Affective Computing: Teaching Machines About Emotion

jbc writes "The L.A. Times is running a story about affective computing, a field in which researchers are programming computers to recognize human emotions through the use of such clues as facial expression, vocal tone, and blood pressure. Some hail it as the dawn of a new era in super-useful machines, while others warn about invasions of privacy."

Comments

Faking it in AI

[Animats]

I've noticed a tendency in the AI community to work on stuff like this when they're not making any progress on the real problems. Stanford's Knowledge System Lab was into this just before they tanked, and you can still see posters for some of the drama-related projects in the abandoned cubicles on the second floor of the Gates Building.

Like Eliza, systems that seem to have emotions generate responses from humans that cause them to be overestimated. Parry, which was developed in the 1960s along the lines of Eliza and simulated a dialog with a paranoid, was probably the first program to have "emotional state". So this isn't new.

Even something as simple as the Furby has that effect. (I'm not criticizing the Furby; I've met the designer, and he's just trying to make a toy kids like. He doesn't make any unreasonable claims for the toy.) It's a great way to get press coverage, because it yields good demos.

Dolls that fake emotions have been around for a while. The classic is Baby Think It Over, the attention-demanding doll from hell used to convince teenagers not to get pregnant. Hasbro marketed, as My Real Baby, a lower-cost (and less obnoxious) version designed by some of Rod Brooks' people from MIT.

And, of course, there are the Sims.

It doesn't take much internal state to fake emotions. It's typically just a few scalar values going up and down in response to inputs.