IBM to pit supercomputer against Jeopardy! champs

[Dimmauk]

IBM to pit supercomputer against Jeopardy! champs
Christopher Null: The Working Guy

It's bad enough that I can barely compete with the geniuses that parade through the Jeopardy! set every evening: Now us mortals have to face the possibility of competing against a computer loaded with trivia, as IBM preps a special supercomputer that could take on human contestants on a special edition of the popular game show.

This is the latest step from IBM in pushing computer intelligence into the world of human intelligence. The latest battleground: Quiz show Jeopardy!, where sometimes easy, sometimes obscure, sometimes tricky trivia is asked of three contestants who have to be the first to ring in in order to get a shot to answer -- and always in the form of a question.

Beating humans to the buzzer and phrasing the question properly will be the least of the worries of Watson -- as the Jeopardy!-playing supercomputer has been named -- nor will it be having a memory for facts. Rather, Watson's big challenge will be in understanding the questions posed on the show in order to determine exactly what is being asked.

For a simple piece of trivia this might not be so tricky: In a category of "World Capitals" a question of "Canada" leaves little confusion for Watson (or human contestants) on how to respond. But a question like "Meaning 'not working properly,' it may date back to a character in the comic strip 'The Katzenjammer Kids.'"* requires a deep level of abstract thought. Watson may have a list of Katzenjammer Kids character names, but will it be able -- on the fly -- to examine that list and compare it to common slang terms to determine the right response? Will it even understand to do that?

And how will it respond to the wordplay clues in categories like "Before & After" or categories which ask for a response that rhymes?

Watson's engineers say they aren't sure either, but they want to try, the goal is to get computers interacting in "human terms," a challenge which began in earnest when IBM pitted a computer against chess champion Garry Kasparov in the mid-'90s. Human and computer both won various matches in those series (though Kasparov claimed cheating when he lost). Ironically, the Jeopardy! challenge presents a much tougher challenge from an engineering standpoint -- but one in which it would be far more effective to cheat. (Any number of quick wits could do well feeding Watson answers; the number of players who could challenge Kasparov at chess is much smaller.)

Now IBM and Jeopardy!'s producers are looking at ways to actually stage this event, who would play, and what exactly would be on screen during the game. I, for one, vote for a CGI avatar with a deep, Scottish accent.

http://i68.photobucket.com/albums/i3...Anus_Relat.png

What's the difference between you and a mallard with a cold? One's a sick duck and I can't remember how it ends, but your mother's a whore Trebek.

[Heian]

IBM to pit supercomputer against Jeopardy! champs
Christopher Null: The Working Guy

It's bad enough that I can barely compete with the geniuses that parade through the Jeopardy! set every evening: Now us mortals have to face the possibility of competing against a computer loaded with trivia, as IBM preps a special supercomputer that could take on human contestants on a special edition of the popular game show.

This is the latest step from IBM in pushing computer intelligence into the world of human intelligence. The latest battleground: Quiz show Jeopardy!, where sometimes easy, sometimes obscure, sometimes tricky trivia is asked of three contestants who have to be the first to ring in in order to get a shot to answer -- and always in the form of a question.

Beating humans to the buzzer and phrasing the question properly will be the least of the worries of Watson -- as the Jeopardy!-playing supercomputer has been named -- nor will it be having a memory for facts. Rather, Watson's big challenge will be in understanding the questions posed on the show in order to determine exactly what is being asked.

For a simple piece of trivia this might not be so tricky: In a category of "World Capitals" a question of "Canada" leaves little confusion for Watson (or human contestants) on how to respond. But a question like "Meaning 'not working properly,' it may date back to a character in the comic strip 'The Katzenjammer Kids.'"* requires a deep level of abstract thought. Watson may have a list of Katzenjammer Kids character names, but will it be able -- on the fly -- to examine that list and compare it to common slang terms to determine the right response? Will it even understand to do that?

And how will it respond to the wordplay clues in categories like "Before & After" or categories which ask for a response that rhymes?

Watson's engineers say they aren't sure either, but they want to try, the goal is to get computers interacting in "human terms," a challenge which began in earnest when IBM pitted a computer against chess champion Garry Kasparov in the mid-'90s. Human and computer both won various matches in those series (though Kasparov claimed cheating when he lost). Ironically, the Jeopardy! challenge presents a much tougher challenge from an engineering standpoint -- but one in which it would be far more effective to cheat. (Any number of quick wits could do well feeding Watson answers; the number of players who could challenge Kasparov at chess is much smaller.)

Now IBM and Jeopardy!'s producers are looking at ways to actually stage this event, who would play, and what exactly would be on screen during the game. I, for one, vote for a CGI avatar with a deep, Scottish accent.

http://i68.photobucket.com/albums/i3...Anus_Relat.png

What's the difference between you and a mallard with a cold? One's a sick duck and I can't remember how it ends, but your mother's a whore Trebek.

Dear god, if they give it Sean Conroy's voice, the universe may implode due to the sheer awesomeness.

[aurik]

Maybe in 10, 20 years. State of computers is in no way advanced enough to do this yet.

[Darus Grey]

You'd be surprised...this happens to be my exact area of expertise, and we're a lot closer to building fully semantic computational models than ever before, we now have working models for concept spheres that have ~97% accuracy developed at University of Toronto, where just last year the best we had was ~44% with an older method.

This means a computer, with a very low amount of power, can now very accurately differentiate between things like "Star(celebrity)" and "Star(celestial object)".

I'm working on a project to assist in metaphor mapping, so computers can accurately determine the source and target domain of common metaphors(I'm in trouble, Do you see the problem?, He was full of life, etc) so that it can more accurately determine the intended meaning of non-literal(the majority of speech) phrases.

We've only seen the first waves of computational linguists graduating with PHDs in the last few years, it's a very new field that's exploding. Traditionally you only had programmers with little to no language training working on these kinds of projects, but now we're finally getting people who are specifically trained in the field and advances are coming that much quicker.

Maybe a bit optomistic, but I'd be surprised if we weren't there in 5 years, it isn't going to take 20...we're nearing the finish line to building understanding learning machines.

[Khamsin]

This means a computer, with a very low amount of power, can now very accurately differentiate between things like "Star(celebrity)" and "Star(celestial object)".

What about star stars, like the dog star or the north star?

[Darus Grey]

ug, got a forum error and ate my long detailed reply >.> I'm lazy so TL : DR version.

North star being a proper noun makes it very easy to detect it in proper context.

[Heian]

Make it go against the blind guy from a couple of years ago, he was nuts. He broke all the records at the time.

[Kyten]

I wanna see how this ends up.

[Kryssan]

It's honestly not that far away. There was a news article recently detailing how a computer discovered new scientific knowledge independently of it's creators.



But personally, I think the first AI should have HAL's voice. Oh c'mon, you know it'd be awesome.


I'm sorry Dave...

[Null]

I like that his name is Null.

[Khamsin]

Make it go against the blind guy from a couple of years ago, he was nuts. He broke all the records at the time.

Even on the video clues?

As for voice, I think it's voice should change depending on the subject. For example, GlaDOS for Science.