Computer Spots Fakers Better Than People Do

Rambo Tribble (1273454) writes "Using sophisticated pattern matching software, researchers have had substantially better success with a computer, than was obtained with human subjects, in spotting faked facial expressions of pain. [Original, paywalled article in Current Biology] From the Reuters piece: '... human subjects did no better than chance — about 50 percent ...', 'The computer was right 85 percent of the time.'"

Comments

In 3 .. 2 .. 1

[OzPeter]

Outlawed by FIFA!

Hollywood

[AlphaWolf_HK]

Perhaps watching faked facial expressions on TV and whatnot has dulled our ability to distinguish them?

Re:Hollywood

[Artifakt]

In the 1970's there was a book called "Four Arguments for the Abolition of Television", or something like that. One of the arguments was the limited image quality of the 512 line scan made even very poorly faked emotions very hard to distinguish from the real thing, and so children who got their learning examples of human expressions from TV would have a hard time telling who was really feeling emotions or just faking them. The author also claimed that emotions such as Rage, Fear, and Strong Suffering would come through better than subtler emotions such as Boredom, Fondness or Compassion, so TV scripts would come to emphasize those emotions which at least somewhat worked and ignore the rest. Perhaps there's something to these ideas.

Re:It wasn't the computer

[ColdWetDog]

That would be an interesting test. I agree, 25 random volunteers really isn't all that high of a bar. Do the same with some experienced clinicians and see what happens.

Moreover, pain is a pretty complex issue - there is acute, nocioceptive (pain receptor) pain - as in the test. Chronic pain is quite a bit different. Visceral pain (from nerve fibers in the abdomen) is different still.

I think that a computer assisted study of emotions has the potential for improving human performance in decoding those emotions, but this is clearly in it's infancy. I don't think there will be an app for that in the near future (a real one, that is).

Re:It wasn't the computer

[Artifakt]

There's a good precedent for your argument that this is a question of instinctual skill vrs trained skill, but it doesn't take anything like a billion examples to train a person in the example I'm considering. A very common way to teach health care personnel to recognize Fetal Alchohol Syndrome is to give them an album with several hundred photos of people in various life stages, all suffering from FAS. This method has worked since the time when the photos were black and white, and in fact, using color shots or video footage doesn't seem to have any impact on success or the number of examples needed. Once someone is trained that way, the success percentage is in the very high 90s, and stays that way, at least for a typical crreer. Similar methods are used for other diseases, for example most people have learned to spot Down's syndrome from just a few examples, but where the syndrome produces only some of the usual appearance effects, spotting the 'borderline cases' with high accuracy can be taught this same way, usually taking about 15 minutes.