IBM hopes 1 million diverse faces can reduce bias in AI

IBM made a million-face data set to help reduce bias in facial recognition technology.

IBM Research

IBM Research on Tuesday released a new data set that contains 1 million images of diverse human faces, with an aim to help advance fairness and accuracy in facial recognition technology.

“For the facial recognition systems to perform as desired — and the outcomes to become increasingly accurate — training data must be diverse and offer a breadth of coverage,” wrote John Smith, an IBM fellow, in a blog post. “The images must reflect the distribution of features in faces we see in the world.”

This comes after artificial intelligence in facial recognition systems has reportedly shown bias. Last week, an MIT study revealed that Amazon’s Rekognition tech had a harder time recognizing the gender of darker-skinned women and made more mistakes identifying gender overall than competing technologies from Microsoft and IBM.

While researchers are already working with attributes like age, gender and skin tone, these features can’t adequately characterize everyone, according to IBM. Things like face symmetry, facial contrast, the pose the face is in, and the length or width of eyes, nose, forehead, mouth and more need to be considered.

IBM’s data set, called Diversity in Faces, has 10 coding schemes, which include features like head length, nose length, forehead height, facial ratios, age, gender, pose, resolution and more.

The million-face data set is available today to researchers around the world on request. 

Source link

Leave a Reply

Subscribe to our newsletter

Join our monthly newsletter and never miss out on new stories and promotions.
Techhnews will use the information you provide on this form to be in touch with you and to provide updates and marketing.

You can change your mind at any time by clicking the unsubscribe link in the footer of any email you receive from us, or by contacting us at We will treat your information with respect.

%d bloggers like this: