> From: "Mason Corinne" <CJM395@psy.soton.ac.uk>
> Date: Mon, 20 May 1996 10:41:10 GMT
> I really don't get what Dan Dennett is talking about in the Reverse
> Engineering paper - are top down and bottom up different approaches
> within reverse engineering or is one forward and one reverse?
> Can you please illuminate further.
Good question: Yes, they are both parts of reverse engineering, but they
represent different theoretical approaches, and, as when people bore a
tunnel under a mountain, it's not always sure they'll meet in between.
Here are a few meanings of "bottom-up":
(1) Study how the brain works, and work your way up to how the mind
works from there.
(2) Study how the sensory and motor systems work first, and work your
way up to "higher" functions from there.
Here are a few meanings of top-down:
(1) Study how the mind works first, and work your way down to how the
brain works from there.
(2) Study how "higher" functions work first, and work your way down to
sensory and motor ones from there.
Neural nets seem more conducive to a bottom-up approach, because they
can start with sensory input. So is sensorimotor robotics, which models
input and output systems. Brain science itself is a bottom-up approach.
Learning systems tend to be bottom-up.
Symbol systems and computation in general are top-down approaches. They
work directly at the "knowledge" level, not worrying about where the
knowledge comes from or how it is connected with the world. Artificial
Intelligence is a top-down approach.
For a long time it was assumed that these approaches could be pursued
independently, and would eventually meet in the middle -- or that
perhaps they were independent "modules," with sensorimotor systems
co-existing with "knowledge systems," but with the functioning of one
not dependent on the functioning of the other.
The symbol grounding problem (which is the problem that there is really
no "meaning" in knowledge systems -- all the meaning is in the minds of
its interpreters) suggests that maybe, after all, the only way to get to
the top is the usual way: starting from the ground...
This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:41 GMT