Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!uunet!tdatirv!sarima
From: sarima@tdatirv.UUCP (Stanley Friesen)
Newsgroups: comp.ai
Subject: Re: What AI is exactly.
Message-ID: <152@tdatirv.UUCP>
Date: 21 Sep 90 16:30:10 GMT
References: <35282@eerie.acsu.Buffalo.EDU> <3851@se-sd.SanDiego.NCR.COM> <146@tdatirv.UUCP> <3893@se-sd.SanDiego.NCR.COM>
Reply-To: sarima@tdatirv.UUCP (Stanley Friesen)
Organization: Teradata Corp., Irvine
Lines: 142

In article <3893@se-sd.SanDiego.NCR.COM> jim@se-sd.SanDiego.NCR.COM (Jim Ruehlin, Cognitologist domesticus) writes:
>In article <146@tdatirv.UUCP> sarima@tdatirv.UUCP (I) write:
>>At a high level this is true, but at a basic neurological level it is not.
>>Neurons operate the same in humans and in buterflies, and at this level
>>learnng most certainly does take place in almost all known animals. 

>Neurons do, but what about cognitive structures?  No doubt we have different
>and more powerful neurally computational (call it mid-level cognition)
>abilities, probably due to some of the specialized neurons in the 
>cerbral cortex 
>We can make up all sorts of cognitive structures that they can't (as
>hypothesized by frames, scripts, etc. etc.).

In the brain concepts &c. are essentially patterns of activity in (very)
large sets of neurons.  In short the mind is a hierarchical structure, the
capabilities at each level are based on composition of lower level functions.
Concpetual learning involves many individual neurons learning to cooperate
in particular ways.  [Actually there are several levels between the neuron
and anything that we would call 'concepts']
 
>>[Tentative definition of learning]
>>	A change in neurological responses due to repeated stimulus
>>	that tends to alter behavior in a way that is responsive to
>>	changing environmental situations.
 
>That's seems to me to be a fair definition for what biologists are trying
>to study.  I certainly don't know enough to discuss it in that realm
>anyway.  But to apply that definition to cognitive science leaves out
>the capability to just sit back and think, to have thoughts or engage
>in cognitive activity with NO change in behaviour.

Ah, but I would call this kind of process abstract reasoning, *not* learning.
In short, it useful to clearly seperate the concept of learning per se from
the other components of intelligence, like reasoning.  This allows each piece
to be studied and characterized on its own merits.

I do admit that I may have made a mistake in including 'neurological' in the
above definition.  Basicly what I have in mind is that any system that modifies
its behavior on the basis of prior experience shows learning.  This concept
can certainly be applied to cognitive science, since it is level independent.
Changes in conceptual structures due to experience, at least in living things,
invariably led to changes in behavior.

> I can sit back and
>enjoy the memory of the date I had last night, or draw conclusions
>about S&L presidents who didn't run the banks properly.  But these don't
>necessarily change my behaviour.  This is learning, as I'm arriving at
>new data (e.g., "Those S&L guys are crooks!"), but I'm not changing
>my behaviour (I don't bank at S&L's anyway).

I rather think that your behavior is changed more than you might think by this.
The changes are likely to be subtle, and hard to link with your opinion of
S&L operators - but they will still be there.  [For instance it might lead
you to say 'Those S&L guys are crooks!', which you would not otherwise have
said]

>Yes, but we have some different neurons (e.g., cerebral cortex vs.
>hypocamus (sp?), and more of them.  The "hardware" is important, but
>what we can do on top of it is what makes learning, or intelligence,
>what it is.

Well, the different types of neurons are not really all that different.
They differ mainly in the type and pattern of connections they make.  They
may also differ in the type of signal they send.  However all of these types
of variation exist in butterflies and even worms - it is only the exceedingly
primitive forms like hydra and planaria that lack internal differentiation
of nerve cell types.
[BTW the hippocampal neurons are essentially identical to the cortical ones,
there is actually more variation *within* the cortex or hippocampus than
between them.  A pyramidal cell is far more different from a stellate cell
than a coritcal pyramidal cell is different from a hipoocampal pyramidal cell]

>>I would distinguish between learning, which is shown by all forms with a
>>nervous system, and intelligence which involves creative behaviors - the
>>initiation of new behavior by mechanisms other than simple trail and error.
>>This includes anticipation, modeling, improvization, recombination of
>>behavioral primitives & c.
 
>You may be right.  But in some of the examples you cite (such as modeling)
>we currently don't have a way to see if cats model internally.  The only
>way we can with humans so far (as far as I know, anyway) is to query
>them as to what cognitive process is occuring.  So while cats might be
>doing just that, we don't know if they really are, and won't until we
>have a more accurate and language-free method of determining if this is
>true.

There are some behaviors which appear to require internal modelling which
can be observed without needing language.  The classic example that I know
of involved an ape rather than a cat.  In this experiment with the ape it
was placed in a tall cage with a bunch of banannas hung on a string from the
top, out of the ape's reach.  Also in the cage were a crate and a stick.
After jumping up and grabbing at the fruit for awhile, the ape went and sat
down and thought for awhile.  Then it got up, placed the box under the fruit,
picked up the stick and knocked the banannas down.  This was done *without*
any significant amount of trial and error - the ape clearly had some idea
of what it was doing.  I maintain that this is proof of some sort of internal
modelling, in which the ape did the trail and error in its head.

>While Man goes beyond this in civilized society, I'll concede this as
>at least the original purpose or function of the ability to learn.

Quite.  This is an example of what biologists call 'pre-adaptation' - a
feature or capability of an organism that evolved for one purpose that, by
accident, is useful for something entirely different.

>I agree.  The reason I go on about learning and intelligence in other
>animals is because we arn't very rigorous about what these things are
>and how to study them.  We often rely solely on behaviour without
>regard to the internal activity going on.  I don't think we should
>be suprised if we find out that while some behaviours look the same
>between humans and animals, the motivations or internal mechanisms that
>cause them are very different.  In other words, we're naturally prone
>to anthropomorphism.

I do agree with this.  I try to be more rigorous than many people here.
I find that many of the sign-language experiments with chimpanzees
particularly wanting here.  However, I am not sure that a difference in
mechanism should necessarily rule out the use of terms like intelligence
or learning.  Certainly, I doubt that the ape above was accompanying his
thought with the kind of running monologue humans tend to use in that
situation.  What I would say is that behavior which involves novelty
or flexibility is evidence of some degree of intelligence, whatever the
internal mechanism.

Certain mechanisms are likely to be self-limiting, and thus unable to
give rise to more sophisticated levels of intelligence.  Thus it may be
that human level intelligence may be achieved by only one mechanism, and
that is why, say, porpoises have not achieved that level of intelligence.
[That is porpoise intelligence may be based on a mechanism that precludes
the level of abstraction humans are capable of].

>I won't argue with the biologist definition of intelligence - if the
>distinction works will for them, that's fine.  As a trained cognitive
>scientist, I take the cognitive approach to the definition.

The main advantage of the biological defintions is that they provide a
cleaner seperation of the concepts into more basic, more easily studied
components.  This allows for greater 'modularity' in working with them.
[You can study learning without also studying conceptualization, or you
can study concept formation without dealing with learning and so on].
I think that cognitive science would be well served by applying this
level of reductionism.