The issue is: how might NNs effectively represent abstract knowledge?
With difficulty!
Okay, to put it in a less facetious-sounding way: It is worth bearing in mind that biological neural nets are _very bad_ at syntactic symbol manipulation; consider the mindboggling sophistication and computing power in a dolphin's brain, for example, and note that it is completely incapable of doing any such thing. Even humans aren't particularly good at it: our present slow, simple, crude computers can do things like symbolic differentiation millions of times faster and more accurately than we can.
The point being, we tend to try to answer "how" questions by looking for simple, efficient methods - but biology suggests (albeit doesn't prove) that the reason we can't see a simple, efficient way for NNs to handle syntactic knowledge is that there isn't one; that researchers trying to use NNs or the like for AGI may have to bite the bullet and look for complex, expensive solutions to this problem.
(My own reaction to this is the same as yours, incidentally: to go straight for symbolic mechanisms as fundamental components in the belief that this plays better to the strengths of digital hardware. That doesn't mean NNs can't succeed, but it does suggest that they'll have to hit this problem head-on and resign themselves to throwing a lot of resources at it, in somewhat the same way that we on the symbolic side of the fence will have to resign ourselves to throwing a lot of resources at problems like visual perception.)
To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
