Today is the day computers took over…game shows. In a brief exhibition match in upstate New York, IBM’s pure language processing computer, Watson, faced the titans of Jeopardy: Ken Jennings and Brad Rutter. Watson won. Sure, it was a brief sparring session, and the real trials won’t air until February, but it looks like Watson has a good shot at becoming the world’s champion of trivia. Alex Trebek didn’t preside over this preliminary match, but he was on hand to watch Watson go toe to toe with some of the top human competitors from the history of Jeopardy. He joked that the computer would soon be on sale on eBay. Not quite, Trebek, but IBM engineers do think that Moore’s law will allow everyone to have their own Watson years from now. Considering how well the computer performed against Jennings and Rutter I’m sure the two of them will be among the first in line. Watch Watson’s stellar performance in the video below. I’ll take The “AIs” Have It for $2000, Alex.
After a brief opening session, the scores were Jennings $3400, Watson $4400, and Rutter $1200.
As mentioned in our previous discussion, Watson draws its information from a personal database, not the internet. It can answer most Jeopardy questions in about three seconds. To do so, Watson uses massive amounts of parallel computing power. Inside the large machine are racks of servers, over 2000 cores, with 15 terabytes of RAM, and about 80 teraflops of processing power. Yet all of this hardware is more or less “off the shelf”. What makes Watson really unique is the way it processes language. IBM developed the DeepQA project (of which Watson is a part) to be able to provide human-like answers to human-asked questions. That means it has to understand the ambiguities and intricacies of human speech – a medium of communication notorious for its acceptable mistakes and imprecision. Using its vast database of literature, scientific reports, and other documents, Watson develops ideas of how often words are associated with other words, and what meanings are extracted from those connections. Add in a few rules about how to best play Jeopardy, and you are most of the way towards building a computer that can defeat humans at their own game.
As you saw in the first video above, there are areas in which Watson does better than others. Or rather, there are areas where humans still outperform Watson. First, it only speaks English (though it appears to have knowledge of words that have been anglicized). The computer may also be more conservative than its competitors in choosing to buzz in early, and perhaps more shy in deciding when it has the right answer. Word play and association categories, such as filling in the blanks in book titles, were still clearly won by Ken Jennings. Probably because humans are pretty good at not fixating on the things they don’t know.
Yet even if Watson is only a great Jeopardy player, and not a perfect one, that hardly matters in the greater plans for its use. IBM executives claim it could read all of the world’s medical files in a few seconds. Watson could likely handle similar stores of data with the same speed. Think of what it would mean to have a computer with such a complete knowledge in a field. Expert doesn’t even come close to describing it. In ten or twenty years, when there’s a Watson in every smart phone (and home), we will all have access to medical, engineering, or legal opinions that in scope (if not quality) exceed anything a human could hope to produce. Applications in intelligence, military applications, and general security are equally profound. If you can provide the database, a computer like Watson can provide the unparalleled expertise. It’s an amazing future.
In the next few days, Watson will take on many other former Jeopardy champions, playing in several rounds of competitions. The event will be aired in a special program in February. Win or lose in that eventual showdown, today’s scrimmage showed us that Watson has the right stuff. Good luck to all the human contestants…you’re going to need it.
[image credit: EnGadget, CrunchGear]
[source: CrunchGear, EnGadget]
