Hi, Warren –
[Warren, this developed into a long essay that others will be interested
in, so I’m copying this to CSGnet]
Do you think that the perceptual
inputfunctions that convert signals from multiple sources in the level
below to form a signal that for example is involved in the 'amount
ofspider’ could be summarised in the balance of gains and delays
withinthe connections and develop through normal reorganisation as the
person interacts in a controlled way with that object; do you think
that some of these balances would have certain starting points
(e.g.presumably low level loops that are nearer would send signals
quicker), or would there be certain other cells who had a function
other than control loops with weighted inputs that would manage
thisprocess?
It took me a while to decode that 109-word question (No, I didn’t count
them, I used an editor that can count words)!
I doubt whether “gains and delays” are quite enough, but I
don’t know. An analog computer capable of doing all kinds of math needs
addition, subtraction, multiplication, and division operations, as well
as integration and differentiation. Also some nonlinear functions are
useful, such as raising to a power or generating thresholds and limits.
In a nervous system we can’t expect to have pure mathematical operations,
but in the analog world a good approximation is all that’s needed. I
showed in B:CP how neurons might do some of those things (Chapter 3). And
maybe limiting the types of computations to what we humans think of as
mathematical operations isn’t really necessary. If you’re asking if such
operations, plus reorganization, are sufficient to create a
spider-recognizing perceptual input function, my answer is
maybe.
Trying to guess in what context you’re asking the questions, I could
guess that you’re asking how new levels come into being. Answer that
question and we can start General Robotics and begin looking for Susan
Calvin (see Azimov). The basic question is what sort of beginning
organization would be needed in order for the reorganizing process to be
able to create specific examples of control systems like
spider-recognizers. And on top of that, what sort of beginning
organization would be needed at each level to assure that the
necessary kind of input and output functions would result from
reorganization? Where do my 11 levels come from, and are others
possible?
The answer many people would like best would be that one generic neurone
type would suffice, with the products of reorganization at each level
coming naturally out of interactions with the environment that involve
the use of lower-level systems already organized.
However, the physical structure of the brain makes that simple answer
look unlikely. The brain isn’t made of just one type of neuron, but of
many obviously different types, like pyramid cells and Purkinje cells –
I haven’t memorized a long list, but there are lots of types, and lots of
typical connections like the broad bundles of axons connecting the
hemispheres (the corpus callosum), and the way the optical and other
tracts cross over to the other side, and many other things that suggest a
large degree of inherited organization. Of course it’s possible that
those types differentiate from much simpler types, even one type, at the
earliest stages of neurogenesis. That has to happen, of course, but to
what extent is it the result an ab initio DNA-directed
organization, and to what extent does it arise from interactions as
development proceeds? See Nature, 2 July 2009, page 35 to move this
question to the level of DNA.
All that initial organization has to arise from or through DNA somehow,
which means that if reorganization is the answer, it has to occur way
back in the genome, and even further back in time: It has to come from
evolution, and an eons-long process of slowly building level upon level.
The tree of life shows the paths that reorganization took; evidently
there are many alternate paths that lead to viable higher levels, though
the multiple forms of life that exist now show that once a successful
kind of underlying organization is found, it persists for a long time in
succeeding forms. We look an awful lot like chimpanzees, and act like
them too.
I attach a paper I wrote some years ago that takes us back to the
primordial ooze. You’ve probably seen it, but just in case …
Having follow that chain about as far back as possible, I could also
guess that you just want to know whether reorganization is enough to
generate a spider-recogizing input function simply from interactions with
spider-objects. My analog-computer answer hints at a “yes,” but
only if there is enough pre-organization to permit reorganization to work
by increasing and decreasing synaptic weightings.
E. coli reorganization, the kind we use in PCT, needs a criterion to work
with: some variable that indicates whether things are getting better or
worse. If things are getting better, keep changing in the direction
you’re going as long as the errors are getting smaller. If things get
worse, tumble and take off in a new direction. If there’s enough
organization there to produce some behavior, with enough parameters to
fiddle around with, this works very nicely. It works way better than just
changing the parameters to new values at random and then seeing if the
resulting system lives long enough to reproduce. As long as that
idea guides our thinking about evolution, the creationists still have an
argument. That kind of natural selection doesn’t have much of a chance of
actually working. In all models I have seen, it works only if you give an
organism permission to reproduce merely because it moved a little
closer to the behavior or form needed to get to the procreation phase
– it doesn’t need actually to get there. I think that’s cheating. And I
think there is evidence in the fossil record that evolution is
reorganization, not saltation. OK, not random jumps.
To get perceptual input functions from reorganization, there has to be
some continuum of effects on the well-being of the organism that depends
on how the input functions are organized. In LCS3 I speculate that one
effect would depend on the amount of conflict that exists in a population
of input functions at the same level. Conflict of the sort I mean is not
just a disturbing effect, it results from having input functions that are
fundamentally incompatible, so in the current environment it’s impossible
to set different reference levels for all the coexisting perceptual
signals. For one perceptual signal to be at some value implies that
others cannot be at certain other values. This guarantees some degree of
conflict between the control systems no matter how their output functions
are reorganized.
When conflict among perceptions is reduced enough, it becomes possible to
choose arbitrary reference levels for all the perceptions at the same
level, and for the output functions to be reorganized so they bring all
the perceptual signals very close to their respective reference levels
with minimum effort. This isn’t an either-or proposition. It just means
that as the input functions come into some degree of conflict, more and
more output is required to achieve zero error even when the output
functions are optimally organized. Conflict becomes damaging when it
reaches a degree at which the output needed to maintain control is more
than the organism can produce or afford to waste. It’s that relationship
between orthogonality of input functions and intrinsic error that
provides the gradient of effects on well-being that is needed for
reorganization to succeed.
This is what makes the reorganizations in demos 7-2 and 8-1 in LCS3 work
– except that the user has to change the orthogonality of the input
functions in 7-2 (by changing the value of the determinant of the matrix
of input functions), and in 8-1 the orthogonality of input functions is
guaranteed. If I ever get to it and have some luck, this is how I will
generate an organism that can reorganize its perceptions as well as its
actions. And that should be very interesting indeed.
If we have an organism that can reorganize its input functions, then we
will find out if there are any inherent levels of organization in the
outside reality. Reorganization will discover them. And everybody who
reorganizes will discover the SAME perceptual entities because they
REALLY EXIST. That includes not just everybody here, but everybody in the
whole universe. I think that proposition would change a lot of things –
whether we proved it true or false.
So, ask a simple 109-word question and get a drink from a fire-hose. But
these ideas are coming together and may lead to the next stage of
developing the models, maybe the stage where our models start looking a
lot more alive.
Best,
Bill P.
origins2.doc (168 KB)