Re: How much intelligence?


Wolf Kirchmeir wrote:
Glen M. Sizemore wrote:
"Curt Welch" <curt@xxxxxxxx> wrote in message
news:20060428005535.234$2A@xxxxxxxxxxxxxxxxx
"chadmaester" <chad.d.johnson@xxxxxxxxx> wrote:
...it's been a while, but oh well - I don't have a lot of time on my
hands cause it's finals.

I think you will find that behaviorists do not believe
the brain stores anything. It simply "changes" in
a way to make it behave differently next time. Or
something like that.
To me, that is simply incorrect. The brain has to be storing things
because it can recall things without being anywhere close to whatever
it is thinking about. For instance, if I'm on the other side of the
world I can easily recall what my house looks like back home - anytime,
any place.

A computer program that recognizes handwritten characters stores
information about the characters. Indeed, it changes the more and more
it trains and learns the characteristics of the characters, but it DOES
have internal representations of the characters - the connections and
strengths of connections between the neurons (artificial neurons, that
is). I think our brains do the same thing, but they have a mechanism
which allows us to put these characteristics together and form mental
pictures (maybe STM; and maybe that's what cognition is). We can create
connections between neurons, and the strengths and placements of these
connections are what give us understanding. I'm not sure what forms a
solid memory - probably the more connections that exist, the more or an
understanding/grasp we have of a concept or concepts (if you know
anything about associative memory models, a single network can store
multiple memories).

Sorry I went off on a tangent there, but do you agree/disagree?
There's no doubt, even by behaviorists, that these sorts of things are
"stored" in the machine, and in the brain. They just don't think the word
"stored" or the concepts connected with the word, are a valid or useful
way
to explain what is happening.

A better way to say it is simply that the animal is changed by exposure to
certain environments and that, as a consequence, the animal behaves
differently. The term "change" is not a synonym for "store". There is no
advantage to speaking carelessly.



For example, many people would believe that saying "God caused the flood",
is not considered a valid or useful way to talk about floods, but the
people that reject the God talk, are not rejecting the ideas of floods.

Behaviorists reject the use of storage talk, to explain the cause of human
behavior that can be better explained using other concepts.

--
Curt Welch
http://CurtWelch.Com/
curt@xxxxxxxx
http://NewsReader.Com/

Talk of "storage" of information tends to come from people with a
computer background. They don't like talk of "changing the animal",
because that reminds them too much of "self-modifying programs", which
they have been taught is a Bad. They tend to think of an intelligent
program as one that accesses, adds to, and modifies a database rather
than one that modifies its own operation, ie, one that accesses,
adds,to, and modifies its own code. So of course they tend to see the
brain as accessing, adding to, and modifying "stored" data.

IOW, 'learning' means changing the stored data, but not how the program
works.*** Hence talk about the brain's being "hardwired" by the genes,
the brain controlling the behaviour, etc. The mistake in this sort of
talk is to assume that control is a meaningful concept - which it
happens to be when discussing a stored-program computer, but not when
discussing animal brains. There are other examples of vague and
inconsistent concepts, but these will do to show that the most common
computationist models of AI cannot work.

Note that a learning program of this type must not only augment the
database, it must also be able to add to and modify the rules of
inferences/interpretation/etc of the stored data. That is, it must be
self-modifying. So the stored-data model of knowledge/intelligence/etc
is inconsistent. Or so it seems to me. Also, "feedback loops" imply
self-modification of the program.

I am not at all bothered by the notion of "a learning program == a
self-modifying program." A multiple-threaded program, running on a large
number of processors, should be able to modify itself quite safely,
since the threads can be written to monitor each other, and so prevent
certain types of self-modification, or prevent their execution, or
correct them, or etc. Not that I believe that a computationist model of
AI is the best approach, but then I don't know enough to decide what may
be the best approach.

Imagine an OS that not only knows how to identify a virus, but knows how
to fix the code that the virus exploited.... A computer running such an
OS would be about as smart as -- a cell.

Just a few rambling thoughts, triggered perhaps by the ingestion of a
large chunk of dark semisweet chocolate. :-)

HTH

Well of course a computer program can be treated as data
because it is stored in the same way as data. Are the formula
typed into a cell of a spread*** data or program? A text
of a BASIC program is data for an interpreter or compiler.


Of course there is field programmable logic which we might
modify (rewire) on the go and maybe the neurons can do
something like that, at least during development.


However I suspect the brain's operating system is embodied
in hardware unlike the operating systems that run on the
von Neumann architecture. Any attempt at duplicating a neural
network on a computer means a virtual architecture written
in software.


Let's say we consider a simple one "neuron" unit brain.

+---+
stimulus -------->| |
| N |-----> response
+--->| |
| +---+
|
+--------------- feedback


The response of this "neuron" at time t depends on
its stimulus/feedback history.


Feedback will "sensitize" the neuron, meaning it
will fire more readily in future. Without feedback
the neuron will "habituate" to the stimulus meaning
it will become less likely to fire (respond).


Thus two twin neurons will end up with a different
response at time t for the same stimulus if they
have had a different stimulus/feedback history.
That is how the behaviorist seem to view it.


But the reason why two different histories produce
different outcomes can only be understood in terms
of the properties of the neuron. And if you want to
build AI you need to know how your units work.


You might say the neuron has been "sensitized" or
you might say it "remembers" the feedback depending
on context. The mechanism may be similar.


In Glen's terminology I guess he might say the neuron
is changed by exposure to certain environments and that,
as a consequence, the neuron behaves differently. But
cause of the change in behavior also involves the
innate properties of the "neuron" unit.


What Glen calls "speaking carelessly" when mentalistic
terms are used may in fact be revealing the computational
basis for those terms.


--
JC

.

Quantcast