This week, the denizens of Twitter began posting photos of themselves with an odd array of labels. Some, like “face,” were confusingly benign, while others appeared to verify harder truths: Your humble writer was declared a cipher, a nobody, “a person of no influence.” Fair enough. But many of the labels were more troubling. There were rape suspects and debtors. A person would be labeled not just black, but “negro” and “negroid.”
The project, called ImageNet Roulette, is an effort by artist Trevor Paglen and researcher Kate Crawford to illustrate the dangers of feeding flawed data into artificial intelligence. It takes aim at one of the field’s seminal resources: ImageNet, the database of 14 million images that’s credited with unlocking the potential of deep learning, the technique used for everything from self-driving cars to facial recognition. The algorithm behind the Roulette tool is trained using images within ImageNet that label people across 2,395 categories, from “slatterns” to “Uzbeks.” “I wanted to crack ImageNet open and look at images that weren’t meant to be looked at,” says Paglen. The experiment, now viral, has plenty of people asking just how those labels got there in the first place, and why they remain [Read more].