Re: Symbols, Images and Neural Nets

From: Harnad, Stevan (harnad@cogsci.soton.ac.uk)
Date: Sun Mar 09 1997 - 18:32:18 GMT


> From: Lowing, Ben <brl196@soton.ac.uk>
>
> Am I correct in remembering from lectures that neural nets aren't
> genuine constructs, but are programmed using algorithms. If this is so,
> wouldn't they suffer from the frame problem as well? ben

Not quite:

(1) A "real" neural net is really a bunch of units and connections
between them, etc. None of the nets you read about are really nets.
They are just nets simulated by computers.

(2) Computers can simulate just about anything; they can also be used
to test theories. So if you have a theory that a net with so many input
units and so many hidden units, and such-and-such a function for
changing connection weights, then a computer can test whether it can
really do what you say it can do. (By the same means it could simulate
and test a new rocket design; in each case all that's needed is a
simulation of whatever it is that you want to test, plus the input
conditions: for a learning neural net, it would be, say, an array of
pixels like the retina, and for a rocket it would be Cape Canaverel,
gravity, and the atmosphere.)

(3) Neural nets don't have a Symbol Grounding Problem or a Frame
Problem because they don't use symbols. A net simulated by a symbol
system still has no symbol grounding problem because the symbols are
only used to simulate the net rather than to simulate cognition.

Ask again if anyone doesn't understand this distinction between
a symbol system as a model for thought and a symbol system used
to simulate something else (like a net or a plane).



This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:51 GMT