18 September 2015
Translation: Artificial Intelligence: When the Machine Imitates the Artist
This article by Morgane Tual was originally published in Le Monde on September 8, 2015 in French.
Some paint like the great masters, others improvise on jazz … Endowed with a kind of imagination, some artificial intelligence programs manage to compete with the human mind.
An algorithm capable of generating works like Picasso’s and Van Gogh’s: it’s what a team of German researchers has recently come up with. They detail their invention in an article published at the end of August 2015. By analyzing the canvases of these prestigious artists, the machine is capable of “learning” their style through deep learning and tracing it on any photo with impressive results in less than an hour.
Are machines in a position to compete with people in the area of creativity? Conscience and emotions, creativity is part of the attributes often put forth in order to distinguish humans from machines. Would a program be able to come up with moving paintings, to imagine thrilling stories, or to rival Mozart’s composition genius? Artificial intelligence (AI) researchers are working on it.
A Recurring Question
But what are we talking about? The notion of creativity is the subject of debate in the research community. Are we talking about artistic ability? Imagination? Invention? Or even the capability to problem-solve? It’s a recurring question in the AI field, where even the notion of intelligence causes debate. “When Alan Turing wrote his foundational article on artificial intelligence, he was confronted with this problem of definition,” explains Mark Riedl, AI researcher at Georgia Institute of Technology. “Instead, he decided to invent a test, which basically says that if the behavior of a machine is impossible to distinguish from that of a human being, then the machine is considered to be intelligent.”
Mark Riedl used the same procedure for creativity, designing the Lovelace 2.0 test in 2014. In this experiment, the judges ask a program to create a work (a painting, poem, architecture …) that has a theme, which has not been defined in advance. No machine has succeeded in passing it for the time being. But Mark Riedl thinks that it’s possible. For him, machines are already capable of creativity, to a certain extent:
“Many people think of the great artists when they use the word creativity. But each human is creative to a certain extent, and this creativity manifests itself daily, dozens or hundreds of times per day. We show our creativity when we play Pictionary, when we use a paper clip to fix a pair of glasses, or when we find another way to get home if the road is blocked. Computers already possess this kind of creativity.”
Combining Already Recorded Elements
But beyond this, a certain number of programs are already able to show their creativity in the artistic field, appealing to a kind of imagination. “All imagination is understood as the recombination of pre-existing elements of memory,” explains Jean-Gabriel Ganascia, researcher at the Paris-VI computer science laboratory, in his book “Ideas Received on Artificial Intelligence.” He cites the example of the unicorn, “produced par excellence by our imagination,” which combines two real beings known to man: the horse and the narwhal. The artistic creativity of machines will generally work in this way, by combining already recorded elements to create new ones. With interesting results.”
Emily Howell composes classical music
David Cope, professor of music and computer science at the University of California, has worked over these last few decades on a machine capable of composing classical music. Its first program, called EMI, “gets its inspiration” from the great composers to create its own music. Basically, David Cope “feeds” EMI several musical pieces by Vivaldi, for example, which it will analyze with the aim of identifying patterns and rules. From these diagrams, it is able to put together its own compositions in a style that approximates that of the composer.
David Cope decided to go further, inventing another program called Emily Howell. It works in the same way, but it feeds off of EMI’s output to come up with its own music, which gives it a more “personal” style, approximating contemporary music. Emily Howell put out its first album in 2010.
Shimon improvises on jazz
On a similar principle, Shimon, designed by Guy Hoffman, is able to improvise on jazz live, starting from a statistical model based on the improvisations of the pianist Thelonious Monk. In this video, the robot adapts in real time to the music it discovers, played by its designer.
Scheherazade writes stories
Nourishing itself on the works of others in order to notice recurrences: this is also what Scheherazade, a program able to write short stories of one or two paragraphs, does. Its designer, Mark Riedl — the inventor of the Lovelace 2.0 test — feeds it writing on bank robberies, for example, if he wants it to come up with a story on this topic. “It does not use any preprogrammed knowledge, it learns everything it needs to know to create its story.” He assures that the writing it produces “is difficult to distinguish from human writing.”
Watson invents kitchen recipes
This technique also allows Watson, the star artificial intelligence program of IBM, the concoct kitchen recipes after having analyzed thousands of them. Chefs have implemented the recipes invented by the machine. They are more or less strange, like a cocktail of cider and pancetta, or the beef burrito enlivened by chocolate and soy beans. The best have been collected into a book, “Cognitive Cooking with Chef Watson.” “This resembles a very strange fusion restaurant,” says in an article a CNN journalist who had the chance to taste this cuisine. An app is even available to allow internet users to “create unique dishes with Watson.” After a weeklong test period, a rather convinced blogger nevertheless clarified that it was necessary to make some adjustments, “Ask yourselves if you really want to add mashed potatoes to lasagna.”
“A Unique Point of View” on Humanity
Artificial intelligence is thus already capable of certain kinds of creativity, but it nevertheless has its limits. “While humans can be creative in different fields, most algorithms concentrate on one thing,” emphasizes Mark Riedl. “A poetry generator cannot draw for example.” But above all, machines perhaps lack attributes of human beings that are potentially essential to match their creativity. For Michael Cook, an associate researcher a the University of London who has created a program capable of inventing video games, machines do not feel emotion, which represents a limit to their creative capacity:
“This kind of thing is really important for creativity — it’s in this way that we feel connected to others, that we are touched by artists. We often understand the work of artists by comparing it to our own lives. Experiencing war, love, having a history with a city, a country … AI has trouble having this kind of impact because we share fewer things with it than with other human beings.”
Still, he emphasizes, this fundamental difference could also present an advantage, “AI can offer us a unique point of view: a view of humanity from the outside. We have not really begun to exploit it, but I believe that one day we will.”
Strange and Unpredictable Results
Ultimately, artificial intelligence could also permit human beings to learn more about their own creativity … This year, a research team from Google has invented a program, Deep Dream, which creates impressive, phantasmagoric, dream-like images that sometimes recall the paintings of the Dutch painter Jerome Bosch. However, this program has not been designed to be creative; Deep Dream is part of a research project on machine learning.
Created by a Google team, it is “fed” millions of images in order to learn to detect forms. Then, by giving it a new image, the engineers ask it, “Whatever you see, we want more of it!” “If a cloud looks a little bit like a bird, the program will make it look even more like a bird,” they explain on a blog. And this provides strange and unpredictable results that, more than two months later, continue to fascinate internet users.
The experiment, with impressive results, has given rise to new lines of questioning for researchers. For them, Deep Dream “could become a tool for artists — a new way to remix visual concepts — or maybe even shed a little bit of light on the roots of the creative process in general.”