A study conducted by Sony’s AI Ethics department shows how artificial intelligence does not take skin tones into account when it comes to facial recognition or even the creation of fictitious faces.
Facial recognition is increasingly using‘artificial intelligence. To monitor fraudsters in the metro or get rid of queues before taking the Eurostar for example. To achieve this, algorithms must be trained Of course. They analyze millions of faces and learn to perfect their identification. The system is not flawless and can even lead to fairly serious miscarriages of justice.
In 2018, a study revealed that recognition was less effective for dark-skinned people. This has led to a change in several companies like Google Or Metawho adopted a more inclusive frame of reference, that of Ellis Monk. Sony believes, however, that a bias remains. A study of his division AI Ethics show that yellowish skin tones are almost ignored by AIin favor of shades going towards the red.
Sony believes AI favors some skin tones over others
The researchers propose to remedy this byautomate skin color analysis by placing it on two distinct scales: from light to dark and of mostly yellow to mostly red. A program isolates the pixels of an image showing skin, converts their RGB value (red, green, blue) into codes used by the repository CIELAB deemed the most relevant, then calculates an average value. For Ellis Monk, creator of the scale used by Google, among others, this would be to ignore a human analysis according to him important in eliminating bias.
Beyond choosing a definitive solution, some firms have already declared that they will study the conclusions of the Sony study. Its implications go beyond facial recognition for surveillance purposes since they also include AI photo creationL’smile identification and even the photo cropping tools. Alice Xiang, head of the AI Ethics department at Sony, has no doubt thatimprove algorithm training systems is necessarily a never-ending task. “We have to keep trying to make progress,” she concludes.
Source: Wired