U of C grad student tries to bring image recognition technology into better focus

University of Chicago grad student Steven Basart and two other researchers collected photos that confused AI systems. | Provided

From his Hyde Park apartment, Steven Basart would spent hundreds of hours scouring the internet for pictures that might trip up artificial intelligence image recognition software.

He was looking for real-world images that would truly test a technology that an increasing number of businesses — think autonomous vehicles and tumor detection — hope to rely on.

The typical set of images used to test the software consist largely of photos that are well cropped, in focus and with little going on in the background that might cause distraction. The result is an identification rate of more than 90%.

Critics said the test was too easy. So researchers began throwing a series of curve balls in the form of unusual computer generated images — some as strange looking as the black-and-white wavy lines that appeared on old tube televisions that weren’t working properly.

But Basart and fellow researchers from the University of California Berkeley and the University of Washington thought: If these machines were going to be operating out in the real world, they should face images that reflect the that world.

So they set about culling photos from the internet that would offer a more realistic challenge.

Artificial intelligence models consistently misidentified this image as a school bus.

AI models mistook this alligator for a unicycle.

“If AI models test well against our data set, they will be more robust in general and perform better in the wild,” he said.

“Background cues and different color pattens and textures tend to fool them,” he said. “For instance, a photo of a squirrel in falling snow was interpreted as a shovel. Or a leopard-pattern couch in a living room was interpreted as a leopard.”

AI models correctly identified the images collected by Basart and his colleagues only 2% of the time.

On July 18, the group posted their research — consisting of 7,500 images — to github, a free website known for hosting code. A paper explaining the research was posted to a separate website.

“Anyone willing to test their models against our data set is free to do so,” said Basart, who worked on the project through the Toyota Technological Institute at Chicago, which is closely affiliated with the U of C.

It’s too early to know if tech giants like Google or Facebook will use the data, Basart said.