MICROSOFT

Microsoft Facial Recognition Reduces Error Rate on Gender Distinction
Facial recognition technology is increasingly used, whether in surveillance systems or simply implemented in smartphone unlock security. But there are problems associated with the effectiveness of recognition between men and women, with greater discrepancy between the lighter and darker shades of skin. Companies have made efforts to train artificial intelligence in error reduction, and Microsoft claims significant improvements in its technology.According to the Microsoft blog, the technology has reduced the error of recognition between men and women with black skin up to 20 times. Among women, the error was reduced by nine times. With these values, the company claims a large reduction of error between different demographics. Microsoft says that artificial intelligence technologies are only effective if they are fed with diverse information on skin tones, but also related to styling, accessories and eye make-up.
To this end, the Azure training team, the Microsoft AI system, made important changes. The database was expanded and the training revised, introducing a new collection of information specifically focused on the skin tone, gender, and age of people. The classifier was further improved to produce results with greater accuracy. The team of researchers also counted on specialists to define strategies and good practices, testing internally the system before launching it, investing in a further diversification of the training.
Finally, the research focused on artificial intelligence tools to help engineers identify weaknesses in training information, particularly in the representation of women's darkest tones, which could lead to AI's system of misclassification of gender, which is considered the main challenge in facial recognition accuracy.



Sapo