Thursday, March 10, 2011

Watson in Jeopardy

From time to time a new computer, robot or piece of software comes by and impresses us so much that I think it cannot be real, must be a fake of some kind.
When Deep Blue faced Kasparov in a series of chess matches, winning 2-1, it is reported that after a particularly unexpected move, the human player stood up and accused the computer of being controlled by humans: a computer could never conceive such a creative move.
When I watched the recent Jeopardy match between the computer (robot? software?) Watson and two champions of the game, I had the same feeling. The first thing that impressed me was that the moderator of the game spoke so fast that the algorithm used to translate the speech into text had to be really great.
Second, translating speech to text is one thing, now how to place this into context? At certain moments the moderator would address Watson and ask questions while in other moments the moderator would talk to the other participants or even the audience, so how did Watson know he was being addressed? There were probably keywords that are used in this game, and the programmers that built Watson probably added these as reference, but by what I saw, in only one occasion Watson spoke out of time, when he incorrectly pronounced the detergent "Cheer" as "Cheering".
Third, the questions of this game are a challenge by itself. I noticed that the two human players took at least one attempt before they understood what the category of questions was about, while Watson didn't miss any of them. How could he "understand" something better than a human? I will reinforce the question: humans have this belief that we are better capable than anything else on Earth of understanding a situation, making sense of it. After all, we can watch tv, which is a projection of a 3D universe into a plane. We can watch a six year old drawing of car and understanding that it means a car. So how come Watson was capable catching the ideas of the questions so fast that the other players would look at each other with that "Ahan..." look?
Fourth, once the scope of the question was defined, how did Watson translate the meaning of the question to the answer? We can conceive that Watson has almost limitless storage capacity and a super-fast processing, but how do you determine that George Bush rhymes with "tush"? Does a computer know what rhyme is? Does Watson have stored somewhere the notion of rhyme, or maybe of sound?
I don't believe so, it appears to me too much, but perhaps I am wrong. What I believe, however is that perhaps if you are fast enough and have enough processing power, you can come up with solutions that were not built with logic reasoning, step by step, as we do. Maybe Watson is not a robot built to think like us, maybe he was not even built to think in first place, perhaps he's just like the room full of monkeys which eventually write down all the work from Shakespeare.
But this rises at least one question? Even if Watson is quick and powerful enough to create a huge number of different answers almost instantly, he still needs to parse those and choose only one, and this implies that he must be capable of comparing the answers and saying which ones are better than the others. Doesn't this imply that he has at least to understand the question?
This hypothetical approach would be similar to a person who cannot solve a differential equation but who can derive expressions, thus he can generate a random list of functions and test each of them as the possible solution for the problem. But again, how does Watson score his answers?
Of course other possibilities for what we saw in Jeopardy, for instance, it is possible that IBM had a thousand different Jeopardy experts playing in a network and they all submitted their answers within the 3 second limit, the answer most voted would then be chosen, but I don't believe this happened. The answer that Watson gave to a tennis term which was also used in races was so out of contest ("automobile races" instead of "rally") that it rose giggles from the participants and audience.
A second possibility is that Watson is connected to a Doomsday machine: every time Watson gave a wrong answer, the universe was destroyed (or at least Earth, or at least me and you reader). In universes where Watson got the right answer we would be left baffled by this wondrous machine. I first read this example of the ultimate computation in a book by Hans Moravek "Mind Children", and of course would only work if the theory of parallel universes is valid.

0 Comments:

Post a Comment

<< Home