AI Project Produces New Styles of Art
Researchers let two neural networks critique each other to create the images
/https://tf-cmsv2-smithsonianmag-media.s3.amazonaws.com/filer/3b/6e/3b6e8a7a-abd1-4b4f-94e6-6dcd4a3581dc/machine-art.jpg)
Artificial intelligence is getting pretty good at besting humans in things like chess and Go and dominating at trivia. Now, AI is moving into the arts, aping van Gogh’s style and creating a truly trippy art form called Inceptionism. A new AI project is continuing to push the envelope with an algorithm that only produces original styles of art, and Chris Baraniuk at New Scientist reports that the product gets equal or higher ratings than human-generated artwork.
Researchers from Rutgers University, the College of Charleston and Facebook’s AI Lab collaborated on the system, which is a type of generative adversarial network or GAN, which uses two independent neural networks to critique each other. In this case, one of the systems is a generator network, which creates pieces of art. The other network is the “discriminator” network, which is trained on 81,500 images from the WikiArt database, spanning centuries of painting. The algorithm learned how to tell the difference between a piece of art versus a photograph or diagram, and it also learned how to identify different styles of art, for instance impressionism versus pop art.
The MIT Technology Review reports that the first network created random images, then received analysis from the discriminator network. Over time, it learned to reproduce different art styles from history. But the researchers wanted to see if the system could do more than just mimic humans, so they asked the generator to produce images that would be recognized as art, but did not fit any particular school of art. In other words, they asked it to do what human artists do—use the past as a foundation, but interpret that to create its own style.
At the same time, researchers didn’t want the AI to just create something random. They worked to train the AI to find the sweet spot between low-arousal images (read: boring) and high-arousal images (read: too busy, ugly or jarring). “You want to have something really creative and striking – but at the same time not go too far and make something that isn’t aesthetically pleasing,” Rutgers computer science professor and project lead, Ahmed Elgammal, tells Baraniuk. The research appears on arXiv.
The team wanted to find out how convincing its AI artist was, so they displayed some of the AI artwork on the crowd-sourcing site Mechanical Turk along with historical Abstract Expressionism and images from Art Basel's 2016 show in Basel, Switzerland, reports MIT Technology Review.
The researchers had users rate the art, asking how much they liked it, how novel it was, and whether they believed it was made by a human or a machine. It turns out, the AI art rated higher in aesthetics than than the art from Basel, and found "more inspiring." The viewers also had difficulty telling the difference between the computer-generated art and the Basel offerings, though they were able to differentiate between the historical Abstract Expressionism and the AI work. “We leave open how to interpret the human subjects’ responses that ranked the CAN [Creative Adversarial Network] art better than the Art Basel samples in different aspects,” the researchers write in the study.
As such networks improve, the definition of art and creativity will also change. MIT Technology Review asks, for instance, whether the project is simply an algorithm that has learned to exploit human emotions and not truly creative.
One thing is certain: it will never cut off an ear for love.