Friday, January 14, 2011

News - IBM's Jeopardy-Playing Watson Computer

This is potentially big news, and at the least a very interesting development, in artificial intelligence:

http://news.yahoo.com/s/ap/20110114/ap_on_hi_te/us_man_vs_machine

IBM has developed a supercomputer, called Watson after the company's founder, that is capable of participating in the game show Jeopardy, in entirely human terms.  It listens to the host in spoken English, breaks the audio signal down into text it can analyze, comes up with a plausible answer from its database, and makes a decision about its confidence level in that answer, as it decides whether and when to "buzz in" before delivering it (in the form of a question, of course) in a synthesized voice.  In a practice round, reported in the AP story linked above, it managed to beat a couple of champion human opponents in some (though not all) categories.

It's not too surprising, really, that a computer can pull this off; the elements involved are not hard to envision, given the current state of the technology.  It's still a fairly narrowly defined problem, as well.  But it's impressive because it's dealing with a considerably more complex problem than the traditional chess-playing benchmark -- especially given that Jeopardy tends to use humor and puns in its answer/question format, something computers have traditionally been very poor at parsing.

I look forward to seeing the serious tournament play broadcast scheduled next month -- will this AI experiment break down and reveal its mechanical workings under extended scrutiny, as has historically been the case, or will it present a convincing illusion of human-style intelligence at work?

The practical applications of such a technology are myriad, but game feasibility is, of course, a long way off.  But I can't help imagining something akin to the classic text adventure, where the hosting system understands not only everything the game designer could foresee, but is capable of comprehending a broader context.  Material properties, traditional object usage, even social conventions could be roughly delineated, leaving it to the system to fill in the gaps based on its own synthetic knowledge of the world, inside and outside of the game, employing what we could come to think of as an artificial imagination.

At a more realistic level, it appears we are approaching a machine that can produce novel responses to novel input -- and that would truly be something new under the sun.

No comments:

Post a Comment