Note (16/1/2010): Moshe has informed me that my enthusiasm is misplaced: actually this kind of real AI is in reasonably common use already. He told me that there's a machine that's already learning for itself how to speak like humans, though he didn't know any specifics so I can't verify that. Still, all that can really be taken away from this post is how out of the loop I am on a subject that used to seem like one of the most important things in the world to me.
I knew Project Natal was impressive
. But I didn't realize Microsoft needed to use real artificial intelligence
to make it. Look at this:http://www.popsci.com/gadgets/article/2010-01/exclusive-inside-microsofts-project-natal
The article is short, but here's what I understand from it. When Project Natal sees an image of a room and identifies a person in it, it then puts together a list of likely guesses as to what that body is doing, assigns a probability to each possible interpretation, and then takes the most likely assumption. Nothing more glamorous is going on here than following rules like "If the leg is two pixels further to the left, then increase the probability of case #1,694.". But those rules were not programmed by people.
Somewhere in Microsoft headquarters is a big network of computers that together form what the project workers call "the living brain". I am not making this up, it's right there in the photo gallery on the Popular Science
page. This computer system is not just programmed but trained
to recognize body positions from images. It was only programmed with a basic knowledge of how human bodies are shaped, which is much like how a living creature has basic functions programmed in as instincts. The "living brain" is given pictures and is told repeatedly what body positions those pictures are supposed to stand for, and then it writes its own rules to make sense of all that. When it's finished, the list of rules it's come up with will be put into the considerably-dumber Project Natal systems, which will not learn for themselves but just follow the rules which have already been learned. And those rules aren't objective laws decided on by some programmer or team of experts, they're the personal views of this particular computer network in Microsoft headquarters. They're rules which are the result of this particular program's design and experiences, with all the imperfections that implies.
The article, like I said, is short. It doesn't say whether this same technique is being used to train Project Natal's recognition of emotions, though I imagine it must be. And I'd really be interested in hearing a more thorough analysis of the way they're getting this program to learn. The article doesn't even say if this sort of thing is common nowadays; I haven't heard of anything this ambitious before, but I don't hear much. What this article does
tell me is that I was wrong about AI systems. Clearly real artificial intelligence does
exist, it's just running on hardware too expensive for end users and still needs to be trained by professionals for very specific tasks. It is