X-Message-Number: 33452
From:
Date: Thu, 10 Mar 2011 01:50:04 EST
Subject: Thomas Donaldson on How We Are Not Computers
How We Are Not Computers, And What That Means
By Thomas K. Donaldson
This article, while not a book review, does depend strongly on two
particular books, "Artificial Life" (ed. CG Langton, 1989) and "Brain
Organization and Memory" (ed. JL McGaugh, NM Weinberger, G Lynch, 1990).
My discussion comes from my own thinking, but one article from each book
has particularly influenced me. From "Artificial LIfe," S. Hameroff and
his colleagues argue (p. 521) that we should not take nerve cells as simple
computational units. Even single cells have complex abilities. Nerve
cells, and even signal transmission between them, has much more complexity
than a stream of single bits. Hameroff reinforces that point. Second, the
article by WJ Freeman and CA Skarda ("Representations: Who Needs Them?",
p. 375, Brain Organization and Memory), helped me clarify my own thoughts
about brain models.
The computer model
The idea of brains as computers has become very popular. It lies
behind the idea that someday we will upload (download?) ourselves into
other computers somehow better than the wetware one we work in now. It
also underlies a lot of mainstream theory about how our brains function:
cryonicists believing a computer analogy have many other scientists to
point to for support. If our Selves are computer programs, uploading
becomes trivial. Before anything else we first need to clarify just what a
computer program and a computer are supposed to be. For the purpose of
this article, I shall consider programmability as the main feature
distinguishing computers from other objects. A program is a combination of
instructions and data which controls the operation of the object. If that
object is a computer, then means exist by which a wide variety of different
programs can (at different times) control its operation.
Clearly devices (and living things) fulfill this criteria to greater or
lesser degrees. Even within the class normally accepted as computers, some
may be unable to perform a set of instructions because memory capacity is
too small. Performance of instructions, in general, requires not just
ability to compute but peripherals such as monitors, printers, and disk
drives. And some devices often thought (perhaps loosely) to be computers
cannot run a wide variety of instructions (ie. embedded devices each with
one single program burned into its ROM). So no object can lie at the
(theoretical) end of the computer side of the rainbow. As for the other
side, for objects totally unable to perform any separate instructions,
rocks or stars fit that description very well. Perhaps some living things
do also.
The analogy of our brains as computers suggests many common ideas.
Behind it lies an image of human brains as objects which are all,
fundamentally, identical. They differ only in the programs they are
running; these programs are Persons, everything that makes you You.
Again, computer programs operate on Data. Data is always a symbolic
representation of some part of the real world. By the computer analogy
when we remember our home town, then, we do so by forming a symbolic
representation of it in our brains. We must distinguish the symbolic
representations we use in talking to one another from those in our brains;
most mammals are quite inarticulate, but somehow find their way around
their environment. A computer analogy for their brains would suggest that
they too have such symbolic representations. If our brains do work as
computers, such representations become essential to their operation.
Furthermore, in principle we might devise ways to transfer specific
memories between one person and another. The language of the symbolism may
differ from person to person; but with the proper translations that becomes
a detail. A computer analogy implies that training of any kind might
someday be compressed into a few hours on a "brain trainer." It also
causes us to spend special attention on these supposed symbolic
representations. The ability to manipulate symbols becomes identified with
intelligence itself. This body of ideas about how our brains work ties in
very well with such efforts as Chomsky's, to explain our spoken language as
a form of translation from a private symbolic representation to a public
one.
It has become, in short, the dominant image of brain operation in the
late 20th Century. Dominance, however, does not make it correct.
Another and different analogy
We can already see signs of another quite different idea. I shall call
this model for memory the Growth Model.
One major theme in the study of memory has been the idea that learning
(even in adults) involves the very same processes by which our brains
develop from embryos. That is, remembering
something long-term means that our brains have formed new physical
connections between neurons; these connections persist by the same
processes maintaining our physical form. This idea easily answers one
major question about long-term memories: why are they so durable? No one
knows a way to destroy a memory without physically destroying neurons
involved in it.
Processes of development also include healing. True, neurons in adult
primates may divide only rarely (although some experiments suggest the
contrary). But healing in our brains includes more than simple division:
it can involve massive rearrangement of circuits, as in recent (unplanned)
work with monkeys with severed nerve connections to their hands.
Development involves not only passage of signals between neurons which
already touch one another, but chemical signals causing growth of a
dendrite or axon toward a neuron not within "touching distance."
Any dominant analogy creates an impression that no other possibility
can even exist. Yet, if followed out, the developmental hypothesis just
sketched above suggests very different conclusions about how our brains
work.
First, long-term memory forms when our brains grow a set of new
connections. These connections would contain new synapses. Hence
(contrary to other estimates based on a computer analogy) to estimate our
capacity for new memories we must do more than count nerve connections. A
maximum capacity would still exist, reached when present connections left
no room for more. How that might happen remains an open question: perhaps
simple crowding, or again single neurons might only support a limited
number of connections.
Furthermore, our long-term memories would consist of the connection
patterns that have grown up between our neurons. They would not be "coded"
into our brains, in any sense of "code." On a gross level our brains do
resemble one another; but when we find out how to look at a brain closely
enough to distinguish memories, we would find these memories identical to
the connectivity itself.
This model resembles the neural net computers that computer scientists
now use, successfully, to make machines to solve problems our brains do
easily. Neural net computers don't store their memories in any one
connection, but in the pattern of all their connections. In this way they
resemble our brains. But neural nets start with a fixed set of possible
connections, some of which are turned on, others off. Unlike neural nets,
brains would form memories by growing new connections. Disused connections
would disappear.
Our thinking would also proceed totally without any symbolic relation
to the world. Instead, our nerve cells have grown connections so that
their total response to any outside event deals with it successfully.
(Note that neural net computers, too, do not use any symbolism in their
computation: in this way they follow brains).
Such systems resist any easy transference of "programs" from one to the
other. In that sense they aren't computers. (Although we certainly can
imagine some massive intervention which reconnects an entire brain. As
before, "computerness" is a matter of degree.) Nor could we make learning
easier simply by separating out a set of skills and knowledge and then
reading it into our own brains, for no single memory can be separated from
any other. And even if someone else could follow all the excitations in
your brain for every neuron and synapse, they would need a long prior
period of observation to read off from them just what it was that you were
thinking. . . other than in the very broadest sense. (That is: whether
you are sleeping or awake, you are thinking something about food or sex,
you are afraid, etc.)
Directions for use
If our brains follow the Growth Model, some common ideas about possible
improvements would require rethinking. Simple transference of our Selves
from one body to another ("Uploading") raises far more problems.
Improvements in learning ability, or transference of particular skills or
knowledge, do also. But here are some ways towards the same aim.
Preservation of alternative copies
This technology should interest every cryonicist. By storing
inactivated copies of ourselves we can survive total destruction of our
main, living copy.
Even if we are not computers, our structure might perfectly well be
stored in a computer system. Graphs give the main data structures needed.
We would store each connection with additional information (just what isn't
fully known yet: transmitters used, its age, and possibly other items).
The practical problem to solve for such a system is how to read out
brain connectivity as rapidly as possible. Fast read-out rates let us
frequently update our stored copies. If you have not been updated in the
last 10 years, then any destructive accident would mean a loss of 10
years. One idea would be to add a system of "watchdog molecules" to each
synapse; these shed copies of themselves constantly. They might then be
gathered together to find how your brain has changed.
Increase in memory capacity
No one yet has faced this problem, but at some stage it will arise. I
believe the most likely brain response (by our unmodified brains) would be
to forget all information least used, rather than to simply stop learning
(some neural nets already do this). In the Growth Model, synapses between
neurons would disappear.
Basically, increased capacity requires unwieldy increased storage
space. Miniaturizing brain circuits (while keeping the same connectivity
system as before) only puts off the problem. True, you might separate your
"extra brain" from your main brain. But even if your "extra brain"
connects with your "main brain" at the speed of light, your memory will
fade significantly if you even go to the Moon.
We can still add off-line storage space, relearning older memories when
needed. Relearning, of course, involves growing new connections.
Note that the same problem arises with neural net computers. One other
point needs stress: despite the limits, by miniaturization and larger
brains we might increase our capacity by at least a factor of 10. That is
still very worthwhile.
(25)
Increase in learning speed
Learning, in this model, involves growth, which takes time, energy, and
materials. Already our brains burn a large share of the calories we eat,
as high as 40%. To cut down energy and materials expense, we might first
use miniaturization. After that, we might add a special cooling system,
perhaps an extension of our present cerebrospinal fluid. Our blood would
bring more materials and also take away excess heat. For temporary periods
of learning at very high rates, we might also imagine special "Learning
Stations" to rearrange our brains, allowing not only increased blood flow
but cooling solutions.
What about increasing "intelligence?" But just what does
"intelligence" consist of? Besides ourselves and other animals, we now
have computers, capable of spectacular feats of processing on some problems
and spectacular stupidity on others. Even other animals can do processing
we cannot (dolphins and sonar, for instance).
The lesson of these examples is that many different kinds of brain
processing exist. In the end, we will want some increase in our learning
ability, but rely on many different systems for other kinds of processing.
These would connect to us in detachable ways, more or less like present
computers. We may even develop special interfaces to attach to them, much
as our hands attach to our machines; but they would still remain apart
from us. Hands were a good idea and remain so. (The problem with making
any kind of ability a permanent part of yourself is that you may not always
use it: one more piece of baggage. With too many additions you grow too
fat, metaphorically and actually).
Why not move ourselves over into computers?
Some would say that by doing this we would become essentially
different, and so lose our Selves and our personality. That may be so,
although I know no logical or experimental means to find an answer.
Instead I will discuss one major practical reason why a growth model may
have prevailed for brain design.
I shall discuss only neural net computers since, so far, they alone can
do some kinds of learning needed. Suppose then a neural net computer with
the same capabilities as our brains in all respects. Neurologists have
classified about 100 or so different processing regions in our brains, each
one a neural (sub)net. The advantage of a neural net computer (with fixed
connections) over a brain would be that all connections had been grown in
advance. (Even at start this "advantage" may mean little: short-term
memory allows a temporary learning until growth has finished. That may
even be its explanation.)
13 13
With 10 neurons in each processing region, each neuron needs 10
synapses to make all possible connections. (Neurons now have a maximum of
about 1000 connections, within a factor of 4.) Let each connection cover
only 1 nm2 of surface. Total area of all connections becomes 10,000
meters2, or 100 meters on a side, for only one neuron. A brain designed
this way carries along one billion (109) times the mass it actually uses.
What about virtual connections? That merely turns one kind of unused
capacity into another (virtual connections use other neurons to transfer
impulses, invisibly to sender and receiver). Virtual connections may not
even work: direct connections between neurons must exist for a reason.
(This is an argument valid for both silicon and protoplasm).
If we try to limit connectivity a priori we find another problem. By
limiting possible connectivity, we limit the kinds of connections our
brains can make. Given that all new connections form on a background of
the old, this means a limitation on connections between responses. Brains
with fixed connectivity, then, will lack adaptability.
These two factors, combined, may tell us why our brains operate by
growing new connections rather than staying solely with the old. Yes,
growth takes longer. But it may also support our mental flexibility, which
is still far more than any computer and may remain so. Perhaps someday we
will modify ourselves to even more flexibility. We can see now how to do
so.
Content-Type: text/html; charset="US-ASCII"
[ AUTOMATICALLY SKIPPING HTML ENCODING! ]
Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=33452