Thursday, February 10, 2011

Rumors of the Death of Hal

As I watched the Nova episode introducing our new overlord,the Jeopardy-playing computer named Watson, I was reminded of an observation my AI professor once made. The history of artificial intelligence is crisscrossed with lines drawn in the sand declaring something like,"machines are not intelligent because only humans can do X" whether X is playing chess or playing Jeopardy. When at last IBM produces a computer that can accomplish the forbidden task, the nay-sayers look under the hood and say, "It isn't really displaying intelligence. It is just some clever programming, etc..." Then they grab a bit of drift wood and start drawing a new line.

I am here to declare that I have seen under the hood and am willing to admit, "It isn't really displaying intelligence. It is just some clever programming, etc..."

There has been a seeming breakthrough in the last few years in the realm of computer translation. The best known example of this is the translation feature at google.com. After decades of research in language parsing and artificial intelligence, has the Holy Grail of AI been achieved? Not at all. Some computer scientists decided that they could do translation without understanding the content of the text they are translating. All they had to do was gather a large amount of text with corresponding translations in other languages and smash them together until, through clever programming, they can map one series of sub-phrases in Ukrainian to a statistically similar series in English. Turns out this is good enough for government work and the scientists involved deserve kudos. But let us be honest, what has happened is that they have thrown hardware at the problem in much the same they calculate the value of pi to rediculous decimal places.

Enter Watson.

How does this clever bit of programming convert the sometimes satirical Jeopardy clues to answers-in-the-form-of-a-question? It smashes vast amounts of text together along with vast numbers of historical Jeopardy clues until it winnows the possible answers to the most likely. Oh yeah, and in order to save Watson's hide prior to its debut on national television, the developers added a subroutine to use the other, human, players' answers to help understand the categories. I realize that the human players use their opponents responses to narrow down possible answers but it seems that Watson is piggy-backing on the human facility with language to glean the crucial factor of context.

The creators of Watson are quick to point out that they are not claiming that the machine is intelligent in any way that humans would recognize. I agree that it is merely a highly sophisticated search engine. However, judging by the way the project leader takes offense at jokes made at Watson's expense, the urge to anthropomorphize a "language using" machine is irresistible.

Although it sometimes seems that artificial intelligence has been thrown over in favor of flashy tools that "play" chess or game shows, perhaps this is good for the field. While Murray Wiggle (I swear that one of the developers looked just like the red Wiggle) labors to get Watson a date with Alex Trebek, researchers in AI are recognizing that no matter how many games of chess they win, they must conquer the things a five year old has mastered - e.g. the motivations of others - before they can arrive at true machine intelligence.

No comments: