Earlier this year, Microsoft released new facial recognition software to developers that offered great promise but turned out to be racist. After the company’s 2016 “Tay” debacle — surely you remember Microsoft’s neat AI Twitter bot that turned into a crazed Nazi in under 24 hours — this was the last thing Microsoft needed. The company’s “Face API” offered developers access to what the company claimed to be advanced facial recognition technology that could be integrated into their own products. But researchers at MIT and Microsoft’s own New York-based research lab found that the company’s new facial recognition tech had some serious problems with race. And gender.
According to a paper published by the researchers, Microsoft’s Face API facial recognition technology was remarkably proficient at identifying white males. In fact, the team found that Microsoft’s tech could identify white male faces with an error rate of 0%. That’s fantastic, but the finding was overshadowed by the face that the error rate skyrocketed as high as 20.8% when the tech tried to identify black women. Microsoft clearly had some work to do to make things right, and in a blog post on Tuesday evening the company claims it has made some serious improvements to its facial recognition tech.
“Microsoft announced Tuesday that it has updated its facial recognition technology with significant improvements in the system’s ability to recognize gender across skin tones,” Microsoft’s John Roach wrote in a post on the company’s AI blog. “That improvement addresses recent concerns that commercially available facial recognition technologies more accurately recognized gender of people with lighter skin tones than darker skin tones, and that they performed best on males with lighter skin and worst on females with darker skin.”
He continued, “With the new improvements, Microsoft said it was able to reduce the error rates for men and women with darker skin by up to 20 times. For all women, the company said the error rates were reduced by nine times. Overall, the company said that, with these improvements, they were able to significantly reduce accuracy differences across the demographics.”
Translation: Face API is now less racist.
Microsoft says the high error rates its tech experienced on female faces with darker skin is indicative of a serious challenge facing the industry as a whole. Artificial intelligence tech needs as broad a dataset as possible in order to be trained well. Of course the blame for this issue still lies squarely on Microsoft: The company failed to provide enough image data on black people to build a useful facial recognition product.
“The Face API team made three major changes,” Roach wrote in his post. “They expanded and revised training and benchmark datasets, launched new data collection efforts to further improve the training data by focusing specifically on skin tone, gender and age, and improved the classifier to produce higher precision results.”
The Azure Cognitive Services team’s updated Face API software is available beginning immediately on Microsoft’s site.