I'm not quite sure why Richard would want to quote Harnad. Harnad's idea of
how the brain works depends on it first processing our immediate sensory
images as "iconic representations" - not 1m miles from Lakoff's image
schemas. He sees the brain as first developing some kind of horse graphics,
for the horses we see,
Then there is an additional and very confusing level of "categorical
representations" which pick out the "invariant features" of horses - and are
still nonsymbolic. But Harnad doesn't give any examples of what these
features are. They are necessary he claims to be able to distinguish between
horses and similar animals.
(If anyone has further light to shed here, I'd be v. interested).
And only after those two levels of processing does the brain come to
symbols - to "H-O-R-S-E" and "C-A-T" etc - although, of course, if you're
thinking evolutionarily, it's arguable that the brain doesn't actually need
these symbols at all -our ancestors survived happily without language.
So Harnad depicts symbols as not so much simply grounded as deeply rooted in
a tree of imagistic processing - and I'm not aware of any AGI-er using
imagistic processing (or have I got someone, like Ben, wrong?)
Richard:
Derek Zahn wrote:
Richard Loosemore:
> My god, Mark: I had to listen to people having a general discussion of
> "grounding" (the supposed them of that workshop) without a single
person
> showing the slightest sign that they had more than an amateur's
> perspective on what that concept actually means.
I was not at that workshop and am no expert on that topic, though I have
seen the word used in several different ways. Could you point at a book
or article that does explain the concept or at least use it heavily in a
correct way? I would like to improve my understanding of the meaning of
the "grounding" concept.
Note: sometimes written words do not convey intensions very well -- I
am not being sarcastic, I am asking for information to help improve the
quality of discussion that you have found lacking in the past.
I still think it is best to go back to Stevan Harnad's two main papers on
the topic. He originated the issue, then revisited it with some
frustration when people starting diverging it to mean anything under the
sun.
So:
http://users.ecs.soton.ac.uk/harnad/Papers/Harnad/harnad90.sgproblem.html
and
http://users.ecs.soton.ac.uk/harnad/Papers/Harnad/harnad93.cogsci.html
are both useful.
I do not completely concur with Harnad, but I certainly agree with him
that there is a real issue here.
However......
The core confusion about the SGP is so basic that you will find it
difficult to locate one source that explains it. Here it is in Harnad's
own words (from the second paper above):
"The goal of symbol grounding is not to guarantee uniqueness but to ensure
that the connection between the symbols and the objects they are
systematically interpretable as being about does not depend exclusively on
an interpretation projected onto the symbols by an interpreter outside the
system."
The crucial part is to guarantee that the meaning of the symbols does not
depend on interpreter-applied meanings. This is a subtle issue, because
the interpreter (i.e. the programmer or system designer) can insert their
own interpretations on the symbols in all sorts of ways. For example, they
can grab a symbol and label it "cat" (this being the most egregiouse
example of failure to ground), or they can stick parameters into all of
the symbols and insist that the parameter "means" something like the
"probability that this is true, or real". If the programmer does anything
to interpret the meaning of system components, then there is at least a
DANGER that the symbol system has been compromised, and is therefore not
grounded.
You see, when a programmer makes some kind of design choice, they very
often insert some *implicit* interpretation of what symbols mean. But
then, if that same programmer goes to the trouble of connecting that AGI
to some mechanisms that build and use symbols, then the build-and-use
mechanisms will also *implictly* impose a meaning on those symbols. under
almost all circumstances (and especially if there is ANY SUSPICION OF
COMPLEXITY IN THE SYSTEM), these two sets of implicit meanings will
diverge. There is simply no reason why they should stay in sync with one
another, so they don't. If there is any conflict, then the grounding of
the system has been compromised. Ideally, the programmer gets out of the
way completely and leaves it to the system to ground its own symbols.
(That, of course, almost never happens).
But now, what happens in practice when people talk about symbol grounding?
They usually take an extremely naive approach and assume that IF a system
has some kind of connection to the outside world THEN it must have
grounded symbols! This is crazy. The fact is that having an outside
connection is a good first step to getting grounded symbols, but it does
not even begin to address all the ways that the grounding can get
compromised. Yet, people who do not really understand the idea of
grounding, but know that it is a cool buzzword, tend to use the buzzword
as if it just meant "connecting your AGI to the outside world".
This certainly happens on this list, but it was also present in many of
the AGI 2006 papers.
At the 2006 workshop (whose theme was something like "Grounding symbols in
the real world") I became more and more frustrated to see that grounding
was being mentioned in this trivial way, and that nobody was stopping to
point out that this was just downright wrong. Remember, this was supposed
to be the *theme* of the workshop! How can that be the theme, and then
everyone (including the workshop convener) not understand that this usage
was trivial and worthless?
This idiotic situation went on and on until the penultimate session of the
workshop, at which point I remember that I stood up in the discssion
period just before the final coffee break and explained that we were not
using "grounding" in a sensible way. Since I had only a few moments to
talk I said that I was looking forward to the roundtable discussion after
the break, because the topic of that roundtable was "Symbol Grounding", so
we would have an opportunity to get down to some real meat and sort the
problem out.
Then, when we came back from the break, Ben Goertzel announced that the
roundtable on symbol grounding was cancelled, to make room for some other
discussion on a topic like "the future of AGI", or some such. I was
outraged by this. The subsequent discussion was a pathetic waste of time,
during which we just listened to a bunch of people making vacuous
speculations and jokes about artificial intelligence.
In the end, I decided that the reason this happened was that when the
workshop was being planned, the title was chosen in ignorance. That, in
fact, Ben never even intended to talk about the real issue of grounding
symbols, but just needed a plausible-sounding theme-buzzword, and so he
just intended the workshop to be about a meaningless concept like
connecting AGI systems to the real world.
I hope that clarifies the issue a little. I have also written about the
grounding issue on these lists, but I don't remember where those posts
are.
Richard Loosemore
-------------------------------------------
agi
Archives: http://www.listbox.com/member/archive/303/=now
RSS Feed: http://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
http://www.listbox.com/member/?&;
Powered by Listbox: http://www.listbox.com
-------------------------------------------
agi
Archives: http://www.listbox.com/member/archive/303/=now
RSS Feed: http://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4
Powered by Listbox: http://www.listbox.com
