Why on GitHub there are no friends. About Robbie Barrat, Obvious and Copyright

    “Computers see such people,” Robbie Barrat told Twitter, showing surrealistic pictures that the neural network he created created.

    Honestly, these works are not a pleasant sight. The characters in the paintings are more like creatures from Silent Hill than people - most have no heads, and if there are, then for some reason they look like strange purple textures.



    And the developer himself does not hide the fact that the result was rather frightening. In an interview for CNet, Robbie calls the work of a neural network "surrealistic drops of flesh with limbs."

    And if earlier neural network artists were discussed in more or less narrow circles, then on October 25, 2018 this topic created a resonance. All due to the fact that a group of French students called Obvious sold at auction a picture created by a neural network, the code for which was written by Robbie. So three students with someone else's code and a bright PR campaign became key figures in the discussions of AI-art. Is it deserved? Let's see.

    Under the cut - the story of Robbie Barrat, copyright and successful marketing, which helped to gain almost half a million dollars.

    Who is Robbie Barrat




    Now Robbie Barrat is 19 years old, he works and teaches at the Stanford Bioinformatics Laboratory.

    Robbie spent his childhood on a parent farm in West Virginia - away from new technology and computer development. In an interview for The Stanford Daily, the guy recalled that computer science at their school was taught by a physical education teacher, and computers needed to be updated. However, this did not prevent Robbie from studying neural networks.

    Neural networks for Barrata were a kind of hobby to which he devoted a lot of time after school. For the past few months, he has lectured on artificial intelligence and neuroscience, and the picture that his neural network created appeared on the cover of Bloomberg.


    Bloomberg cover with AI painted landscape. Source: 1fuzz.com

    “I arrived in California almost immediately after graduating from high school to work at NVIDIA. They noticed on GitHub a project in which I taught the neural network to write rap, and offered me an internship. I recently left NVIDIA and came to work at the laboratory. ”

    AI benchmark and Kanye West tracks


    Robbie Barrat created his first AI at age 17 to win an argument. Then the participants in the school programming club, where Robbie was the chairman, argued about whether artificial intelligence could perform tasks better than people. It seemed to Barrat that the answer was obvious, but several peers defended the theory of “human superiority” to the last.

    “Suddenly, I had an idea to create a neural network that could read rap music,” Barrat recalled. A week later, when the club met again, he presented the song already written by the AI.

    AI rep was the first serious project of Robbie Barrata. The developer claims that he wrote most of the source code in the evening, and spent the next few days optimizing the results. The last iteration learned from 6,000 Kanye West tracks and could generate speech and pauses, which almost always sounded in the subject.

    “Initially, the AI ​​simply rearranged the phrases from the songs, but now he can write his own words,” said Barrat.



    Link to the open source code of the project on GitHub - here .

    AI couturier and show from Balenciaga


    The songs were just a start. In August 2018, Robbie posted on Twitter his new ideas - an AI-created clothing collection based on photos of models from the Balenciaga fashion show.

    Robbie Barrath collected images of Balenciaga clothing models from lookbooks, advertising campaigns, shows and online catalogs for the last two months, and then used them to train the pix2pix neural network.

    At the exit, the images turned out to be nothing worse than the originals of Balenciaga. “Since the neural network is not limited to human taste, style and history, it comes up with things that would never occur to humans. I like the way the network ignores symmetry, ”Barrat wrote on Twitter.

    For example, the neural network offered pants in which the pocket will be located on the shin. Robbie liked this idea so much that he decided to make it for himself and other fashion lovers.



    Even if the ideas of AI do not always look realistic, Barrat believes that designers can use such a tool to search for inspiration.

    Fat folds and navels: what AI sees naked people


    Even before the project with Balenciaga, Robbie created a neural network that can generate pictures. At first it was landscapes and portraits of people. They were followed by a collection called Nudes, generated on samples of paintings with naked people.

    Neural networks are able to create pictures, invent clothes, and write songs, largely thanks to Ian Goodfellow. Goodfellow's GAN paper, published in 2014, is a legend of AI. They say the idea to create a generative-adversary network came to Jan after several glasses of beer while he was sitting with friends in a pub. Then Goodfellow wondered: what if neural networks can compete with each other?

    Paintings created using the GAN, have a special aesthetics. With their help, we can see how the algorithms process information.

    Networks know how to copy basic visual templates, but have no idea how to combine them with each other. This is also reflected in the paintings of Barrath’s neural network - at first glance it is depicted as something similar to human bodies, but their boundaries are fuzzy, the figures melt into each other, and the rules of anatomy are absent in principle.



    “Creating good high resolution pictures takes about two weeks. The discriminator looked through 30,000 paintings over and over again. He probably scanned the entire data set at least 10,000 times. So, it took several million samples to train the neural network, ”explained Barrat.

    “At the beginning of training, neural networks did not work well. The discriminator did not distinguish the pictures, and the generator simply created noise. But while the generator received feedback from the discriminator, the latter also practiced: he had access to both the real paintings and the fake ones created by the generator.

    The Nudes collection of paintings turned out to be terrible, but in a good way. The generator simply tried to get a good score from the discriminator. He found out that the discriminator pays attention only to low-level functions, such as fat folds and navels. When the generator filed a picture with such components to the discriminator, the testing network counted it as real. The generator fooled the discriminator by feeding him with these terrible drops of flesh with folds. The discriminator said: “yes, this is a portrait of a naked body,” and at that moment the training stopped. The generator no longer needed to improve.

    When the pictures did not work, I was upset. But we must admit that the results are quite entertaining. If the discriminator missed only beautiful images, they would have attracted attention for ten minutes. We have already seen a lot of naked portraits - people painted most of their history. The neural network creates a new style of painting, although it is unusual. "

    Link to the open source code of the project on GitHub - here .

    So whose code?


    The pictures drawn by AI are not a new direction, but they have only recently been announced about art. On October 25, the portrait “Edmond de Belami, from the family of the de Belamies”, created by GAN, was sold at Christie's auction for $ 432,500. Auction House representatives said that this amount is almost 45 times higher than the one in which they rated the work previously.

    The authors of the work were a group of three French students called Obvious. As they noted after the auction, the painting was dedicated to Jan Goodfellow: from French “bel ami” means “good friend”. By the way, the author of the original code, Robbie Barrata, was also mentioned in a thank-you speech, but very indirectly.

    “We would like to thank the AI ​​community, especially those who first used the technology. This is Ian Goodfellow, the creator of the GAN algorithm, which inspired us to the name of the series “Famille de Belamy” and Robbie Barrat, who had a great influence on us. This is an exciting moment, and we hope that this sale will draw attention to the work that our predecessors and colleagues are doing, ”said Hugo Cassel-Dupre, technical leader of Obvious.

    But if the work of Obvious intrigued the art lovers, then it rather confused the AI ​​developers.

    In an interview for The Verge, Robbie spokethat Obvious addressed him with a request to use the "component" of his code. But after agreeing, they began to produce works identical to those laid out on his GitHub. From this point on, AI developers began to argue who owns the code: Barrat or Obvious?

    The American portal The Verge devoted a whole investigation to this issue. He was joined by other AI artists.

    Mario Klingemann, an AI artist from Germany, was shocked that the auction sold the painting for hundreds of thousands of dollars. In an interview for The Washington Post, he said that for some time he considered this a "joke among oligarchs."

    “For me, this is amateurish work, equivalent to children's projects that only parents can evaluate,” said Klingemann. “But I think that for people who have never seen anything like this before, this may seem interesting.”

    In a letter to The Verge, the artist wrote: "Probably 90% of the actual work was done by Barrat." Tom White, an academic and AI artist from New Zealand, also told The Verge that the work is very similar.

    “I knew that people would use my code as components of their projects. But I never thought that someone would sell it - if only because my work is not yet so good. I wrote this code in my free time when I was still in school, ”Barrat said in a comment for The Washington Post.

    Why, then, does Obvious earn money in pictures, but Robbie Barratt does not?


    By the way, Obvious never denied that their work was based on someone else's code. Hugo Cassel-Dupre confirmed that the elements were borrowed from Barrat: "If we talk about the code, it did not suffer big changes, but we spent a lot of energy on working with the neural network itself."

    From all this, the question arises: how could a semi-tenant work, semi-borrowed work, ever fall into the hands of Christie's? After all, this is not just an auction house - it is the leader of the global art market, whose total revenue only at the end of 2015 was £ 4.8 ($ 7.4 billion).

    Probably one of the reasons for Obvious's success was their ability to file themselves.

    The press materials that the trio sent in early 2018 sounded like the following slogans: “Creativity is not only for people”, and also “Artificial intelligence has managed to create art”. Such news from the media was snapped up.


    The first page of the press release that Obvious sent out to the media. Source: The Verge.

    Although beautiful phrases and not quite true, Christie's were ready to buy it. And, as it turned out, not lost.

    Most likely, Barrath, if he wishes, will still not prove anything - but he doesn’t seem to want much. Is it that posts in Twitter stinging posts:



    PS They will talk about the application of artificial intelligence at our two conferences.

    First - at the AI Conference Kyiv , which will be held on November 14.
    Then at the AI Conference in Moscow , which will be held on November 22.

    Also popular now: