Xref: utzoo comp.ai:3173 talk.religion.misc:10438
Path: utzoo!attcan!uunet!lll-winken!ncis.llnl.gov!helios.ee.lbl.gov!nosc!ucsd!orion.cf.uci.edu!uci-ics!venera.isi.edu!smoliar
From: smoliar@vaxa.isi.edu (Stephen Smoliar)
Newsgroups: comp.ai,talk.religion.misc
Subject: Re: Elementary AI Philosophy
Keywords: Understanding and Comprehension, Reality and Modeling
Message-ID: <7346@venera.isi.edu>
Date: 23 Jan 89 16:15:29 GMT
References: <18464@santra.UUCP> <1241@arctic.nprdc.arpa> <904@ubu.warwick.UUCP> <9423@ihlpb.ATT.COM> <43763@linus.UUCP>
Sender: news@venera.isi.edu
Reply-To: smoliar@vaxa.isi.edu.UUCP (Stephen Smoliar)
Organization: USC-Information Sciences Institute
Lines: 59

In article <43763@linus.UUCP> bwk@mbunix.mitre.org (Barry Kort) writes:
>
>There *can* be "something inside the computer that `understands'",
>but that something need not be thought of as a "homunculus".
>
>If we substitute the word "comprehend" for "understand", we have
>a better chance of seeing how a computer can have an idea of how
>things work out there in the real world.  The verb "to comprehend"
>means "to capture with".  In my mind, and in the mind of my
>computer, I construct models which replicate the structure and
>behavior of real-world objects.  I capture (comprehend) reality
>with such models.
>
>As to intelligence, that step follows easily after I have a working
>model.  I can now do "thought experiments" on the model to find
>out what will happen if I diddle the controls on the model, or
>if I perturb the operating environment in which the model is
>embedded.  I call this "cognition".  (Some people call it
>model-based reasoning, or modal logic.)
>
I'm not sure that things progress quite so "easily" as that.  Let
us assume, for the sake of argument, that you have "something inside
the computer" that can construct the sorts of models you have in mind.
It is not a simple step from having those models to performing thought
experiements on them.  What sorts of issues remain?

	1.  First of all, there is the problem that you are going
	to have to manage lots of models.  Thus, there are questions
	of storage management.  There are also questions of retrieval:
	How do you know what model to access and when to access it?

	2.  Next, there is the issue of manipulating those models.
	Metaphorically speaking, any model is going to have lots of
	"knobs" on it.  Observing a model's behavior is a matter of
	knowing how to twist which knobs when and what to look for.

	3.  This then brings up the issue of "thought experiements."
	Like laboratory experiments, thought experiments must be planned.
	Furthermore, those plans are designed in response to hypotheses
	to be investigated.  Thus, before we can talk about thought
	experiments, we have to talk about some agent which "thinks
	them up."  Thus, while it may be true that you do not need
	a homunculus to build your models (and I probably would be
	willing to contest THAT point, too, given more time to think
	about it), it would seem that your scenario ultimately depends
	on having a homunculus to manipulate them.

Is there any way to escape the homunculus?  I would argue that it can only be
escaped by rejecting the premise which fostered it:  the idea that
understanding can be attributed to a "something inside the computer."
Both Minsky and Edelman have proposed views of comprehension in which
understanding emerges from the interactions of components.  The important
point here is that no individual component may be said to embody understanding,
nor is it the case that all components are subjected to some kind of
"intelligent control."  It is simply that there is this behavior which
is a byproduct of interaction which manifests the characteristics of
understanding;  and because it is an emergent behvior rather than a
"thing," it makes no sense to try to isolate it in a single component
of a system.