Keith Sutherland:
.... But let me ask Pat one general question: why do you choose computational theory as the magic bullet to rescue a "bottom-up" theory of perceptual binding? What use is a computer without a program, and if you are using this as your model, it sounds pretty darn top-down to me, with the programmer as good a candidate for the homunculus controlling the system as any I can imagine.The programmer(s) helped create the system, along with the chip designers and thousands of other people. But once running, its the program which is influencing the behavior of the hardware, not the programmer. You can shoot the programmer and the machine carries on without a hiccup.
Theres a persistent misunderstanding lurking here: that a program is like a recipe which the computer 'obeys' blindly. Some are like that, but others arent. Some programs are more like populations which evolve, some work by performing inferences in organised ways, some do operations which look like shufflings and sortings of data, some lie in wait and pounce when patterns appear, some of them adapt and learn, some of them send themselves from place to place around the world leaving special trails in networks. They do all kinds of things in all kinds of machine architectures. ...
There have been many attempts to argue that contemporary models like neural networks and genetic algorithms operate along more "biological" principles. Again, I'm not qualified to comment, but would refer to Sam Salt's Tucson presentation (1996) where he pointed out that the differences are exaggerated -- all programming models ultimately reduce to a series of binary digits going through a Turing Machine.Well, replace 'Turing Machine' with 'hardware' (a Turing machine can't do anything, its just a mathematical abstraction), and I'll agree. But this is like saying that all brains ultimately reduce to masses and masses of quarks and leptons exchanging photons: true, but not very useful. This doesnt account for how programmed machines work, how they do what they do.
And just why is everyone getting so excited about neural networks? One doesn't need a very long memory to recall a similar euphoria for symbolic AI in the 50's and 60's. So why is it that the same people who rallied to Minsky's McCarthyite attacks on the Perceptron are now all of a sudden converts to the connectionist faith?They arent: its a new generation which has come along. Also, notice that MInksy and Papert's demolition of the Perceptron depended crucially on that machines having only two 'levels'. At the time it was generaly assumed that this wasnt really important, but we now know that it was. Three levels and you can do all kinds of wonderful things, quite escaping the M&P critiques. It took a good few years to realise this.
One is tempted to say that this particular road to Damascus was just the result of the realization that GOFAI was a blind alley.Well, this 'blind alley' still has an amazing amount of traffic on it. I wonder where you get this impression that GOFAI, as it is affectionatly known, has somehow died or gotten refuted? Let me invite you to the next AAAI or IJCAI or ECAI or IAAI meeting to get a better impression about what is going on.
Only time will tell whether the current euphoria for neural nets will have the same half-life. As Winograd (1991) states, "connectionism, like its parent cognitive theory, must be placed in the category of brash unproved hypotheses."Of course. All hypotheses in this area are brash and unproven. Great!
Pat Hayes
Phayes@uiuc.edu
jcs-online menu