Re: Neural Nets Vs. Symbol Systems

From: HARNAD Stevan (harnad@cogsci.soton.ac.uk)
Date: Tue Jun 04 1996 - 22:40:13 BST


> From: "Stanford, Graham" <gjs295@soton.ac.uk>
> Date: Fri, 24 May 1996 19:30:50 +0100 (BST)
>
> Neural nets are one possibility for how we might explain the mind
> through reverse engineering. Like the brain they have interconnected
> units, between which they pass activity, and the strength of these
> connections may vary. However they do not have axons, dendrites, action
> potentials or graded potentials, unlike the brain. Also, they are not
> well suited to reasoning and logical inference, language or
> calculation, in which respect they are obviously also unlike the
> brain.
>
> Symbols are arbitrary, basically they have no direct similarity to the
> object they depict, they represent it. A symbol system, examples of
> which could be mathematical formulae, recipes or even sentences, can be
> manipulated using algorithms. For instance one formula such as y =
> 91x-23 can easily be rearranged to be for x in terms of y. Likewise a
> sentence can be re-worded up to a point without changing its message or
> correctness.

It is symbols that are rulefully manipulated, not symbol systems; a
symbol system consists of the symbols and the symbol manipulation rules
(syntax; algorithms).

> The key difference is that nets cannot manipulate in this way because
> outputs are contained within its units. It cannot be manipulated to
> construct new orders or sentences, to do this it would have to learn
> the complete set of rules, for instance those governing algebra.

Not bad, but the last paragraph is a bit too cryptic for kid-sib: Give
examples, and. for an A, relate to the symbol grounding problem and
reverse engineering.



This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:45 GMT