The facial recognition effective… if you’re White

QMI agency

Sunday, 11 February, 2018 18:56

UPDATE
Sunday, 11 February, 2018 18:59

Look at this article

The programs facial recognition is effective when asked to identify the genre of a white man, but is wrong more regularly, they need to recognize the kind of black people, shows a study from the Massachusetts Institute of Technology (MIT).

To achieve this result, Joy Buolamwivi and his team have analyzed the performance of three software packages from Microsoft, IBM and the chinese company Megvii (Face++). The programs had to try to recognize the type of 1270 people, men and women from three african countries (South Africa, Senegal and Rwanda) and europe (Finland, Iceland, Sweden).

On the one hand, and the software have all three identified much more easily the kind of men that that of women, the chinese software up to get 99.3 percent of correct answers with the men, but only 78.7 per cent with the women.

Things are, however, worse taking into account the skin color of the face. Systems, in particular, were unfit to recognize the type of black women, with a success rate ranging between 65.3 per cent for IBM and 79.2% at Microsoft.

In total, the majority of errors made by the software during the test were of black people, primarily women.

“You can’t have artificial intelligences that are not inclusive. However, those who create the technology are the ones who implement the standards,” stressed Ms. Buolamwivi, in an interview with the New York Times.

“We risk losing the gains achieved with the civil rights movement because of our false impression that the machines are neutral. We need to ask for more transparency,” added the author of the study on a website that is dedicated to him.

The researcher sent the results of its study to the three companies. IBM and Microsoft have responded by indicating that they are seeking to improve their software.

Recognition by software and by categories

Microsoft

• White man: 100 %
• Man in black: 94 %
• White woman: 98,3 %
• Black woman: 79,2 %

Face++

• White man: 99,2 %
• Black man: 99,3 %
• Woman, white: 94 %
• Black woman: 65,5 %

IBM

• White man: 99,7 %
• Black man: 88 %
• White woman: 92,9 %
• Black woman: 65,3 %