IBM's 'Jeopardy!' match more than game playing
Big Blue sees Watson driving strides in analytics
- By Nick Wakeman
- Feb 10, 2011
The battle between man and machine opens a new chapter on Monday when an IBM computer named Watson takes on two "Jeopardy!" champions in a test of speed and trivia.
For the developers of Watson, named for the founder of the 100-year-old company, the "Jeopardy!" game is the culmination of a four-year project that had plenty of naysayers when it started.
The early knocks were that a computer could not understand the nuances of language such as puns, sarcasm and wordplay fast enough to match the human brain.
How did Watson fare on Day 1 of Jeopardy!
Other related story
IBM's Watson vs the human brain
Now many of those in the computer science world who said it couldn’t be done are now hailing Watson as a technological marvel.
But for IBM, Watson and the innovation behind it is much more than just playing a syndicated game show, said Dave McQueeney, vice president of IBM Research.
The company, which was founded in 1911 by Thomas J. Watson, hopes the success of Watson will drive more interest in computer science as a career.
“There is the potential to wake up a lot of young kids and reinvigorate interest in these technical fields,” McQueeney said.
Second, and probably most applicable to the government market, Watson will usher in a new era of analytics tools, he said.
Watson is powered by 10 racks of IBM Power 750 servers, running Linux. It uses 15 terabytes of RAM and 2,880 processor cores.
To prepare for "Jeopardy!", it has been loaded with a huge amount of information about books, movies, history, plays, music, current events, and the list goes on.
For each question, Watson evaluates information from about 200 million pages of content, or 1 million books, McQueeney said.
And it has to do it in 3 seconds.
To achieve that speed, Watson runs thousands of processes at once, using hundreds algorithms and Natural Language Processing and Statistical Machine Learning technologies. It also relies on Unstructured Information Management Architecture.
Unlike Watson’s chess playing cousin Deep Blue, Watson is built using technology that is commercially available today, McQueeney said.
Within a few years, Watson-like computers could be affordable for government agencies that need to analyze vast amounts of data, he said.
“There are these classes of problems where people want to reason over very large collections of data and pose questions,” McQueeney said.
Uses in government range from health care analytics to looking for patterns in eligibility decisions to being a tool for streamlining processes.
“It really is designed to collaborate with people,” he said.
To play "Jeopardy!", Watson breaks down the clue looking at things such as grammar, context and relationships and generates hundreds of possible interpretations. It then runs those against its databases and builds an evidence profile to determine what are the most likely correct answers. Each answer has a confidence level.
Viewers watching the "Jeopardy!" match will see the top three answers and their confidence levels.
For some questions, one answer will have a high confidence level. This is when Watson is most likely to buzz in. For other questions, none of the answers will have a high confidence level and Watson will not buzz in.
“That’s the interesting thing,” McQueeney said. “The machine knows when it doesn’t know the answer.”
In a real world application, this type of computing makes Watson a collaboration tool for humans who want to work with large sets of data. For example, in the medical field, Watson would not generate the diagnosis, McQueeney said.
“The machine’s performance and the human’s performance would amplify each other,” he said. “The human understands the subtly, the depth and the reasoning. The machine then brings a tremendous breadth of analysis but it understands something of the human reasoning.”
The doctor would pose a question and include the information such as medical history. The machine would come back with possible interpretations. The doctor would prune those and do more analysis and come back with possible answers and the confidence level for each answer.
The confidence level would include links to the data such as medical journals and other reports that the doctor could consult before making a diagnosis.
“There are so many things that people want to get out of these huge stores of data,” McQueeney said. “They know the reasoning, but they haven’t had the tools to handle all the data. That’s what Watson does.”
The "Jeopardy!" showdown between Watson and past champions Ken Jennings and Brad Rutter will air Feb. 14, 15 and 16, in a tournament of champions format. Two matches will be played over three days.
Any of Watson’s winnings will be donated to charity. Jennings and Rutter have agreed to donate half of their winnings to charity as well.
Nick Wakeman is the editor-in-chief of Washington Technology. Follow him on Twitter: @nick_wakeman.