HOW THREE FRENCH STUDENTS USED BORROWED CODE TO PUT THE FIRST AI PORTRAIT IN CHRISTIE’S

n Friday, October 25th, Christie’s will conduct a very unusual sale. As part of a three-day Prints & Multiples event, it’s auctioning offthe Portrait of Edmond Belamy, a canvas in a gold frame that shows the smudged figure of what looks like an 18th century gentleman. It’s expected to fetch a modest price, somewhere between $7,000 and $10,000, but the artwork’s distinguishing feature is that it was “created by an artificial intelligence,” says Christie’s. “And when it goes under the hammer, [it] will signal the arrival of AI art on the world auction stage.”

But for members of the burgeoning AI art community, there’s another attribute that sets the Portrait of Edmond Belamy apart: it’s a knock-off.

The print was created by Obvious, a trio of 25-year-old French students whose goal is to “explain and democratize” AI through art. Over the past year, they’ve made a series of portraits depicting members of the fictional Belamy family, amplifying their work through attention-grabbing press releases. But insiders say the code used to generate these prints is mostly the work of another artist and programmer: 19-year-old Robbie Barrat, a recent high school graduate who shared his algorithms online via an open-source license.

The members of Obvious don’t deny that they borrowed substantially from Barrat’s code, but until recently, they didn’t publicize that fact either. This has created unease for some members of the AI art community, which is open and collaborative and taking its first steps into mainstream attention. Seeing an AI portrait on sale at Christie’s is a milestone that elevates the entire community, but has this event been hijacked by outsiders?

WHOSE CODE IS IT ANYWAY?

To understand the concerns about attribution and expertise that have cast a shadow on the Belamy auction, you need to know a little about the tools used in the AI art world. The most important of these is the generative adversarial network (GAN), a type of algorithm first designed by Ian Goodfellow, a researcher now working at Google. The name “Belamy,” chosen by Obvious, is a tribute to him; it’s a translated pun from the French “bel ami” meaning “good friend.”

Goodfellow’s work on GAN is the stuff of AI legend. The story goes that over a few beers with friends one night in 2014, he posed the simple question: what if neural networks could compete with one another?

The basic idea of a GAN is that you train a network to look for patterns in a specific dataset (like pictures of kitchen or 18th century portraits) and get it to generate copies. Then, a second network called a discriminator judges its work, and if it can spot the difference between the originals and the new sample, it sends it back. The first network then tweaks its data and tries to sneak it past the discriminator again. It repeats this until the generator network is creating passable fakes. Think of it like a bouncer at a club: sending your drunk friend away until they act sober enough to get in.

ART MADE WITH GANS IS FULL OF MELTING SHAPES AND DISTORTED BOUNDARIES

This basic concept of dueling networks proved to be immensely powerful, and GANs are now a cornerstone of contemporary machine learning. They’ve been particularly fruitful in the AI art world, and pictures created using GANs have a distinct aesthetic that reflects how the algorithms process information. The networks know how to copy basic visual patterns, but they don’t have a clue how they fit together. The result is imagery in which boundaries are indistinct, figures melt into one another, and rules of anatomy go out the window. This aesthetic even has a tentative name, proposed by Google AI engineer François Chollet: GANism.

Robbie Barrat@DrBeef_

Exploring the latent space of AI-generated landscape paintings today instead of doing work.

Please do not tell my manager.

2:07 AM - Mar 20, 2018

• 1,653
• 579 people are talking about this

Twitter Ads info and privacy

Barrat has been a leading light in the world of GAN art, generating headlines with his surreal nudes and landscapes. He also shares the algorithms he uses to create these images on GitHub, helping fellow artist-coders to get the neural networks up and running. This is how Hugo Caselles-Dupré, the tech lead for Obvious and a machine learning PhD student in Paris, found Barrat’s algorithms and used them to generate the Belamy portraits.

Exactly how much of the work can be credited to Barrat or Obvious is a difficult question. Generating images using GANs is a multistep process. First, you collect training data with a “scraper” that your network will replicate. Then, you construct the generative algorithm, which is the most time-consuming and difficult part. Next, you start running the algorithm and sort through its output, picking the best examples from the hundreds or thousands generated.

Speaking to The Verge, Caselles-Dupré readily admits that Obvious borrowed elements from Barrat (like the “scraper” used to collect images), but he says they tweaked the code to produce their portraits to their own liking. “If you’re just talking about the code, then there is not a big percentage that has been modified,” says Caselles-Dupré. “But if you talk about working on the computer, making it work, there is a lot of effort there.” (Some of that effort seems to have involved bugging Barrat. In discussions on GitHub, you can see Caselles-Dupré, alias Caselles, asking Barrat to make changes to his code.)