Microsoft retrains facial recognition AI to better recognize gender across skin tones

Microsoft has improved the capability of its facial recognition technology to recognize the gender of people with darker skin tones, reducing error rates by up to 20 times, according to an announcement in a blog post.

Error rates for all women were reduced by nine times, the company says, significantly reducing the differences in accuracy across different demographics for its Face API, delivered via Azure Cognitive Services. Microsoft’s Face API team expanded and revised the datasets used for training and benchmarking, launched new data collection efforts to improve the diversity of training data, and improved the classifier to produce more precise results.

M.I.T. Media Lab Researcher Joy Buolamwini and Microsoft researcher Timnit Gebru found that leading facial recognition systems from Megvii, Microsoft, and IBM returned results with major differences in accuracy between different demographics in a study released earlier this year.

“We had conversations about different ways to detect bias and operationalize fairness. We talked about data collection efforts to diversify the training data. We talked about different strategies to internally test our systems before we deploy them,” said Microsoft Senior Researcher Hanna Wallach, an expert on fairness, accountability and transparency in AI systems.

Microsoft has also been working on a tool for engineers to automatically detect bias in AI algorithms.

The tech industry has been forced to work to reduce bias in algorithms by increased public awareness, which has also resulted in a declaration by human rights groups calling for an ethical framework based on international human rights law to be applied to machine learning systems, and hearings on the issue by the Subcommittee on Information Technology of the U.S. House Oversight and Government Reform Committee.

Article Topics

accuracy  |  AI  |  biometrics  |  facial recognition  |  Joy Buolamwini  |  Microsoft