Thursday, December 10, 2009

Reading Faces

Willl computers soon be able to read facial expressions better than humans can? I don’t think so.

  • ·         Our brains have evolved specifically to, among other things, recognize facial expressions.
  • ·         So far, we’re not very good at teaching computers to interpret anything much. Have you used an online translator lately? I would think that accurately translating meaning from one language to another is sort of similar to extracting meaning from facial expressions (tell me why I’m wrong about that one), and I’d further think that translation from one language to another is generally more deterministic than is reading facial expression.
  • ·         If a computer program is well understood, then it should be reasonably easy to fool. It might be as easy as simply using an extreme expression (say a really big smile) to swamp a computer program’s algorithms with false data. People have the advantage of being able to spontaneously create new rules and hypotheses to explain the facial expressions they observe, and are therefore more robust to this kind of deception.
  • ·         Many insights into interpretation of facial expression could be used by people as easily as by computers, so computers would only gain an advantage if an effective face-reading technique was discovered that plays to the abilities of computers, and against the abilities of humans. Currently I think most face-reading techniques play to human strengths.

To me, the real advantage would be if computers could tell us something about someone else that we can’t tell on our own (I’m considerably less interested in the idea that computers are simply cheaper per hour at reading faces than are live humans).


  1. Robert, I suspect you're right. Three, semi-related things come to mind:

    1) Jonah Lehrer's (_How We Decide_) comparison of poker to chess or backgammon in terms of building a computer that can defeat a human opponent. Poker is so much about reading people's faces and mannerisms, that it's unlikely a computer will ever seriously compete.

    2) Pareidolia, specifically seeing faces in inanimate objects and patterns. Our brains are seriously wired to see faces.

    3) Facial Action Coding System (FACS) - amazing stuff. People can be trained to be incredibly good at it, but could a computer? Like you, I doubt it.

  2. Bob,

    Excellent points!

    So, why am I wrong that translating meaning from one language to another is like reading faces? I've got to be wrong about that one, and you're the kind of guy who can tell me why!