The Artist and Their Algorithm: Artificial Intelligence in Service of Art
At the beginning of November, Christie's auctioned off a piece generated by an artificial intelligence algorithm. Sold for over 480,000 CHF, the Portrait of Edmond Bellamy sparked debates across the art world and the research community. Some applaud the audacity of the creators, proclaiming the dawn of a new era in the art market. Others see it as a marketing stunt, inflated by technical approximations and superficial debates about the nature of art, its production, and its value.
The Portrait of Edmond Bellamy, 2018, created by GAN
The Artistic Formula
To generate this portrait, the creators, a Parisian collective named Obvious, used an algorithm called Generative Adversarial Network (GAN). The fictitious subject's name, Edmond Bellamy, is nothing more than a literal translation of the name of the inventor of this generation method, Ian Goodfellow. Inspired by game theory, the model "stages" a competition between two players. The first, called the "Generator," aims to create a fake—imagine, for example, a counterfeit banknote. It is constrained by the fact that it can never see examples of a real one. With no idea what the banknote should look like, it attempts a random combination of shapes and colors and submits it to a judge, the second player, called the "Discriminator." The judge responds to the forger's attempts by either rejecting the fake or accepting it, considering it sufficiently close to a real one.
A GAN learning to reproduce numbers
This operation is repeated many times until a balance is reached between the two players. In the best case, the generator will have succeeded in replicating a real banknote and manages to fool the second player. The generator can then create an infinite number of new examples based on its understanding of what the discriminator expects. If this reconstruction is perfect, it will find its place in many applications. We have already discussed the crucial importance of data in developing machine learning algorithms. The ability to generate, on demand, an infinite number of examples to improve an algorithm is a very enticing prospect.
Advances in GANs since their invention in 2014
Unfortunately, the generator rarely achieves a perfect reconstruction. One of the greatest challenges of current models lies in their ability to generate examples originating from different classes. In the example of counterfeit banknotes, it is relatively easy, after many tries, to generate a given type of banknote (for example, a 50 CHF note). However, being able to generate any banknote requires much more effort in abstraction. Most of the time, the model will not perfectly separate the specific characteristics of a 50-franc note from those of a 20-franc note. As a result, it will often generate a mix of the two, a sort of interpolation between two known examples.
Interpolation between a dog and a jellyfish by BigGAN
Instead of replicating banknotes, the creators programmed the generator to produce portraits, inspired by 15,000 different paintings from the 15th to the 18th century. Unable to separate the thousands of subtleties between these examples, the generator ends up producing a mix of its influences. The result easily reveals this imprint, with typical colors and near brushstrokes, whose digital origin seems visible only through blurred areas, digital artifacts of the algorithm.
Who Is the Artist?
Like many works before it, this piece sold at Christie's is signed. Yet, upon closer inspection, the signature is somewhat unusual. It expresses in mathematical terms the relentless battle between generator and discriminator.
The mathematical formulation of a GAN, signature of the portrait
So, is this formula alone the author of this painting? Shouldn't its creator receive all the credit? Or is this formula just the brush with which the artists expressed themselves?
Technically speaking, the mathematical formulation is just the instruction manual. To achieve the result presented at Christie's, it is the data itself—the paintings that inspired the algorithm—that produced this unique piece.
However, this uniqueness is peculiar, as it seems closer to randomness than to a stroke of genius. As we described in the previous article, the algorithm is incapable of understanding. It does not see faces, master paintings, does not consider the context or impact of its stroke, knows nothing of art history, and is incapable of expressing or conveying emotion.
For many, the distinction lies precisely in the point of intention. The artist, too, is a product of their influences, their environment, but their work is primarily the result of their intent. Some, like Paul Gauguin, invoke the divine, while others, like Victor Hugo, see it as a fatalistic expression of the human condition. André Malraux, on the other hand, reminds us that art is an exchange between humans. Yet, all seem to agree on the primary importance of intention in the artistic process.
From this perspective, the Portrait of Edmond Bellamy is indeed human. Humans chose the images on which the algorithm was based. Humans invented and implemented the algorithm. Humans see and interpret the work. Finally, humans assign value to it, sell it, or buy it.
The Brush of the Future
The introduction of technological advances to the art world offers extraordinary possibilities to creatives from all backgrounds. The opportunity to experiment with new tools has always expanded the realm of possibilities for artists. Technology once again serves humanity where we feared it might replace it.
Abstraction work by Tom White
Since the late 1960s and the advent of computers, generative art has captivated an entire generation of artists. Like many other disciplines, computer-assisted art is now undergoing a second digital revolution, that of data. Instead of generating complexity from simple rules, today's artists attempt to do the opposite: distill the simplicity of a complex visual language.
Mario Klingemann was one of the pioneers in the field, using neural networks to augment or transform existing works. Other artists use these methods to stylize their own sketches, like Helena Sarin. Tom White attempts to extract abstract representations from the way neural networks interpret objects presented to them. The first incursion of these algorithms into the realm of art dates back to Google's "Deep Dream" project, led by Mike Tyka, producing psychedelic representations that fascinated the general public.
Left: 'The Fall of the House of Usher' by Anna Ridler. Right: 'The Butcher's Son' by Mario Klingemann.
It is worth noting that the auction also sparked additional controversy, with a case of dispute over the exact authorship of these portraits. Young artist Robbie Barrat was disturbed by the strange resemblance between his creations and those of the Parisian collective. It was revealed that the three young Frenchmen had freely drawn inspiration from a model shared under an open-source license by the American artist. While they denied explicitly copying the model used, they admitted in interviews that their intervention had been minimal at best. As always, the question of attribution takes on a different dimension when colossal sums are at stake, not to mention a place in posterity as the first algorithm-generated painting to be sold at a major auction house.
An excerpt from Robbie Barrat's work
Machine learning researchers are well aware of the possibilities offered by these ultra-powerful models. A concerted effort to make these models accessible to as many people as possible is bearing fruit. One of the pioneers of this effort, David Ha, captured the imagination with his Sketch-RNN project, which invites you to make a simple sketch while watching a neural network attempt to complete the drawing in progress. Google now allows anyone to experiment with the latest iteration of their GANs, offering the chance to create image blends. As in the example of the portraits, this involves interpolating between several examples. Anyone can try transferring the style of a painting to a photo or transforming a horse into a zebra.
A recurring theme in this column will be the power of cooperation between humans and machines. Some argue that the advent of algorithmic art signals the death knell of artistic creation. Others see in it a tool of near-divine power. As often, reality lies in nuance, between a tool that stimulates creativity and a questioning of the nature of art and creation. In any case, these questions and their potential answers will only make sense in their human context, by and for humanity.