Fractals and Forgeries: Teaching AI to Identify Real Jackson Pollock Paintings

It can be tempting to look at an abstract painting and think almost anyone could do that. When the work in question could be worth hundreds of millions of dollars, like one of the paint-splattered canvases of the abstract expressionist Jackson Pollock, the temptation can apparently be overwhelming.

In the late 1940s, Pollock revolutionized the art world with a distinctive technique of pouring or dripping cans of paint directly onto the canvases. While there are 189 verified Pollock “drip paintings” in existence, there are probably twice that many convincing forgeries in circulation, says Richard Taylor, a professor of physics, psychology, and art at University of Oregon. Many have been involved in high-profile controversies and scandals over their authenticity.

Taylor first became fascinated with Pollock when he came across images of his paintings in an art book at school as a child. He started using computers to examine art in the 1990s, and today his interest in Pollock is both academic and aesthetic. 

“We want to make sure the legacy of a great artist is protected,” Taylor said. “We don’t want any fakes to be mistaken for masterwork.” 

That’s why Taylor and former UO doctoral students Caleb Holt and Julian Smith (no relation to the writer of this article) decided to see if they could use AI to tell the difference.

Taylor and his colleagues used a neural network called ResNet, short for Residual Network, which was introduced by Microsoft in 2015 and is often used in computer vision tasks, especially for recognizing objects in images. To train the model they started with a data set of images of all the authenticated drip paintings.

For examples of faux Pollocks, they used works by other famous artists imitating his style, such as Max Ernst, as well as paintings made for the 2000 film Pollock by actor Ed Harris, who studied the artist’s techniques intently for the role. Other examples were made for the study with a device the researchers designed called the Pollockizer, which used a dripping paint container suspended from a line over a canvas. Some came from “Dripfest” events they held for adults and children to try their skills.

The final total of 588 images still wasn’t enough for robust statistical analysis, however. From previous work, Holt and Smith knew that Pollock’s drip paintings were fractal, meaning they repeated similar patterns at smaller and smaller scales. “In a way, a big Pollock painting is made up of all of these miniature Pollocks,” Taylor says. When they divided up the training images into tiles, the total rose to 250,000, which boosted the strength of the model significantly.

When the analysis was finished, the neural network had learned to distinguish between authentic and imitation Pollock poured works with 98.9% accuracy. There were a few false positives, including one by the Pollockizer and another by the French artist Henri Michaux, a contemporary of Pollock. But overall, Taylor says, “It was better than I was expecting.”

The study, published in the journal PLOS ONE, also found that Pollock’s visual signature remained consistent over the period he was making the drip paintings, which has been a subject of debate in the art world. “We were also able to look at the signature at different locations on the painting and show that it deteriorated just slightly towards the edge,” Taylor says. “So the AI tool is not just a technique for authenticating, but it’s also a good way of learning new things about Pollock.”