technology that’s behind deepfakes in general.)