Microsoft Improves Face API to Reduce Racial and Gender Bias

This week, Microsoft announced changes to its Cognitive Services Face API that will improve the technology's ability to accurately determine gender across various skin tones. The improvements reduce error rates for men and women with darker skin tones up to 20 times. For women of all skin tones, error rates are reduced nine times. The improvements address a gender and racial bias that is revealing itself across AI and Machine Learning technologies.

"[Microsoft] had conversations about different ways to detect bias and operationalize fairness," Hannah Wallach, a Microsoft Senior Researcher, commented in a company blog post. "[Microsoft] talked about data collection efforts to diversify training data. [Microsoft] talked about different strategies to internally test [Microsoft] systems before we deploy them."

Microsoft, among others, has found that its facial recognition technology has traditionally worked best when the subject is a lighter skinned male. Microsoft researchers attribute this bias towards the datasets used to train its machine learning algorithms. By increasing the number of woman and darker skinned individuals in training datasets, the AI improved.

Microsoft made three major changes to the API. As mentioned, the datasets were improved to increase underrepresented samples. Second, new data collection efforts were implemented to improve training with a specific focus on skin tone, gender and age. Finally, Microsoft improved the classifier. To learn more about Microsoft's efforts, and AI-industry concerns, read through Microsofts announcement.


Be sure to read the next Artificial Intelligence article: Daily API RoundUp: CryptoStandardizer, Voxprox, Meqem, Seobility