December 15, 2017 -
Google researchers have discovered a way to improve the company’s smile detection algorithms across genders and races, which reaches a state-of-the-art 91 percent on the Faces of the World dataset, according to a research paper published on arXive (PDF).
Technology Review reports that this represents a 1.5 percent improvement over the previous best accuracy.
The researchers used four racial and two gender classifications to improve accuracy while preserving privacy, according to the paper. Their method involves utilizing transfer learning and “some of the last hidden layers of a face recognition model” to train the demographic classifiers, “then combining the last hidden layers from these classifiers into a third model trained for the task of smiling detection,” according to the paper.
The paper includes discussions of ethical considerations and fairness in machine learning, including the concepts of “Fairness through Awareness,” which suggests that awareness of “sensitive characteristics” like gender and race is important to models that work well for different demographics, and “Equality of Opportunity” and “Equality of Odds,” which require equal rates of false negative and false positive results, respectively, across subgroups of the population.
Last month, University of Surrey researchers developed a facial recognition system with accuracy improved by recognition of different races.