[ExI] Watson On Jeopardy
eugen at leitl.org
Wed Feb 16 13:10:30 UTC 2011
On Wed, Feb 16, 2011 at 07:55:56AM -0400, Darren Greer wrote:
> Annihilated them Spike. App. 35,000 for Watson, 4000 for Jennings and 10000
> for Rutter. He got final Jeopardy wrong but was parsimonious with his
> wager -- just 900 odd dollars. Alex Trebek laughed and called him a
> 'sneak' because of the clever wager. The category was which U.S. city has
> an airport named after a war hero and a WWII battle. Watson said Toronto. I
> got a good laugh. I didn't know we'd been annexed. Another
> interesting detail. Ratings for Jeopardy have soared into the
Not to mention bring Big Blue back into the limelight.
> stratosphere because of Watson. It moved into the number two spot in TV
> land behind a Charlie Sheen sitcom last night.
By the way, Watson is not nearly as dumb (and far more usable) than I
thought. According to
February 09, 2011
Must See TV: IBM Watson Heads for Jeopardy Showdown
Michael Feldman, HPCwire Editor
Next week the IBM supercomputer known as "Watson" will take on two of the
most accomplished Jeopardy players of all time, Ken Jennings and Brad Rutter,
in a three-game match starting on February 14. If Watson manages to best the
humans, it will represent the most important advance in machine intelligence
since IBM's "Deep Blue" beat chess grandmaster Garry Kasparov in 1997. But
this time around, the company also plans to make a business case for the
technology. Trivial pursuit this is not.
And impressive technology it is. On the hardware side, Watson is comprised of
90 Power 750 servers, 16 TB of memory and 4 TB of disk storage, all housed in
a relatively compact ten racks. The 750 is IBM's elite Power7-based server
targeted for high-end enterprise analytics. (The Power 755 is geared toward
high performance technical computing and differs only marginally in CPU
speed, memory capacity, and storage options.) Although the enterprise version
can be ordered with 1 to 4 sockets of 6-core or 8-core Power7 chips, Watson
is maxed out with the 4-socket, 8-core configuration using the top bin 3.55
The 360 Power7 chips that make up Watson's brain represent IBM's best and
brightest processor technology. Each Power7 is capable of over 500 GB/second
of aggregate bandwidth, making it particularly adept at manipulating data at
high speeds. FLOPS-wise, a 3.55 GHz Power7 delivers 218 Linpack gigaflops.
For comparison, the POWER2 SC processor, which was the chip that powered
cyber-chessmaster Deep Blue, managed a paltry 0.48 gigaflops, with the whole
machine delivering a mere 11.4 Linpack gigaflops.
But FLOPS are not the real story here. Watson's question-answering software
presumably makes little use of floating-point number crunching. To deal with
the game scenario, the system had to be endowed with a rather advanced
version of natural language processing. But according to David Ferrucci,
principal investigator for the project, it goes far beyond language smarts.
The software system, called DeepQA, also incorporates machine learning,
knowledge representation, and deep analytics.
Even so, the whole application rests on first understanding the Jeopardy
clues, which, because they employ colloquialisms and often obscure
references, can be challenging even for humans. That's why this is such a
good test case for natural language processing. Ferrucci says the ability to
understand language is destined to become a very important aspect of
computers. "It has to be that way," he says. "We just cant imagine a future
But it's the analysis component that we associate with real "intelligence."
The approach here reflects the open domain nature of the problem. According
to Ferrucci, it wouldn't have made sense to simply construct a database
corresponding to possible Jeopardy clues. Such a model would have supported
only a small fraction of the possible topics available to Jeopardy. Rather
their approach was to use "as is" information sources -- encyclopedias,
dictionaries, thesauri, plays, books, etc. -- and make the correlations
The trick of course is to do all the processing in real-time. Contestants, at
least the successful ones, need to provide an answer in just a few seconds.
When the software was run on a lone 2.6 GHz CPU, it took around 2 hours to
process a typical Jeopardy clue -- not a very practical implementation. But
when they parallelized the algorithms across the 2,880-core Watson, they were
able to cut the processing time from a couple of hours to between 2 and 6
Even at that, Watson doesn't just spit out the answers. It forms hypotheses
based on the evidence it finds and scores them at various confidence levels.
Watson is programmed not to buzz in until it reaches a confidence of at least
50 percent, although this parameter can be self-adjusted depending on the
To accomplish all this, DeepQA employs an ensemble of algorithms -- about a
million lines of code --- to gather and score the evidence. These include
temporal reasoning algorithms to correlate times with events, statistical
paraphrasing algorithms to evaluate semantic context, and geospatial
reasoning to correlate locations.
It can also dynamically form associations, both in training and at game time,
to connect disparate ideas. For example it can learn that inventors can
patent information or that officials can submit resignations. Watson also
shifts the weight it assigns to different algorithms based on which ones are
delivering the more accurate correlations. This aspect of machine learning
allows Watson to get "smarter" the more it plays the game.
The DeepQA programmers have also been refining the algorithms themselves over
the past several years. In 2007, Watson could only answer a small fraction of
Jeopardy clues with reasonable confidence and even at that, was only correct
47 percent of the time. When forced to answer the majority of the clues, like
a grand champion would, it could only answer 15 percent correctly. By IBM's
own admission, Watson was playing "terrible." The highest performing Jeopardy
grand champions, like Jennings and Rutter, typically buzz in on 70 to 80
percent of the entries and give the correct answer 85 to 95 percent of time.
By 2010 Watson started playing at that level. Ferrucci says that while the
system can't buzz in on every question, it can now answer the vast majority
of them in competitive time. "We can compete with grand champions in terms of
precision, in terms of confidence, and in terms of speed," he says.
In dozens of practice rounds against former Jeopardy champs, the computer was
beating the humans with a 65 percent win rate. Watson also prevailed in a
15-question round against Jennings and Rutter in early January of this year.
See the performance below.
None of this is a guarantee that Watson will prevail next week. But even if
the machine just makes a decent showing, IBM will have pulled off quite
possibly the best product placement in television history. Open domain
question answering is not only one of the Holy Grails of artificial
intelligence but has enormous potential for commercial applications. In areas
as disparate as healthcare, tech support, business intelligence, security and
finance, this type of platform could change those businesses irrevocably.
John Kelly, senior vice president and director of IBM Research, boasts,
"We're going to revolutionize industries at a level that has never been done
In the case of healthcare, it's not a huge leap to imagine "expert" question
answering systems helping doctors with medical diagnosis. A differential
diagnosis is not much different from what Watson does when it analyzes a
Jeopardy clue. Before it replaces Dr. House, though, the machine will have to
prove itself in the game show arena.
If Jennings and Rutter defeat the supercomputer this time around, IBM will
almost certainly ask for a rematch, as it did when Deep Blue initially lost
its first chess match with Kasparov in 1996. The engineers will keep stroking
the code and retraining the computer until Watson is truly unbeatable.
Eventually the machine will prevail.
More information about the extropy-chat