Generative adversarial networks
Ian Goodfellow and colleagues in Yoshua Bengio's Montreal lab trained an image generator by pitting it against a second network that tried to spot its fakes, the method behind the realistic synthetic faces of 2017 and 2018.
What it was
"Generative Adversarial Nets" was posted to arXiv on June 10, 2014 by Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville and Yoshua Bengio, all listed at the Université de Montréal's Département d'informatique et de recherche opérationnelle. It was presented at NIPS in December 2014. MIT Technology Review reported in 2018 that Goodfellow came up with the idea during a night out at Les 3 Brasseurs, a Montreal bar, when friends asked for help with a program that could create photos by itself.
A GAN trains two networks against each other. A generator turns random noise into samples, such as images, and a discriminator estimates the probability that each sample came from the real training data rather than from the generator. The generator is trained "to maximize the probability of D making a mistake," which the paper framed as "a minimax two-player game." Both networks could be trained with ordinary backpropagation.
What it changed
Within three years GAN variants were producing faces that were hard to tell from photographs. Nvidia researchers' progressive GANs generated imaginary celebrities in 2017, and in February 2018 MIT Technology Review quoted Yann LeCun calling GANs "the coolest idea in deep learning in the last 20 years" and Andrew Ng calling them "a significant and fundamental advance."
The same capability became a security worry. The February 2018 report "The Malicious Use of Artificial Intelligence," whose 26 authors included Dario Amodei and Jack Clark, illustrated AI's recent progress with a row of GAN-generated faces from 2014 to 2017, beginning with Goodfellow and colleagues' paper.
The arguments it moved
What are the biggest risks from AI?
GANs gave the argument about AI misuse a concrete case in synthetic images of people who do not exist. "The Malicious Use of Artificial Intelligence" (February 20, 2018) wrote that "AI systems can now produce synthetic images that are nearly indistinguishable from photographs, whereas only a few years ago the images they produced were crude and obviously unrealistic." In MIT Technology Review's February 2018 profile, the Dartmouth forensics researcher Hany Farid said of detecting GAN fakes, "We're fundamentally in a weak position," and Goodfellow, then leading a Google team on machine-learning security, said of the security problem, "Clearly, we're already beyond the start." OpenAI's February 2019 GPT-2 announcement extended the worry to text, writing that its findings, "combined with earlier results on synthetic imagery, audio, and video, imply that technologies are reducing the cost of generating fake content and waging disinformation campaigns."
Sources
- Generative Adversarial Networks (arXiv, June 10, 2014)
- The GANfather: The man who's given machines the gift of imagination (Martin Giles, MIT Technology Review, February 21, 2018)
- The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation (arXiv, February 20, 2018)
- Better language models and their implications (OpenAI, February 14, 2019, archived)