On Mon, Nov 14, 2005, Uri Even-Chen wrote about "Re: [off topic] A new project - automatic translation": > On the other hand, neural networks are not really intelligent. You > can't teach them to play chess, for example. They are not suitable for > everything.
This is getting wildly off-topic, but... Ray Kurzweil, in his book "The Age of Spiritual Machines: When Computers Exceed Human Intelligence", makes the following observation about human thought, and how computers can immitate it. If you'll allow me to put what I remember from his writing into my own words: He believes that human thought has two modes: "computation" and "patern recognition". Examples of the former include an arithmetic computation, or thinking about several options one after another, and an example of the latter includes face recognition. He believes that chess is an example where both modes are used: a chess expert, like a chess novice, goes in his head through many of the possible moves and his oponent's possible reactions (this is the computational mode), but unlike a novice, he also does some "patern recognition" on each of the resulting boards, and instantly (without sequential "computation") recognizes situations which are good, or bad, for him. This final recognition is the part of their thought-process that chess-players can't really explain, and is often called "intuition". Kurzweil argues that a chess-playing program could act similarly - walk the the (fantasically huge) tree of possible moves and counter-moves, sequentially, and at every junction apply a neural network that recognizes "good boards", and prune the tree at that junction if the neural network decides to that this move is not worth it. This technique, of walking huge trees with a *heuristic function*, are well known in AI (look up "A star", "Minimax", etc.) and are not Kurzweil's invention. But his interesting insight is that Neural Networks are useful but SHOULD NOT (not only "CAN NOT") be used directly to solve every problem, but rather should be combined with other computational techniques. It is arguable that similarly, neural networks should not be used directly to parse language. It is very possible that language understanding and generation has both a "computational", or sequential, aspect (reading the words one by one, following some sort of state machine in your head), and a pattern recognition aspect. > If you're interested in what I did with neural networks - I used them to > compose music. If you want to see more details, look at "Speedy > Composer": http://www.speedy.co.il/composer/ The music is very nice :) -- Nadav Har'El | Monday, Nov 14 2005, 13 Heshvan 5766 [EMAIL PROTECTED] |----------------------------------------- Phone +972-523-790466, ICQ 13349191 |A computer without Microsoft is like a http://nadav.harel.org.il |chocolate cake without mustard. ================================================================= To unsubscribe, send mail to [EMAIL PROTECTED] with the word "unsubscribe" in the message body, e.g., run the command echo unsubscribe | mail [EMAIL PROTECTED]
