[From Bill Powers (920914.1530)]
Chris Love (920914.1300) --
The problem: By allowing those n-1 level ECS' to provide the
imagined percepts to your (focus) n-level ECS you are getting
(undesired?) influence from neighboring n-level ECS'?
Let's see, when could this cause a problem? I guess it would be when
there are environmental disturbances that actually cause trouble at
the level n-1 systems in the imagination mode, which of course those
systems don't report to the higher level as a change in the
perceptions. So the systems at level n that think they are dealing
with real inputs are getting partially-made-up inputs, and therefore
fail to take corrective actions they should be taking. Of course
that's true in spades for the level-n system that wants to be
imagining, but it's a mistake to cause other systems at that level to
deal with false information.
Perhaps there has to be some "context" such that a whole group of
level-n systems has to use imagined information at the same time. Or
perhaps -- and this is just as realistic -- real systems DO get into
this kind of trouble! This thing didn't come with an instruction
manual.
Delta weight = learning rate* connection wt. * error signal
New wt. = Old wt. + Delta wt.
Comments:
The learning rate is typically set to a small value between
0.1 and 0.3.
The connection weight is the exisiting weight (wt.) of the
connection of interest.
The error signal is that signal eminating from the comparator.
I think there's something to the GENERAL Hebbian idea of learning, but
the concept that more strength in a connection is better just doesn't
hold up. When you get around to designing control systems that have to
behave in an environment with somewhat physical properties -- in which
the effects of actions don't happen instantly -- you'll find that
"more" output is a good thing only up to a certain point, after which
more still will lead to instability. Also, trying to use Hebbian
concepts in adjusting weights will work only as long as the optimal
condition is all weights at maximum or zero. For computing many
perceptions, some weights have to be adjusted upward and some downward
to get the optimal result. I think in general that some _particular_
weight is the correct one for the situation, not _maximum_ weight.
I've seen some work (Harry Klopf, on the "selfish neuron") in which
all the working models that used this Hebbian sort of learning failed,
because their only stable state occurred when all weights went either
to maximum or zero. To get them to settle down anywhere else, the
modelers had to introduce all sorts of ad-hoc limits and
nonlinearities. Just the right ones, of course, to produce the results
they wanted. When you have to fiddle with a model at that level of
detail to make it work, something is probably basically wrong with the
model. You end up with a "right-result-producer" that only works when
it's tweaked _just so_.
What does this do?
Well, as I see it it correlates particular references with
particular percept signals.
These words don't make any sense to me. Why would we want to correlate
a reference [signal?] with a percept signal? Do you mean that the
resulting error signal will cause lower level system to alter the
reference signal toward the value of the reference signal? I just want
to make sure we're talking about the same organization.
As to your Almighty ECS that you mentioned to Rick, you're going to
have to be almighty careful about what kind of knowledge you give it.
After all, it doesn't know whether it's controlling categories or
intensities. Whatever you build into it has to work exactly the same
way no matter what it's controlling. Also, if you try to build in all
levels of functioning of all kinds into every ECS, you're going to
have a tremendous duplication of function, sort of like giving every
cell that uses oxygen a pair of lungs.
···
----------------------------------------------------------------------
Penni Sibun (920914.1200) --
the first ``official'' (it's got a lot of publicity and money
attached) turing test was held last year. i've seen some transcripts
and the system that won seemed pretty credible (it ws judged human
about half the time). however, the task was limited not only to
teletype conversation, but to a particular topic.
Even with just a particular topic, it shouldn't be too hard to tell
whether the other guy was a machine if you were free to ask questions.
For example, you could take any two statements by the other entity and
ask "Can you show me that these statements really relate to this
topic?" or "Are these statements consistent with each other?"
Actually, it would be interesting just to respond to any particular
statement by saying "No, you're wrong about that." Anything to elicit
some signs of higher-level perceptions or an attempt to control for
something.
----------------------------------------------------------------------
In your reply to Rick you say
no. the sex test looks at a person and guesses whether they are
female or male. period. the sex test doesn't care about what's
inside. it only has access to what you consider ``superficial''
things.
I guess that as usual big disagreements rest on little words, in this
case "are." When you say you're trying to guess whether they "are"
male or female, the normal interpretation would be that you're trying
to guess what they REALLY are, not just how they look. In other words,
you can only tell whether your guess is "right" by (blush) peeking.
If you mean ONLY judging by the readily-visible characteristics, there
isn't much guessing about it, is there? You just classify your
perceptions the way you're used to classifying them.
----------------------------------------------------------------------
--
Best to all,
Bill P.