Professor creates algorithm to rate creativity of art work
Ahmed Elgammal, a professor in the Department of Computer Science, is the director of the Art and Artificial Intelligence Laboratory at Rutgers University. The lab was formed five years ago and works on aspects of perception and cognition related to human creativity in order to develop artificial intelligence.
Recently, Elgammal gave a lecture titled “The Shape of Art History in the Eyes of the Machine,” where he discussed the ways that computer science and artificial intelligence can be used to analyze art.
Elgammal said that he often hears people say that art is subjective while science is objective, so art cannot be studied with scientific methods — but if everything in the universe is governed by laws, then so is art history.
Elgammal said that his lab is motivated from the artificial intelligence point of view rather than trying to replace the work of art historians or people in that field.
“When humans look at art, it involves perceptual, cognitive and intellectual abilities. And ultimately that is the goal of AI. That’s why we train or program the machine to look at art, then it can tackle these kinds of problems,” he said.
In 2015, Elgammal, along with members of his team, came up with an algorithm that measures the creativity of a work of art.
There are certain tasks deemed to be creative that only humans can do, such as creating art or poetry, making a joke or creating a story. For a long time, humans have been the judge for creativity, but now the question remains if whether machines can do the same, Elgammal said.
Creativity itself is hard to narrow down to a single definition. Of the many definitions available, Elgammal’s algorithm defines creativity as something that is novel and influences future works of art.
This is the hard part about creativity because people can define creativity in different ways, Elgammal said. There is psychological creativity, which is when a person creates something that is novel for them individually even though somebody else did something like that before.
Then there is the historical creativity in which one must look at the whole history of human art for example and try to see if that product creative at the time it was made, he said.
“The most agreed upon definition of creativity is something that is novel, different from prior work and at the same time it is influential — moving forward, people will appreciate the creativity in that product and will try to copy it or make variations of it,” he said, “So we tried to model these novelties and influences in a group of products.”
The algorithm formulated falls under a class of algorithms known as Network Centrality, Elgammal said. These types of algorithms are used in multiple applications that involve a large network of connections such as in the analysis of traffic or epidemics.
In order to demonstrate the performance of the algorithm, it was applied to datasets with over 62,000 painting, according to the lab page.
During the process, the algorithm looks at the paintings to see how similar they are. If something is not similar to prior work and similar to works looking forward, then it is novel and influential, Elgammal said.
The algorithm is successfully able to identify artworks that were influential at that time.
“This is really amazing because it tells us an objective measure on which this art was created. It tells us also this creative works that even art historians know are creative are not just subjective judgment, it's basically a very objective judgment. It's coming based on a formula and we know exactly how it works, nothing subjective here,” he said.
Currently, the algorithm involves two criteria — novelty and influence. The algorithm could be adjusted to put more weight on one factor over another and could include other specific elements of art as well.
“When it comes to art, we judge creativity based on certain elements. Are you looking at creativity of composition or color or subject matter? All these elements can be used to make the judgment,” he said. “You can control of all these things and based on them you can really understand art history in a quantifiable way.”
Elgammal said that this algorithm could be applied to other pieces of work such as poetry, literature and music as well. Having a background in computer vision, he chose to focus on art to verify his creativity algorithm.
The team specifically chose art history because there is more information on the art movements, the styles and who was important, Elgammal said. When the team gets some results, they can always compare it to this knowledge which is much easier because it is more agreed upon.
“When we look at art history, people understand exactly who is who and who did what and who’s influential in our time,” he said.
Elgammal said that humans are good at looking at the micro level, but the machine is good at the macro level. Humans can look at a couple of paintings at a time, but cannot look at millions or hundreds of thousands of paintings and analyze them.
The team hopes this machine can help in finding out things that maybe a human has never seen before.
“A few years ago, we found out that two paintings, one by a French artist from 1880’s and another painting by American Artist Rockwell made in 1950. More than 70 years apart but when you look at them, they have striking similarity. Nobody ever looked them side by side, because they’re from different places and time zones, so no noticed this. These are the kinds of things machines can bring to our attention,” he said.
Elgammal and his team are also working on other projects that allow the machine to classify paintings into certain categories, such as renaissance. Another project includes determining which paintings were influenced by which previous artists.