Training computer vision systems to recognize the facial biometrics of a range of ethnicities can help to improve their accuracy, suggests a new Google study.
Google’s researchers focused their efforts on algorithms designed to detect smiles in images of faces, training them on four race subgroups – Asian, black, Hispanic, and white. The algorithms set a new standard for accuracy in detecting smiles, at 91 percent, reflecting an improvement of 1.62 percent over the previous record.
That may seem like a relatively minor gain on its face, but it’s important because of the larger implication that training AI for facial recognition across multiple ethnicities can lead to a better result than training that does not address a variety of racial features. When it comes to facial recognition, this is a serious concern, with civil rights advocates like the ACLU objecting to government use of the technology in part on the grounds that studies have suggested it can be less accurate on minority ethnic groups, and thus lead to discriminatory outcomes. Meanwhile, Apple, in seeking to establish its new infrared facial recognition system as the gold standard for user authentication on smartphones, says it trained its technology on a billion images representing a range of ethnicities, ages, and genders to help ensure that it performs exceptionally well.
If biometrics and computer vision specialists take this research seriously and apply its approach in their own work, that exceptional performance could become more the norm in the future – an outcome that would benefit everyone subject to facial recognition technology, whether by government observers or their own mobile devices.