After 'Jeopardy!' Win, Watson Goes to Med School

In February, IBM's Watson computer system made mincemeat out of two all-time Jeopardy! champions, racking up more than $77,000 in the two-day challenge — its nearest human competitor earned just $24,000. The next logical question is, "Where does Watson go from here?" Some of us would say "Wheel of Fortune." IBM says medical school.

This isn't the first time IBM has used games as a challenge to advance computing prowess. In 1997, the company's Deep Blue system defeated world chess champion Garry Kasparov in a six-game match. But chess is a game that is largely mathematical, playing to the computer's strength. Jeopardy is much more verbal, with clues and categories filled with puns and subtle references. And the body of knowledge required to succeed at Jeopardy! is enormous — everything from the Kardashians to the Caspian Sea. How did the computer cope?

The first hurdle is simply understanding what information the question is asking for. Many of us remember being pleasantly surprised when a computer game could understand typed-in commands like "Go north" and "Take the axe." Thirty years later, Watson was taught to understand complex sentence constructions and vocabulary, making sense of "The 'Ancient Lion of Nimrud' went missing from this city's national museum in 2003 (along with a lot of other stuff)" and correctly answering "Baghdad."

Another major hurdle was "understanding" a vast database of information. Simply storing millions of individual facts isn't the problem — it's making meaningful connections between them in a matter of seconds. For this, IBM refined a technique called "machine learning." The traditional approach to building a computer expert system is for humans to create a set of rules for analyzing data, which just isn't practical with Watson's 10 million documents covering a vast array of topics. Instead, Watson analyzed hundreds of thousands of previous Jeopardy! questions, looking for patterns. This learning-by-example approach let the system create its own rules, which it then used to process the question at hand and tease out the most likely response from its database.

One more key technology is worth mentioning here: the ability to set a level of confidence in a given answer. In Jeopardy!, if you're unsure of a particular response, you don't want to buzz in, since you'll be penalized for a wrong answer. Watson takes this into account, coming up with a percentage of certainty and holding back from ringing if its confidence isn't high enough.