Yeah, this is basically just Newton's method of approximation, seeded with a random initial value. It will eventually converge on a local maximum defined by the facial recognition function.