Watson / Jeopardy 'robot', whats so cool?

...

To entertain engineers.

(...)

That is what is so cool and mostly the point.

Perhaps you have been totally nonplussed by a question put to you outside of your current 'context space'? I do that two or three times a day, (but I'm old).

Note to self: She asked 'Soup or Salad?' not 'SuperSalad?'

--Winston

There was an interesting program on PBS's NOVA on this project.

Presumably, the same as the point of Deep Blue, that kicked that chess champion's ass. I saw that clip. The chess master (was it Kaspaov?) was heard to say, "Well, at least it didn't enjoy it."

That was no more allowed that it would have been for a human contestant.

They even rigged up a mechanical hand to press the same button as the humans.

And it's about to win a million dollars, which IBM is going to donate to charity.

And it's just WAY KEWL!

Hope This Helps! Rich

The point is that it took a lot of arm waving and head scratching to get a machine to make the same logical/illogical word associations that a human does - and learn from its mistakes, deal with puns, associate disparate clues, etc..

It is way more complicated than playing chess where there is a very limited list of rules and a more finite number of possibilities.

Yeah Google might help, but then it wouldn't be artificial intelligence, just a fancy text to speech engine.

The project is motivation for the IBM researchers to push the envelop in natural language processing. The show just makes it fun, while at the same time, acting as PR for IBM.

It has no speech recognition. The questions are fed to it as text files. That was explained at the beginning of the first show. It seems to be very good text to speech like you say however. But even my GPS is pretty good at reading the names I type into it.

Neither Google or wolfram alpha even come close to being able to answer Jeopardy questions. Wolfram alpha is still pitiful and mostly useless - it only looks "good" in a canned demo. Google is a fantastic tool to allow humans to find relevant web pages from keywords, but it doesn't have the level of context based understanding that Watson seems to have. That is, Watson is very good at correctly identifying what subject the question is about from the limited clues that are in typical Jeopardy questions, and can correctly deduce what type of answer is correct, as well as what answer is correct.

In addition, Watson is fast. It was built to win the button pushing race that's a key part of the game. And it doesn't just push the button and then hope it will get the answer before it's asked for. It comes up with the answers before it decides to push the button, and will only push the button, if it's statistical probability is right. But yet, it is still winning the button race most the time. It's basically making the human champions look like first graders.

Yeah, that's the question. How well can the algorithms that they used to make it good at Jeopardy questions be applied to other areas. The IBM video they ran on last nights show implied they would be able to put it to use in the medical field to allow doctors to quickly find relevant answers to diagnoses questions.

The question remains, how much English language facts ended up being hard coded into the Watson to help prevent it form making stupid mistakes and how much of it's knowledge base was learned just by scanning (reading) documents on its own?

Well, I assume they wouldn't have actually put together the match unless they had gotten it to the point that they expected it would be able to win. So my bet was on Watson as well. At the end of day one, Watson was not in the lead. He was tied with Brad. But now, at the end of day two, Watson is way out in front. Maybe something like $36K in winnings vs $12K for the next closest guy? The other guys spend most the show watching Watson answer questions.

But Watson does make mistakes. Most the points the other guys got was from when Watson wasn't sure of the answer, and didn't buzz in at all. They are showing a graphic at the bottom of the screen with Watson's three top answers and a bar graph showing Watson's estimated probability of the answer being right. It also has a vertical line showing the buzz-in limit. If the top answer is not past that line, Watson won't push the button.

Most the time, Watson's top answer is above 90% and most the time it's right. And when it's wrong, it's often a much lower probability. But a few of Watson's wrong answers were way up in the high 90% range - so it was "sure" it was right, even though it was dead wrong. I think that happened maybe once on the first night and once on the second night.

It also made a stupid mistake which it was obviously not programed to avoid on the first night. Ken won the buzz-in, and produced a wrong answer. Watson then got to try next, and answered the same as Ken - Watson obviously had no way of knowing what Ken had answered so they both made the same mistake.

It is somewhat amazing to me how much knowledge it seems to have, and how small the server farm is. It's something like 15 racks of 10 servers each with no connections to the internet or other systems. But yet, this small set of servers seems to have encyclopedic knowledge of general human affairs and has it coded in a way that it can extract highly relevant answers to the Jeopardy style questions.

I've seen nothing about where Watson get's it's information from? Does it digest normal written text documents on it's own (I believe so - but I'm not sure). And did they let it loose on the internet, or did they hand feed it high quality information documents - such as maybe wikipedia?

However, Jeopardy questions are normally answered with a single word. So it's clearly a very specialized type of information extraction. Finding the single word that best fits the clues, is not the same thing as being able to produce a language answer to a complex question such as "explain to me how you work Watson.".

With Google being a part of our lives for years, I have to admit that you are right that this next "milestone" in machine vs computer is not nearly as interesting a challenge as chess was. Even though Google is not tuned to answer Jeopardy questions, you do have to feel Watson is not doing anything that much more special than what Google does. Google ranks web pages from clues instead of ranking words, but otherwise, the technology is no doubt very similar.

I think that Jeopardy!, at least partially, jumped the shark when they let winners go on indefinitely, rather than retiring them after five consecutive wins.

Watching the human "machine" Ken Jennings performing in those dozens of wins became boring. It's much more fun to watch the program to see what you might know and the contestants do not, and vice versa.

Jeopardy! probably achieved higher ratings from attracting and holding the yahoo's that said to themselves "Gaw-lee, he's a right agin!"

Now it's a physical machine against at least one human "machine".

--- Joe

Evidently you're in the minority. It's ratings were soaring at the time.

I take it that you didn't watch it? (BTW, I didn't either)

Jeopardy! is on at the same time as the first half-hour of reruns of "Two-and-a-Half Men", Wheel of Fortune, on the same channel as Jeopardy!, is opposite the second half-hour of reruns of "Two-and-a-Half Men". Most of the time I watch only the second half-hour of "Two-and-a-Half Men".

This week I watch both half-hours of the reruns.

I hope old Charlie recovers and straightens out enough to give us more first-run episodes.

--- Joe

My point was that evidently the human "machine" sucked.

I really don't care about the woes of the self-absorbed Hollywood nut-baskets.

It is ratings?

THERE IS NO APOSTROPHE IN THE POSSESSIVE ITS!!!!!!!!

Thanks, Rich Grise, Self-Appointed Chief, Internet Apostrophe Police.

I don't see where *you* made that point. I did say it was boring.

My point: I'd rather have Charlie Sheen's comedy than boring, shark-jumping, yahoo-oriented episodes of Jeopardy! He does have to be somewhat healthy to make more episodes. Otherwise, I wish they wouldn't even report on the personal lives of any show-biz types.

--- Joe

Insufficient alternatives.

Don't listen.

I knew I screwed that pooch as soon as I hit send. Over-editing does it every time.

...

Its so annoy'ing when peoples' dont do the r'ight thing with aposrophsesses' itsnt it;

What were you doing?.... too buzy jerking off!

It was a simple comparison - not a list of alternatives.

You were the one who said "I really don't care about the woes of the self-absorbed Hollywood nut-baskets." I was only addressing your off the point remark. My mistake.

--- Joe

"Neither" is a better alternative, even if there aren't any.

Translated: I don't.

"Off point"? Wow! That's rich!

Hey Shawn, found a job yet?