212
PARTISAN REVIEW
If
we look at the human brain, it works almost entirely by pattern
recognition. We don't have time to think very much in real time about new
situations. So, for example, if you look at chess master Garry Kasparov,
he doesn't have time to think in real time, "Well, if I go here, he's going to
go there, and then
I'll
go here, and he'll go there." He can only think of
maybe a few dozen or a few hundred of those situations in the time he has
to move. Deep Blue can actually think of three hundred million of those
kinds of board situations per second; so how does Kasparov have a
chance against a computer?
Well, he uses his pattern recognition to access a mental database of
situations he has thought about before, that is to say, during his entire
life. That's why it takes so long
to
become a chess master: he has mas–
tered a hundred thousand positions and thought about them in some
depth. So while he's playing, he sees a situation and his pattern recogni–
tion says, "Oh, this is like that board position that grand master so-and–
so had last year when he forgot to protect his trailing pawn. I better
protect my trailing pawn." I actually suggested
to
Murray Campbell,
who is the head of the Deep Blue team, that he combine this very fast
brute force method of thinking about move-counter-move situations
with a neural net that would actually be able
to
do pattern recognition,
and train that neural net on every master game of the century. He
thought that was a good idea and we set up an advisory panel to design
the project using this idea, but IBM canceled the project. But this is the
essence of human intelligence, and at least in this one field of artificial
intelligence, we use those techniques.
I have a project right now called Fat Kat-that stands for Financial
Accelerating Transactions from Kurzweil Adaptive Technologies–
which is applying neural nets and evolutionary algorithms
to
predicting
the stock market. We feed in all of the data, a lot more of it than a
human being could look at, and it's very chaotic data-there's a lot of
"noise" in that information.
If
you look at, let's say, commodity prices,
we say that the signal has a lot of "noise," which is to say there's a ran–
dom element which is unpredictable. But that doesn't mean it's all noise,
or all randomness; there is a signal amidst the noise and these powerful
mathematical techniques can find that signal and use it. We actually sim–
ulate evolution in the computer: we have a million little investors, with
their own software-based "DNA" that defines their rules for investing
money and we actually have them compete on real-market data. Then
we dispense with the ones that didn't exceed market averages and leave
the ones that did well, and have them procreate through a simulated
sexual reproduction where we have progeny that take a piece of genetic
