|
|
Post by Pikachu on Feb 10, 2011 18:15:04 GMT -5
|
|
|
|
Post by buhwhyen on Feb 17, 2011 4:18:51 GMT -5
IBM challenge video. Basically the continuation (successor) of Deep Blue is Watson; a supercomputer that, based on probabilities, claims to be able to be faster and more accurate at answering any question, ie Jeopardy. But the most interesting part of their claim was that the computer could interpret English and formulate answers based off of just data held in it's own harddrive, it had no outside resources to pool (ie the internet). It also had to be able to push the plunger to answer the question as well as be able to say it's own answers aloud. The one drawback to the design is that it doesn't have a voice recognition system, so the questions have to be typed to be interpreted. The big draw back of not having voice recognition is that it can't benefit from hearing the wrong answers of other players, which is seen in the day 1 video, but aside from that...lawl technology. This is particularly interesting to me because this is one field of study that I've been doing a decent amount in, and currently taking a (capstone-esque) class where I'm doing a bunch of the programming to do similar probability classifications. There was also a test game staged a month? ago that you can probably search for on youtube as well. www.youtube.com/user/TheArcticEcho#p/c/29F1B52DC6B443D3 |
|
|
|
Post by dyuman on Feb 17, 2011 21:15:08 GMT -5
skynet kthnx.
|
|
|
|
Post by Pikachu on Feb 17, 2011 21:32:28 GMT -5
Yeah, NPR has been doing some coverage on this. Ken Jennings was going around promoting the show and talking about what it's like to compete against it.
The repetition of Ken's incorrect answer was VERY interesting actually, because figuring out why and how the computer would come to an incorrect conclusion would probably lead to some new insight (either linguistically or logically).
Far from Skynet, though. Advanced Turing box is not sentience... not yet at least.
|
|
|
|
Post by buhwhyen on Feb 19, 2011 11:54:23 GMT -5
TLDR;
Watson is a hop, skip, and maybe a jump away from really cool star trek-esque computers. Sentient? It depends how you define sentience. It can operate on it's own and while it doesn't necessarily innovate or "have it's own mind" it operates and makes it own decisions based on inputs. The only significant difference I see is that it requires an input, more training data = more able to handle new situations -> not just a Jeopardy answer box. Watson is actually a bigger milestone in computer technology than getting to the moon is for space exploration. It's a very real possibility we will now have to start thinking seriously about limiting computer abilities rather than not in our lifetimes.
****************************************
To be fair. NPR's coverage on TJ Watson was pretty lacking. Its really just quotes Ferrucci, IBM person. The way they quoted him actually is very misleading. It...kinda makes sense if you don't actually have background or knowledge about machine learning, but you can't really get a good idea of what anything really means.
In contrast to what Josh said, I'm actually not at all surprised by Watson's repetition of Ken Jenning's answer. And honestly, I don't think it can give any significant or even moderate insight.
Why? Because Watson isn't actually "interpreting" language in the sense that we would think of "interpretation." To be more precise, Watson, and other computers for that matter, are completely unable to "interpret" language. They can do calculations, estimates, and comparisons to come to conclusions about what is most probable...but aside from that they're pretty limited.
A simplified overview of the process is something as follows:
-you have stored data
-you receive an input (question)
-you have to find key words/phrases to search for
-you make a big list of possible definitions of words and words related to your key words/phrases
-search data for related information to key words/phrases and related terms
-you calculate tons of probabilities based on known data and the input
-Watson lists the 3 most probable answers
A more detailed overview is as follows:
-you have data that is stored among however many servers Watson holds.
-Watson receives a query/question/input.
-The input/question or whatever you call it, has to be classified. This is actually the most important step, and realistically can't really be done in just this step alone, it's an iterative process that is constantly adjusted based on results of the searches that follow. Wrong classification = definite wrong results.
-To classify the question, the first step is to break the question into keywords, or phrases. Using these key words/phrases, you can both narrow your search field, populate a list of potential definitions and related terms.
-In a crude sense, related terms are stored similarly to how a "linked list" is in C. Each "node" or "term" has a bunch of associated words that range from words that share the same meaning to words that are commonly used together.
IE. Bat can refer to the animal or object used to hit things. A related term might be cave, nocturnal, etc.
-After getting both potential definitions and related terms of the input, you can calculate a number of probabilities based on your known values and potential values (potential value is basically a random variable - if you speak in probabilities. Knowns are just classical probabilities)
This step is where you will get probably about 50% of your error (the other 50% should be split between mis-classification of the problem class and illogical links of data points).
-output answer (the fact Watson lists the 3 most probable answers is really just for the programmers to know how far off the probability calculations were if the most probably answer was wrong).
So to fully answer why I'm not surprised by Watson's answer...If you were to search for both parts of the question (in google) the answer would clearly point to the 1910s. So why didn't Watson return the same results as google search engine? That's due to 2 things, first search engines such as google use word counts as the main tool to determine which webpages to display (Watson does not do this, obviously). Secondly, it is highly likely that the terms (and related terms) used in the question, such as "modern," "crossword," "puzzle," etc were more often related to the 1920s than the 1910s. And as such, the probability that the correct answer was the 1920s. As mentioned earlier, the big limitation of interpreting language is that it really can't be done in the same way human do. Computers have to interpret information based on narrowing down potential answers and then calculating the most probably answer based on given known information (conditional probability). Watson's answer is just an erroneous calculation (similar to how humans remember stuff incorrectly from time to time), stemming form either misclassification of the problem or incorrect association b/c more related terms to insignificant keyword/phrases in the input. Error rates will never be 0%, I would be more afraid of the under 60 second search and calculation time more than I would one wrong answer.
Also keep in mind that Watson is both blind and deaf. It cannot see or hear, but it can speak and push the plunger to answer questions.
**********************
random shit
-Watson being able to answer Jeopardy questions is actually a much bigger milestone than I think most people really think. It's NOT NOT NOT about being able to be an answer box. Given "x" input, spit out "y" output. It's more that we've found a way to properly classify and search for information based on an input text string without restriction. Jeopardy questions come with many various references/puns/etc that have to be interpreted, which was previously the "wall that you had to climb" so to speak. Watson basically has scaled that wall almost completely, and just 4 years ago we were barely clawing at the bottom of it.
-We've absolutely crushed the major barrier we had previously..so what's the difference between Watson and Skynet? Skynet's sentience, I define to just be independent thinking and decision making, was actually showcased pretty convincingly in the Jeopardy game, since it operated on it's own and was able to assess the situation where there was only 1 question left in each category, last question of the round, daily double, final Jeopardy, when it was taking a guess, etc. These can mostly be programmed and say, well if you're above/below "this" threshold value then do "this other" command. But that's really not a convincing argument because if you just give it a larger number of examples to pull from (training data), then it can infer an answer on it's own with stupidly good accuracy. Once computers stop requiring inputs to make decisions and do operations (being prompted to do "x"), I will shit myself.
-The one major component missing is the ability to process sound and interpret that, but to be honest, the technology already exists, see voice commands in just about every cellphone/pda/car now days. Merging the 2 is really not a big deal at all.
-The error rate for wrong answers was astoundingly low. I didn't count the actual numbers of wrong answers, but it was probably under 10% overall (I include questions when he didn't buzz in as well). And 10% error is actually pretty damn solid for probability classification. Most classification techniques are heavily dependent on using an appropriate model based on given data, but I'm not aware that it is possible to model language in any way that would be beneficial to a computer in this fashion.
-In actual classification, each data point is assigned a "label," this label tells you what class the data point belongs. By doing estimates and probabilistic calculations, you can compare your unknown values (inputs) to known values (usually referred to as training data, but in the NPR article refers to them as "Jeopardy clues plus the correct solutions").
|
|
|
|
Post by Pikachu on Mar 2, 2011 18:33:24 GMT -5
For those of us in class of '01 and are not terribly Facebook active, we've all been added to a group for the 10 yr. reunion. Let me state for the record that you're all old and I am not.
|
|
|
|
Post by Pikachu on Mar 9, 2011 15:45:13 GMT -5
|
|
|
|
Post by feedback on Mar 10, 2011 2:21:29 GMT -5
|
|
|
|
Post by dyuman on Mar 16, 2011 10:03:34 GMT -5
Pretty good. I went 10-9-11 on it
|
|
|
|
Post by feedback on Mar 20, 2011 14:20:39 GMT -5
|
|
|
|
Post by Pikachu on Apr 1, 2011 2:02:34 GMT -5
shoryuken.com/content/interview-w-broly-legs-disabled-ssf4-chun-li-player-4088/Two players from SoCal, Gootecks and Mike Ross, put together a show called CrossCounterTV, where they talk about some stuff going on with fighters, stream shows of themselves playing online, and interview players in the scene. This time, they interview a guy who basically can't walk and can only use one hand, yet plays SF4 and Smash. You guys absolutely need to check this out.
|
|
|
|
Post by buhwhyen on Jun 2, 2011 15:53:34 GMT -5
For those who have been living under a rock, IPV6 day is June 8th. The world will not end (lawl at people who thought it would end 2? weeks ago). It's just a global testing day for IPV6 stuff mostly for larger companies (see google). Hooray progress.
|
|
|
|
Post by Pikachu on Jun 5, 2011 21:32:45 GMT -5
Apologies to anyone to whom I told SSF4AE was coming out in another week and a half. It's coming out June 7, as DLC for like... $15? Fun times to be playing Capcom games.
|
|
|
|
Post by Pikachu on Jun 16, 2011 0:38:09 GMT -5
www.youtube.com/user/crosscountertvThe current video clip you should see (as of the posting date) is Gootecks and Mike Ross going around the Capcom booth in E3 looking at two things: Street Fighter X Tekken and SF3 Third Strike Online Edition. SF X Tekken looks really interesting actually. It's almost like a calm Guilty Gear with a Marvel touch. 6 buttons as well. I have really high hopes for this game. By now everyone knows how I feel about Third Strike. On the other hand, the extra features in Third Strike are incredible (which is an understatement). There are so many that it's tl;dr to list in a post, but watch the video whether you intend to purchase this or not... because Capcom is using this as an experiment to see what the gamers' reactions will be to these features. All I can say is that if they do a CvS2 for PS3 with all these features, I will cry tears of happiness.
|
|
|
|
Post by buhwhyen on Aug 7, 2011 9:55:35 GMT -5
|
|
