The first table created using artificial intelligence to be auctioned was sold in New York for US $ 432,500, 40 times more than its estimated value of US $ 10 thousand. The work, Portrait of Edmond de Belamy (2018) was sold by Christie's after more than six minutes of continuous bidding.
According to the Christie's catalog, "painting" is part of the group of portraits of the Belamy fiction family and created by an artificial intelligence apparatus trained by Obvious, a collective based in Paris. Its members, Hugo Caselles-Dupré, Pierre Fautrel and Gauthier Vernier, explore the area of art and artificial intelligence using a set of algorithms. According to Caselles-Dupré, the algorithm that was used for this work is composed of two parts. "On the one hand is the generator, and on the other the discriminator," he told Christie's. "We fed the system with information on 15 thousand portraits painted between the 14th century and the 20th century. The generator creates a new image based on that, and the discriminator tries to locate the differences between an image created by humans and an image created by the generator, "he explained.
Who is the creator?
This auction opens a philosophical debate about who should be accredited by this work of art: the creators of the algorithm, who implemented the base code, who modified that code or the Obvious company for using open source code to generate art? This moment indicates a fundamental change in thinking about creativity and art. Artificial intelligence does not show an exact creator.
Thousands of portraits
Hugo Caselles-Dupré, member of Obvious, told Christie's that the system was fed with 15 thousand portraits painted between the 14th and 20th centuries. Then, the program imitated the portraits and created several thousand paintings. Of all of them the company chose the best 11.
How does it work
Generative Adversarial Networks, or GAN, the technology used to create the portrait, has been used in art since 2015. The basic idea of a GAN is that a network is trained to look for patterns in a specific data set and is generated to generate do you copy. Then, a second network called a discriminator judges its work and if it can detect the difference between the originals and the new sample, it sends it back.
The first network adjusts its data and attempts to pass them back beyond the discriminator. Repeat this as many times as necessary. Networks know how to copy basic visual patterns, but have no idea how they fit together. The result is an image in which the boundaries are confusing. This aesthetic even has a tentative name: GANism.
The signature of the "creator" appears below on the right and it is the algebraic formula that was used for its creation: Min (G) max (D) Ex [log (D (x))] + Ez [log (1-D (G (z)))].