"But such training typically requires images that are labeled, so that the computer can tell whether it guessed right or wrong and refine its concept of a human face accordingly." . . . .