DeepFace: face recognition near human level
In June 2014 Facebook AI Research and Tel Aviv University presented DeepFace at CVPR: a network of over 120 million parameters, after a three-dimensional alignment of the face, reached 97.35% on Labeled Faces in the Wild, cutting the error by over 27%.
Why it matters
Face verification in unconstrained photographs came close to human level, which the paper itself gives as over 97.5%. Its introduction says outright that the machines' lag had so far been a buffer against the social and cultural implications of face recognition.
Trained on the SFC set: 4.4 million faces of 4,030 people, 800 to 1,200 each. The face is aligned by a three-dimensional model and 67 fiducial points. On YouTube Faces the error fell by over 50%. The human level on LFW is taken from other work, not measured.