Afroscanned
Machine learning is a hot topic in the tech and business industries. With everyone scrambling to stay on top of the emerging trend, any potential problems are often overlooked. Former artificial intelligence (AI) research assistant and current founder of Paere, a creative technology startup with a focus on inclusion and accessibility, Lequanne Collins-Bacchus, found an issue with a kind of machine learning that classifies images.
Lequanne came to MozFest to discuss image classification and algorithmic biases that ensue. Convolutional neural networks scan multiple layers of pixels to determine what the components of an image are. These pixels have a positive value if they’re white and a negative value if they’re black. That association is already problematic, Lequanne pointed out during her MozFest session “Afroscanned.”
Algorithms learn how to identify based on the information they are fed. Lequanne pointed to an incident in which Google Photos categorized two black people as gorillas and a Nokia camera malfunction in which every photo taken of Asian faces prompted the message “Did someone blink?”
The explanation for these racist errors relates to the white and black pixels automatically scanned during every image classification. “Computer vision,” Lequanne explained, “describes the process of automatic extraction, analysis, and understanding of useful information from a single image. It relies heavily on white values.”
Get MozFest’s stories in your inbox
Join Medium for free to get updates from this writer.
Though there was no quick fix solution to racial discrimination in artificial intelligence, Lequanne did offer participants a 3D selfie using a laser scanner and her iPad. The technique, known as photogrammetry, captures several images of the same object. An algorithm then combines the information into a 3D model.
Through this thorough process, photogrammetry “gathers reliable information about the world around us,” Lequanne said.
Lequanne is a black woman, representing a voice that makes up only 3% of the whole tech industry, according to the National Center for Women & Information Technology. Having that perspective on artificial intelligence is important at a place like MozFest. One “Afroscanned” participant voiced interest in hosting digital inclusion workshops in New York that highlighted the issues Lequanne raised. Another participant plans to incorporate Lequanne’s observations into existing AI projects.
“The digital is ubiquitous and we need to constantly critique how we experience it in order to analyze the many ways it can be unhealthy, discriminatory, or opaque,” said Lequanne. You can get in touch with Lequanne here.
