Facebook’s new artificial intelligence team reports that it has developed software that can match faces with 97.25-percent accuracy, compared to human beings’ 97.53-percent accuracy. The DeepFace software is an application of deep learning, in which networks of simulated neurons are used to learn to identify patterns in large amounts of data. DeepFace uses a two-step technique to process face images, and in the first step it corrects the angle of the face so the person in the image faces forward, using a three-dimensional image of an “average” forward-looking face. In the second step, the simulated neural network generates a numerical representation of the reoriented face, and if the software yields sufficiently similar descriptions from two different images, it concludes they must show the same face. Although DeepFace executes facial verification rather than facial recognition, AI team member Yaniv Taigman says some of the former function’s underlying methods can be applied to the latter. University of Washington researcher Neeraj Kumar says Facebook’s results demonstrate that finding enough data to feed into a large neural network can facilitate substantial improvements in machine-learning software.