Almost all flagship smartphones support face recognition, but the conclusions about the high accuracy of modern systems are only partially true. In practice, the result of the work of algorithms strongly depends on the sex and color of the human skin.
Researcher Joy Buolamvini of MIT Media Lab has created a compilation of 1270 politicians from countries with a large number of women in power. This set of data, she "fed" the three face recognition systems from Microsoft, IBM and Megvii (China). Artificial intelligence incorrectly defined sex in less than 1% of white men and 7% in light-skinned women. In the case of dark-skinned men, the error was 12%, and with the recognition of black women, the error was already 35%.
Here it is worth remembering the story with the Google Photos service, which in 2015 marked a photo of African Americans with the tag "gorilla". The search giant acknowledged that they removed the forbidden word from the database in order to avoid getting into similar situations due to algorithm errors. Likewise, the images were no longer displayed in the search for "chimpanzee" and "monkey".
Why is that?
Most likely, the difficulty in recognizing women and blacks is not related to the limitations of computer vision technology. Data solve everything, and the "bias" of the algorithm appears at the learning stage. In one of the widely used image bases, the proportion of men exceeds 75%, while 80% of the photos show light-skinned people. For the same reason, Asian-developed face recognition systems better distinguish Asians.
To avoid such problems in the future, algorithm developers need to consider different demographic and ethnic groups.
Source: The Verge
Read in Russian: Алгоритмы распознавания лиц хуже различают женщин и чернокожих