PromoTheRobot Posted February 14, 2011 Posted February 14, 2011 Long article but something to ponder. PTR
stevestojan Posted February 15, 2011 Posted February 15, 2011 This was one of the coolest half hours of TV I've seen in a long, long time. The computer, Watson, is not connected to the internet. He gets the clues via text document sent to him. To see a minor explanation of how he deciphers the words, and figures out WHAT is being asked for, was pretty incredible. If you can HULU tonight's episode, or if you're on the west coast and it hasn't come on yet, I'd highly recommend it. One part shows the "progression" of Watson, in a practice session where a clue was asking for a "non dairy creamer" and watson said "Milk" Then they showed how he could take a question similar and really "figure out" what the clue was getting at. Cool, and terrifying all at once.
eSJayDee Posted February 15, 2011 Posted February 15, 2011 Thanks for the explanation Steve. Receiving the question as ASCII DRAMATICALLY simplifies the task of the computer. This, in my mind, makes for an unfair completion. I read that the computer handily won a "scrimmage" a month ago. I assume it faired equally well tonite? Still, it's pretty darn impressive that it can do this. I'd be curious as to how the computer searches for the "answer". I assume does it basically just scans a huge database (obviously w/ some sort of indexing/hashing involved).
Just Jack Posted February 15, 2011 Posted February 15, 2011 Thanks for the explanation Steve. Receiving the question as ASCII DRAMATICALLY simplifies the task of the computer. This, in my mind, makes for an unfair completion. I read that the computer handily won a "scrimmage" a month ago. I assume it faired equally well tonite? Still, it's pretty darn impressive that it can do this. I'd be curious as to how the computer searches for the "answer". I assume does it basically just scans a huge database (obviously w/ some sort of indexing/hashing involved). Your questions are basically answered in the show tonight. I can't remember specifics so I'm not even going to try.
stevestojan Posted February 15, 2011 Posted February 15, 2011 Thanks for the explanation Steve. Receiving the question as ASCII DRAMATICALLY simplifies the task of the computer. This, in my mind, makes for an unfair completion. I read that the computer handily won a "scrimmage" a month ago. I assume it faired equally well tonite? Still, it's pretty darn impressive that it can do this. I'd be curious as to how the computer searches for the "answer". I assume does it basically just scans a huge database (obviously w/ some sort of indexing/hashing involved). The score is $5000, $5000, $2000 currently (computer tied at $5000). Its going to take 3 shows to complete the game as there's background info they show. What amazed me is how it "knows" what the clue is even looking for. Sure, there are key words, but it could be SO MANY possibilities if you just look at the words. Anyway, take a look.
boyst Posted February 15, 2011 Posted February 15, 2011 (edited) NOVA did an amazing piece on Watson last week. I am trying to find the link to it, anyone know where I could find it? I'd like to share it. It breaks down piece by piece how the system works, etc. Easier then I thought, http://www.pbs.org/wgbh/nova/tech/smartest-machine-on-earth.html <--to view episode Edited February 15, 2011 by jboyst62
birdog1960 Posted February 15, 2011 Posted February 15, 2011 Receiving the question as ASCII DRAMATICALLY simplifies the task of the computer. i was disappointed that voice recognition wasn't used, mostly for selfish reasons...i'd love an electronic medical record system that i could dictate directly into accurately with the computer intelligently parsing the info. looks like that's still a ways off (even on the research level). that would have also made the machine appear much more human which might have been alarming. i agree that it seemed an unfair game with the machine getting the questions in a different form. but its reassuring that computers aren't ready to replace humans just yet.
The Senator Posted February 15, 2011 Posted February 15, 2011 Thanks for the explanation Steve. Receiving the question as ASCII DRAMATICALLY simplifies the task of the computer. This, in my mind, makes for an unfair completion. I read that the computer handily won a "scrimmage" a month ago. I assume it faired equally well tonite? Still, it's pretty darn impressive that it can do this. I'd be curious as to how the computer searches for the "answer". I assume does it basically just scans a huge database (obviously w/ some sort of indexing/hashing involved). Thank you. I was disappointed to learn that the questions are electronically entered as data to ' Watson ' rather than using voice-recognition technology - in my opinion, this renders the display of IBM's AI technology far less impressive. Apparently Watson uses a series of algorithms to ' decide ' which of the multiple choices in ' his ' huge database is the ' best ' answer to the question at hand, and also some sort of ' level-of-confidence ' metric to determine whether to ' buzz-in ' - all cool stuff, but not as impressive as speaking your query to the Star Trek computer about the USS Enterprise, I think.
shrader Posted February 15, 2011 Posted February 15, 2011 I wonder if the computer can possibly screw up in final jeopardy when it has the extra time to give its answer. I'll also be curious to see how it determines what to bet going into that round or for daily doubles (did it get any of those yesterday?).
/dev/null Posted February 15, 2011 Posted February 15, 2011 I didn't see the episode but I did catch a clip. I'm calling Shenanigans because I recognize the voice. That's not really a computer, it's Professor Steven Hawking
Just Jack Posted February 15, 2011 Posted February 15, 2011 I'll also be curious to see how it determines what to bet going into that round or for daily doubles (did it get any of those yesterday?). It did get a daily double. If I remember right it was the second answer revealed, so it choose to wager $1000 and got it correct.
stevestojan Posted February 15, 2011 Posted February 15, 2011 I'm confused why lacking voice recognition is a downfall for you guys. I'm not arguing, just maybe need some clarification. Isn't voice recognition a fairly "old" technology? I mean, I remember Dragon Naturally Speaking was pretty accurate 5+ years ago. And that program cost $99 or so. I have an iphone app that can dictate what I say that costs $.99 this year. It's "fairly" accurate. I would think for a couple grand (and I'm guessing IBM spent hundreds of thousands if not millions on this project) they could get a dynamite voice recognition system. Since all players have to wait to buzz in until Alex is done reading, is this really an issue? Maybe someone can correct me here.
Just Jack Posted February 16, 2011 Posted February 16, 2011 I'm confused why lacking voice recognition is a downfall for you guys. I'm not arguing, just maybe need some clarification. I can see it both ways. First, by getting it via text, it can start searching immediatly for an answer. BUT, with voice rec, it would need to hear the whole answer first, before searching. BUT, the other contestants can see and read the answer while being read. So in my mind, it's almost a non-issue.
birdog1960 Posted February 16, 2011 Posted February 16, 2011 I'm confused why lacking voice recognition is a downfall for you guys. I'm not arguing, just maybe need some clarification. Isn't voice recognition a fairly "old" technology? I mean, I remember Dragon Naturally Speaking was pretty accurate 5+ years ago. And that program cost $99 or so. I have an iphone app that can dictate what I say that costs $.99 this year. It's "fairly" accurate. I would think for a couple grand (and I'm guessing IBM spent hundreds of thousands if not millions on this project) they could get a dynamite voice recognition system. Since all players have to wait to buzz in until Alex is done reading, is this really an issue? Maybe someone can correct me here. it's still way too slow and too inaccurate, at least in my hands. have you ever watched closed captions on tv? i assume they're still using people to do it as the errors are not systematic (and very prevalent). why aren't they using voice recognition (and if they are,it's doing a terrible job)? i think it's because it works too slowly. i have colleagues that dictate into dragon and i've watched them demonstrate it. even after the program learns their voice and cadence (which takes a significant time investment), they still have to talk very slowly and there's still about a 5% mistaken transcription rate. so one then needs to carefully proof the product with a fine tooth comb. my human transcriptionist makes much fewer mistakes and leaves blanks when she's occasionally unsure, making it much easier to find the problems. i dictate extremely quickly (at least 2X as fast as my dragon colleagues) and it still takes about 40 minutes per day. doubling that adds up pretty fast and that's not including the extra time for a very careful proof read. one would think fixing these issues would be trivial but apparently it's not. in the case of the show, i suspect the machine would be at a disadvantage because of these inaccuracies and lag time but maybe i'm wrong.
Mr_Blizzard Posted February 16, 2011 Posted February 16, 2011 I'm sorry Dave, I can't do that . . . (HAL in 2001 - A Space Odyssey)
ofiba Posted February 16, 2011 Posted February 16, 2011 I wonder if the computer can possibly screw up in final jeopardy when it has the extra time to give its answer. I'll also be curious to see how it determines what to bet going into that round or for daily doubles (did it get any of those yesterday?). You got your wish tonight. It answered Toronto for final Jeopardy and the topic was "US Cities". It wasn't very confident though and only wagered 900 bucks.
stevewin Posted February 16, 2011 Posted February 16, 2011 (edited) I wonder if the computer can possibly screw up in final jeopardy when it has the extra time to give its answer. I'll also be curious to see how it determines what to bet going into that round or for daily doubles (did it get any of those yesterday?). Here's a link that explains how much it chooses to bet: Watson's wagering strategies Here's an explanation of why it missed so bad on final jeopardy last night: Toronto?? Finally - a lot has been discussed about the way the data is received - this article discusses that How Watson sees, hears and speaks Thought it was very interesting that Watson actually physically presses a button - with comment that humans can and do buzz in faster than his best possible reaction time. I think people who are so focused on the way the question is input are really missing the whole point. The technological breakthrough is what is done AFTER the question is received - the extent of comprehending the meaning/nuance of complex/free-form language using AI. It is absolutely NOT just a search through a big database - it is MUCH MUCH more complicated than that. Stevewin (IBMer - 25 yrs) Edited February 16, 2011 by stevewin
The Senator Posted February 16, 2011 Posted February 16, 2011 Here's an explanation of why it missed so bad on final jeopardy last night: Toronto?? I'm sure the IBM'ers could throw in a few KLoCs to give him a very nice bedside manner too, but I assume ' Watson ' could screw-up just as badly on a life-or-death medical query, or make an equally catastrophic financial decision should it be similarly ' confused ' ?
birdog1960 Posted February 16, 2011 Posted February 16, 2011 Here's a link that explains how much it chooses to bet: Watson's wagering strategies Here's an explanation of why it missed so bad on final jeopardy last night: Toronto?? Finally - a lot has been discussed about the way the data is received - this article discusses that How Watson sees, hears and speaks Thought it was very interesting that Watson actually physically presses a button - with comment that humans can and do buzz in faster than his best possible reaction time. I think people who are so focused on the way the question is input are really missing the whole point. The technological breakthrough is what is done AFTER the question is received - the extent of comprehending the meaning/nuance of complex/free-form language using AI. It is absolutely NOT just a search through a big database - it is MUCH MUCH more complicated than that. Stevewin (IBMer - 25 yrs) watson is an extemely impressive accomplishment. i certainly get that the nuances of language are being appreciated by the computer, albeit through text. but it would be that much more impressive (and useful) to have it done the way humans do it...through "hearing". after all, the "natural" and thus most efficient interface with humans would seem to be the spoken word. as you can tell, i've been frustrated with the pace of speech recognition technology improvements in the last 20 years. i'm sure your scientists are as well. i was hoping for a major breakthrough resulting from a project such as this. alas, science doesn't usually work that way.
stevewin Posted February 16, 2011 Posted February 16, 2011 watson is an extemely impressive accomplishment. i certainly get that the nuances of language are being appreciated by the computer, albeit through text. but it would be that much more impressive (and useful) to have it done the way humans do it...through "hearing". after all, the "natural" and thus most efficient interface with humans would seem to be the spoken word. as you can tell, i've been frustrated with the pace of speech recognition technology improvements in the last 20 years. i'm sure your scientists are as well. i was hoping for a major breakthrough resulting from a project such as this. alas, science doesn't usually work that way. Regardless of your desire for advances in speech recognition, that simply wasn't part of this particular project. The goal was not to take the spoken language and interpret it - it was to take language (however inputted) and interpret it as a (super smart, super fast) human would. Realize also - discussions of the potential application of this technology are focused more on mining/understanding thousands/millions of stored written documents, not interpreting the spoken word (as I said before - the real work/technology is applied AFTER the prompt is received).
Recommended Posts