There is also a very interesting note by Gabriele Scheler
at the Salk Institute about neuromodulation in the brain
and how "the subject area may really change our views of
brain-style computation." He points to neuromodulation as
further evidence for "global control mechanisms" in the
brain and provides a list of references. And Mark Bickhard
at Lehigh University, John Karwatzki at Kingston
University, Talib Hussain at Queen's University and others
refer to computational architectures in their work that
explicitly use control theoretic ideas.
I also received two responses from Ron Blue that were long.
Those interested can contact him directly at <rcb5@msn.com>.
A copy of my original memo is included in the Appendix for
reference. Further comments/questions are welcome.
Asim Roy
Arizona State University
============================================================
>From PETER J. ANGELINE at Natural Selection,Inc.
<angeline@natural-selection.com>
I read with interest your recent post. Let me make a few
comments on training NNs that maybe counter to your
perspective and offer my opinions at the end.
You state correctly that NNs have been created for some
time with an implicit control theoretic model. This, in my
view, provides absolutely no justification for any
assumptions related to how neural wetware is created. The
limitations of those NN researcher's imaginations on
developing standard training algorithms has no bearing on
the organization and operation of neural wetware. All that
need be done is to identify a single algorithm that does
not use global information to train NNs to call this
assumption into question.
Evolutionary computations can be used to induce equivalent
NN structures without any global processing but merely
through local competition. This has been born out in a
number of studies in the literature. A similar, more
complete perspective is offered in Gerald Edelman's book
"Neural Darwinism" which is specifically an investigation
into how a Darwinistic approach can be used to explain
brain organization and human learning.
In addition, I believe, there is a paper by John Kolan and
Jordan Pollack in some NN journal a few years back that
shows how a backprop algorithm could be implemented in
wetware without the need for global information.
None of the above is intended to dissuade you, just to
inform you that your original assumption is not a
necessity.
My own perspective is that there is no centralized control
for learning in the brain. Besides being eerily similar to
the long discarded humunculus argument, this perspective to
me appears too simplistic to explain brain organization and
operation, and probably impossible or extremely expensive
to implement in biological practice. More likely, the
mechanisms in the brain have been opportunistically
organized into loosely mutual "controllers" at best. Some
portions might partially "control" certain aspects of
others but it would be more harmful than helpful to
identify any one as a "controlling" entity for all.
At best, control is distributed, but even this seems
simplistic to me. Your example of "limbs controlling the
brain" neglected to mention that there is some processing
done in the limbs that does not require brain intervention.
"Control" then of the limbs is at best distributed between
the brain and the limbs. If it exists in the limbs it
probably exists in the brain as well. Again, I feel this is
even too simplistic a concept to be useful.
My perspective comes from my work in inducing executable
structures (programs, NNs, FSAs) with evolutionary
computations. The processing in such structures is
extremely difficult to describe or characterize given the
opportunistic method of evolutionary induction, a method
that is most probably characteristic of the evolution of
the brain.
We as humans have a great need to describe and explain
phenomena and will occasionally force an explanation on a
complex system even when it is grossly inaccurate or
demonstrably false merely to satisfy this basic desire. The
basic concepts of "controlling" and "controlled" are much
too limiting to be, in my opinion, worthwhile in models of
the operation of the brain. A much more subtle concept will
be needed to advance our understanding here, one that
recognizes and embraces the opportunistic nature of the
natural process that originally gave rise to the brain's
organization.
------------------------------------------------------------
MY REPLY TO PETER J. ANGELINE:
A very interesting note. I have not looked at neural
wetware, but I would be curious to know of the inputs to
your algorithm for any given learning problem. Perhaps
comparing it to back prop, for example, and showing how it
is done differently might be helpful in understanding
what's going on in your algorithm.
------------------------------
PETER J. ANGELINE's reply:
My first work on inducing recurrent neural networks with an
evolutionary computation is published in the IEEE Trans on
Neural Networks, January 1994 (v5 n1) and is titled "An
Evolutionary Algorithm that Constructs Recurrent Neural
Networks" with co-authors Greg Saunders and Jordan Pollack.
Here conventional sigmoid nets are induced. My current work
is a generalization of this work dealing with the induction
of symbolic expressions for dynamic systems, either
discrete or continuous. The first journal article on this
new work is coming out soon in Cybernetics and Systems.
The inputs for my evolutionary induction methods are
identical to that used by a NN learning algorithm except I
do not supply the architecture of the net. Nodes and
connections are added at random to individuals of the
evolving population using various mutation operations. The
implied dynamic is that architectures that provide a
benefit to progeny learning the task more quickly will be
selected by the method. In this way both the architecture
and the weights are simultaneously induced rather than
iteratively as in most NN architecture manipulating methods
such as Upstart Optimal Brain Damage or
Cascade-Correlation. The "worth" of an individual is
typically measured by how well it performs the task,
although I also look at competing individuals against each
other which entails only a concept of "better" be defined
rather than the whole task a la supervised learning.
In my most recent work, the model of a neuron is expanded
to be an arbitrary symbolic expression (for instance
sin(log(IN1+IN2/IN3))) which allows for the simultaneous
induction of a set of symbolic equations describing the
dynamical system of interest. I have experimented with many
discrete time systems and am now moving into continuous
time systems. Typically the symbols available to create a
neuron are chosen to be reflective of the task to some
degree although not always.
If you are interested, my work can be found on-line at
http://www.natural-selection.com under my personal pages.
The technique is called MIPS (Multiple Interacting
Programs). I am happy to supply any other information
regarding this work if you are curious.
---------------------------------------------------
MY REPLY TO PETER J. ANGELINE:
>The "worth" of an individual is typically measured by how
>well it performs the task, although I also look at
>competing individuals against each other which entails
>only a concept of "better" be defined rather than the
>whole task a la supervised learning.
Pete, "who" is it, beyond the network itself that you are
trying to design and train, that is measuring the worth of
an "individual?" And "who" is it that is looking at
competing individuals or solutions? And "who" is it that is
"trying" out these different network designs/solutions? Do
you think that you might be using some "control mechanism"
to do all this, to guide your algorithmic processes?
----------------------------------
PETER J. ANGELINE's reply:
>Pete, "who" is it, beyond the network itself that you are
>trying to design and train, that is measuring the worth of
>an "individual?" And "who" is it that is looking at
>competing individuals or solutions? And "who" is it that is
>"trying" out these different network designs/solutions? Do
>you think that you might be using some "control mechanism"
>to do all this, to guide your algorithmic processes?
You are a wise man. %^)
Yes, often *in practice* evolutionary algorithms also fall
into the same trap. This type of issue comes up in biology
as a debate on the teleology of evolution which is
summarily ignored in the practice of evolutionary
algorithms. But there are ways around this problem to
some degree.
First, one could argue philosophically that there is a
"physics" to the evolutionary process as implemented by the
fitness function (the worth metric) and the selection
mechanism. The "physics" dictates that "less fit
individuals are less viable" and consequently die off prior
to reproduction (are not selected). The fitness function is
a shorthand processing "hack" for the actual physics in the
"competitive environment". Unless you are willing to call
physics a "controller" of our evolutionary development this
is a philosophically safe but ultimately unsatisfying argument.
A second more satisfying approach is to try to actually
create a physics which is not centralized, focusing on
"emergent" properties. Thomas Ray has a system called
Tierra that moves in this direction. ECHO developed by John
Holland (and others) and SWARM by Chris Langton have
similar philosophies as does some work by Walter Fontana.
They all decentralize the processing (and also remove the
teleological aspects) and just let things run to see what
developes. Ray showed the emergent development of
parasites, hypre-parasites and other interesting phenomena
with biological relevance. This approach is one of the
mainstays of (good) artificial life work.
Emergence is the key to this approach and emergence's
younger brother is opportunistic exploitation. And yes
"emergence" is as undefinable as "intelligence" so I will
spare you raising that argument in your next reply %^).
Allow me to haphazardly define emergence for the sake of
argument as the closure of properties of a "physics" that
is consistent, complete, and self-reinforcing. In other
words, take a "physics" (e.g. a dynamical system) and run
it. The set of properties of the attractor of the dynamical
system, and the transients along the way, can be considered
"emergent properties" of the underlying "physics".
The results of a process employing emergence are a pure
consequence of the closure of the "physics". There is no
controller. And arguably, if we choose a control theoretic
shortcut to compute those properties it doesn't invalidate
that they exist and are the result of an uncontrolled
process.
A quick example might help here. I did some early work on
inducing modular programs with evolutionary techniques. I
had a mutation that modified a LISP program to remove a
random subtree and use it to define a new function which
replaced the subtree in the individual. This was apart of a
larger technique that had already be shown to evolve LISP
expressions. Nothing in the "physics" specifically selected
for particular modular organizations, modular programs
emerged in the resulting programs because they provided a
benefit to the individual being acted upon by the "physics".
I am interested in your comments on the above.
---------------------------------------------
MY REPLY TO PETER J. ANGELINE:
I think the problem might be that the evolutionary people
are looking at a system A and then concluding that that's
how system B works, although system B is completely
different from system A in many different ways. I think the
brain comes prebuilt with a lot of good machinery
("controlling mechanisms") and we have simply overlooked
that in our theories so far. I hope work in neuroscience
over the next few years will gradually reveal how some of
this machinery works. The point is, we need to look at
system B more carefully, rather than infer the
characteristics of B by studying another system A.
Also, as I noted before and as you acknowledge, neither
connectionist nor evolutionary/emergent systems can escape
from using a "control theoretic" structure. But there might
be other problems with the evolutionary idea. The
evolutionary approach constructs the learning problem as a
NP-hard problem. But nobody believes that the brain (of
insects, birds and so on) is solving NP-hard problems.
============================================================
FROM GABRIELE SCHELER, AT THE SALK INSTITUTE <gabi@salk.edu>
I think you raise a very important point here, which
actually touches upon the relevance of neuromodulation to
connectionist and cognitive brain models, and I believe
that this is a core point in trying to further evolve our
models of the brain on the network or systems level. Of
course a lot has been done here on the level of the single
cell and compartmental modelling, but I believe compact
models built on the basis of integrate-and-fire models (or
rather SRN models, as in Gerstner 1997) are also becoming
amenable to realistic models of neuromodulatory action.
Basically, there is a set of small areas in the midbrain,
i.e. in the basal ganglia, which contain neurons that
produce (a) dopamine (b) serotonin (5HT), (c) acetyl
choline and (d) noradrenaline. These areas have widespread
afferent projections, which make them suitable to "monitor
perceptual states" (DayanMontagueSchultz1997), and also
specific efferent pathways to certain brain areas, such as
the striatum or the prefrontal cortex. In addition to
synaptic transmission which does occur (i.e. an
axon-dendritic coupling), there is also "volume
transmission", i.e. a rise in availability of the
neuromodulator in extracellular space, which together with
pre-synaptic reuptake mechanisms, fits the notion of global
parameters being set in order to provide control variables
of the functionality of some network very nicely, I think.
(I advocated this idea also in Scheler(1998a)). There are
a number of network models to date, which attempt to link
certain changes on the single-cell level to a different
functionality of the network, and correspondingly to
experimental, behavioral data.(E.g., LismanFellousWang98,
Scheler98b, Hasselmo95, Cohenetc.90 etc.,
cf. FellousLinster98). Of course the full picture (as
nearly always in (neuro-)biology) is more complicated than
that. In addition to those distant "control" centers
there may also be local effects of glutamatergic (fast
transmission) on "terminal release" of neuromodulators,
i.e. ongoing processing providing additional refinement of
the distant "control" signal. Also, and this is a very
important point, one has to look closely at what the
change in the level of a neuromodulator in a given network
really means. For each neuromodulator, there is a range of
typically 4,5,6 or so different receptors, which instigate
different actions within the single cell, and may actually
have adverse (paradoxical) effects. These receptors may be
ionotrope (meaning if activated they are capable of opening
ion channels on their own), but more typically they act
through second messengers (5HT and acetyl choline have both
types of receptors, but dopamine, noradrenalin (and
adrenalin) have only second-messenger receptors). For
instance, D1-receptors (dopamine) act on adenylate cyclase
and raise the cAMP-level inside the cell, while
D2-receptors (also dopamine) act on adenylate cyclase in
the opposite way. It seems, that these receptors are
preferentially located on interneurons (D2) or pyramidal
cells(D1) in prefrontal neurons, which the implications
that this has on the network functioning as a whole. The
second messengers, which act inside the cell, are part of a
host of fairly complicated biochemical reaction chains, and
there is no way one can discuss the whole story in this
framework. Then again, there are the data from in vitro,
and in vivo studies exposing brain cells to neuromodulators
or a number of receptor-specific substances (agonists and
antagonists) and recording the alteration in their
behavior. Again, I cannot summarize all the different
findings that people have made in that respect. But (and
this is what I was getting at) it may well be that our
notion of what may be a suitable and valid control
parameter may be upset by the actually observable behavior
- and even more interestingly, suggestions on what a
control parameter SHOULD do, may in some cases be
reconcilable with the electrophysiological evidence. For
instance, if you believe the tuning of the refractory
function is an interesting parameter in a network model,
you may reconcile that with the role of CA^2+ influx and
afterhyperpolarization effects which are known to be
subject to neuromodulation (Scheler98b). Or, it turns out
that dopamine stimulation is capable of blocking or
attenuating AMPA-receptors in contrast to NMDA-recptors,
which obviously will have effects in long-term potentiation
(LismanFellousWang). Etc.pp. I think there is whole lot of
interesting and stimulating subjects "out there" and (quite
possibly I'm not aware of a number of other network models
that already have been done) but the subject area may
really change our views of "brain-style computation" very
much. I have not mentioned yet the issue of interactions
among the central neuromodulatory areas (such as
serotonin-dopamine), nor the mechanisms of up-regulating
and down-regulating receptor densities in the time span of
days, weeks, months; nor the intricate way at which the
action of the central neuromodulator-producing neurons is
controled for instance in conditioned stimulus settings, or
in pre-pulse inhibition etc. Thanks for raising this issue,
I'm looking forward for all the other responses.
References:
Gerstner(1998): Spiking neurons. In: Gerstner/Bishop(1998):
Pulsed Neural networks. MIT press.
LismanFellousWang(98): A role for NMDA channels in working
memory. Nature neuroscience 1(4).273-275
Durstewitz, D.(1998) PhD thesis, Ruhr-Universitaet Bochum.
Scheler,G.(1998a) A theoretical view of dopamine
modulation. In: International Conference on Computational
Intelligencs and Neuroscience, Research Triangle Park,
October 23-28.
Scheler, G.(1998b): Dopaminergic regulation of neuronal
circuits in prefrontal cortex. Submitted to Neural networks.
DayanMontagueSchultz(1997): A neural substrate of
prediction and reward. Science 275, 1997.(1593-1599)
Houk, Davis, Beiser(1995): Models of Information
processing in the basal Ganglia. MIT 1995.
Hasselmo(1995) Neuromodulation and cortical function:
Modeling the physiological basis of behavior, Behavioral
and Brain Research67(1);1-27.
Cohen, J. Servan-Schreiber: A network model of
catelochamine effects. Science 1990.(vol. 249)
============================================================
FROM JOHN KARWATZKI AT Kingston University <karwatzki@KINGSTON.AC.UK>
> Overall, it appears that a very convincing argument can
> be made that there are subsystems within the brain that
> control other subsystems. This "control theoretic" notion
> would thus allow external sources to directly control a
> cell's behavior.
I agree. I have previously published a proposal for brain
architecture (AISBQ No 73, Summer 1990) which involved
massive neural feedback mechanisms around a focal area.
Because feedback systems have the ability to become
unstable it would be essential to have local and global
controllers to control levels of neural firing. In a neural
architecture it would be quite likely that these controllers
are actually neural systems themselves (probably using gating
techniques to control neural thresholds and/or synaptic
plasticity).
============================================================
FROM MARK H. BICKHARD AT Lehigh University <mhb0@Lehigh.EDU>
I have been developing a model of representation and
cognition based precisely on the kind of control structures
that you mention in your cogpsy posting. I have also looked
into at least some of the principles of neuromodulation
that you mention.
A control theory framework and model of representation can
be found in:
Bickhard, M. H. (1993). Representational Content in
Humans and Machines. Journal of Experimental and
Theoretical Artificial Intelligence, 5, 285-333.
Discussions of connectionism, brain processes (including
non-local modulation), learning, and many additional
issues, projects and frameworks can be found in:
Bickhard, M. H., Terveen, L. (1995). Foundational Issues
in Artificial Intelligence and Cognitive Science: Impasse
and Solution. Elsevier Scientific.
There is quite a lot of other stuff in this framework. My
own publications are listed in my web pages, and some are
available there (including the Representational Content
paper mentioned above). References are in the book
and in recent publications.
I was following your discussions in the connectionist
community, but didn't join in since I hold that standard
connectionism indulges in a common underlying error with
symbol manipulation approaches when it attempts to model
cognitive phenomena. As models of dynamic processes,
however, including control relationships, connectionist
frameworks have - in my judgment - a great deal of
potential. Such points did not seem to be relevant to the
main discussion on the connectionist net, so I didn't join
in (though I did at one point email you directly). Your
statement to the cognitive science community, however, is
much closer to what I argue for than I had realized
from your connectionist postings - thus this reply to you.
============================================================
FROM REZA FARIVAR AT University of Victoria <rezaf@UVic.CA>
Your point regarding the final "controller" is a very
interesting and fundamental one. It elevates the problem
of the environment shaping the neural connections, because
it does not ask how the environment achieves such a task,
but rather how is the task achieved at the neuronal level.
I too have been searching for some answer on this, because
I feel that any controller must be operating
extra-neuronally, serving as a feedback mechanism by which
a neuron can learn that it has made the proper connections
and the proper response has been achieved. My reasoning for
why it must be extra-neuronal is that anything
intra-neuronal will result in even more processing that has
not been controlled. What I mean here is that the process
by which a neuron in the higher layers may be teaching the
neurons in the lower layers that a certain connection is
invalid, this process itself is open to training, and thus
requiring a trainer or controller. As a result, the
controller must be extra-neuronal.
Some biological phenomena do occur at the neuronal
level that allow for this sort of feedback to occur. It is
unknown, however, to what extent they do this task. One
item that I recall reading is the creation of calcium waves
following the activation of a neuron. If I recall
correctly, when a neuron becomes activated, a fluctuation
occurs in the level of Ca2+ in the glial cells that
surround the neuron. This fluctuation actually follows a
very specific mathematical pattern. It is helical and its
periods can be calculated. To further my speculation,
Ca2+ can have a multitude of effects on a synapse, either
inhibitory, or excitatory, or otherwise. Thus the flux
observed in Ca2+ levels may have the ability to control
neuronal functioning. To what extent this fluctuation may
be able to control neuronal activity is not fully known and
it may be impossible to know, because glial cells (in which
these fluctuations are observed) are extremely small and
very difficult to observe in action.
============================================================
FROM TALIB HUSSAIN AT Queen's University
<hussain@QUCIS.QUEENSU.CA>
YES! I was very pleased to see the argument as you
presented it. It agrees with the core aspects of my
research, which I will be presenting later this month:
Hussain, T.S. and Browse, R.A. (1998) "Including control
architecture in attribute grammar specifications of
feedforward neural networks", to be presented at the 4th
Joint Conference on Information Sciences (Oct 23-28,
Research Triangle Park, North Carolina).
I have been developing a common representation
framework for multiple neural network models. One guiding
reason behind this has been the need to develop a genetic
representation of neural networks that will allow the
evolutionary process to explore among different learning
behaviours as well as network topologies. A second reason
is that I have always found neural network models to be
very haphazard in design. Many aspects of the network
behaviour are actually hidden in the description and those
hidden aspects often involve the manipulation of global
information. Thus it is often difficult to compare two
models as "neural" networks since it is difficult to
identify their common assumptions.
My primary philosophy has been that a neural
network specification should be purely local and consist
only of neurons (of varying type) and their connectivity.
Thus, the learning behaviour and how and when signals are
processed should be all expressed in terms of neuron
behaviours. A desirable consequence of such a
specification framework is that the genetic code of a
network will only require a single, generic interpreter,
but still be able to produce networks which exhibit
learning and processing behaviours available in a variety
of existing models, and possibly even hybridize those
behaviours.
The result of my work has been the development of a
representation framework in which the "control" aspects of
the network are explictly included in the network
specification. To accomplish this, I have relaxed only one
assumption of the "fundamental" neural definition that most
practitioners follow --> My basic model of a neuron
permits a neuron to output signals of different and
multiple types. In other words, it may have more than one
"axon", where each axon sends a signal of a unique type.
This simple change permits the development of neurons
which output signals that may be used by other nodes as
pre-requisites for performing certain actions. The
remainder of my model is purely local. Neurons have access
only to the information in the signals they receive and
their internal memory. In particular, neurons have no
knowledge of the network topology (i.e., where their input
signals came from and where their output connections are
going).
To briefly give the idea of how this may be applied to
implement a complete neural model, consider a network of
nodes in which learning may take place only after
"activation" has settled. Let us consider two types of
signals. 'A' for activation and 'L' for "perform learning".
We may develop a "stable_detector" node which accepts 'A'
signals from all the other nodes and outputs an 'L' signal
only if all those signals are identical to the ones it
received the previous time cycle (and has stored in local
memory). Those 'L' signals are sent to all the other nodes,
which use it as a pre-requisite to initiate learning.
I have developed a network simulator which allows the
developer to specify nodes with arbitrary internal
functionality and networks with arbitrary topology. The
simulator will generate and execute the network with only
minimal I/O control assumptions. To date, I have only
developed a specification for back-propagation and Kohonen
networks. However, new specifications are easy to
implement. (The hard work is trying to extract the hidden
control behaviours in current specifications)
I have obviously tackled this problem from a
computational point of view, and make no strong claims as
to biological plausibility. However, it is clear to me
that the biological neural system is physically expressed
in terms of purely local mechanisms (theories of quantum
consciousness aside), even if some of those mechanisms may
have wide reaching effects. It is my hope that the
exercise of developing explicit control sub-systems and
following strict local information encapsulation will lead
us to the consideration of new ideas that will provide us
insight into biological control mechanisms.
============================================================
FROM MIKE KEITH AT ROCKWELL <mjkeith@ra.rockwell.com>
I think you need to be a little more precise about your
terms. What specificallly makes something a controlling
function ? What makes something NOT a controlling function?
How does our model of an individual neuron tie into the
above? How does our model of a network (1 or more layers or
other configurations) tie into the above?
============================================================
FROM AARON HOSFORD at University of Texas at Dallas
<hosford@utdallas.edu>
I would first like to note that when you mentioned
backpropagation as using external control in its
model, you were confusing the model with the modeller. The
structure of a neural network could just as well be built
directly into the chip, and I don't think such a network
would be controlled in the sense that you have used. The
structure of a neural network under backpropagation is not
something that is necessarily altered or improved as the
network learns, and if it remains static as the network
learns then there is no real control happening.
However, I do agree with your other examples and with your
stated idea in general. I have often supposed that emotion
guides learning in the brain. As far as I know (I am not
well educated in the actual functioning of the
brain), emotions are triggered by the release of chemicals
from particular areas of the brain which are specialized
for that function. My conception is that, upon
positive/negative emotional response to a stimulus,
recently active connections between neurons will be
strengthened or weakened by the influence of those
neurochemicals which were released by the emotional
response. As for memory, when a particular event is
simulated in the brain, its appropriate emotional response
is stimulated, thereby increasing the effect upon neuron
connection strength. I imagine this greatly improves the
rate of learning in the brain, but it is not completely
necessary for learning to occur.
As an illustration, imagine yourself experiencing a mildly
traumatic event -- formatting your hard drive for example.
You are put into a bad mood and your thoughts are simply a
repetetive cycle through the memory of the event. The
neurochemicals responsible for the bad mood remain in your
brain for a period of time as the patterns of brain
activity leading up to the event are recreated, each time
lessening the strength of the connections which were part
of that pattern. At some point the neurochemicals are
eliminated and you resume normal brain activity.
==========================================
APPENDIX: BRAINS INTERNAL MECHANISMS - THE NEED FOR A NEW
PARADIGM
Perhaps a more careful observation of living objects
as a physical system is needed to get a deeper insight on
how the brain actually works. Unlike other physical
systems, like a hurricane or an ocean or a volcano, some
living objects are quite unique in the sense that they
possess a physical entity called the "brain" which no other
systems have. And one of the most unique characteristics of
these systems is that the brain "controls" the behavior of
other entities in the system, such as the movement of its
various limbs, the eye movements and so on. Some scientists
might argue that it is equally valid to say that the limbs
inform and therefore control the brain. But saying that the
limbs inform and therefore control the brain would be the
same as saying that a car or a train or an airplane
"controls" its driver or pilot instead of the other way
round. Or that a nuclear power plant "controls" its central
operating station and its operators, and not the other way
round. Or that a country's central bank does not "control"
its economy by setting the interest rates (just tell Alan
Greenspan that he is not in charge and you will have no
place to hide). This is not to argue that there is no
substance to this viewpoint - one is fully justified in
viewing all feedback control systems in this manner, since
the control signals are essentially a function of the state
of the system. However, this viewpoint is not legitimate
for understanding the internal mechanisms of any such
system, because it overlooks the fact that the "control
signals" are actually determined by and originate from a
certain subsystem (a driver, a pilot, a control station, a
central bank) of the overall system and that these are the
signals that cause the state of the system to change. Such
a viewpoint also overlooks the fact that there are no other
entities in the system capable of computing and generating
those particular "control" signals.
The main argument above is that certain living objects, and
humans in particular, are physical systems very unlike a
hurricane or an ocean or a volcano. When these living
objects are observed from the point of view of their
various subsystems, the "controlling" function of the brain
is obvious; there exists controlling subsystems within
these system. However, there are no such "controlling"
subsystems in a hurricane or an ocean or a volcano or
similar physical systems. The statements made here do not
constitute any profound new scientific discovery by any
standard. They are a simple restatement of some ordinary
facts of biology and nature.
Some very logical questions that arise naturally are as
follows: Could there exist similar "controlling"
substructures within the brain? In other words, could there
be parts of the brain that in some sense guide, govern and
control other parts of the brain? Could it be that synaptic
strengths of cells are actually controlled by other
entities within the brain than the cells themselves? If a
"controlling" subsystem can be observed at a higher macro
level in these living objects, why couldn't they exist at a
lower micro level, in the brain? If one were to propose
such a possibility, it should not be an illogical one; it
would be very consistent with verifiable external
characteristics of these living objects at the macro level.
And I think it is also supported by growing evidence from
neuroscience that extrasynaptic neuromodulators
(extracellular signals from sources external to the cells)
influence synapses, that they can effect permanent changes
in the synaptic strengths and that they originate from
certain fixed sources in the brain (I would love to get
some more insight and recent references from
neuroscientists on this).
Overall, it appears that a very convincing argument can be
made that there are subsystems within the brain that
control other subsystems. This "control theoretic" notion
would thus allow external sources to directly control a
cell's behavior. It would not be fair if it is not
acknowledged here that such control theoretic notions are
already in use, in one form or another, in almost all
connectionist learning systems. For example, the various
constructive learning algorithms, such as adaptive
resonance theory (ART) and radial basis function (RBF)
networks, use non-local means to "decide" when to expand the
size of the network. And the back-propagation algorithm
itself depends on a non-local, external source to provide
it with the design of a network in which to learn. So
connectionist systems inadvertently acknowledge this
"control theoretic" idea, by using a "master or controlling
subsystem" that designs networks and sets learning
parameters for them. In other words, as baffling as
it may sound, the control theoretic ideas have been in use
all along; they are nothing new. However, although these
control theoretic ideas have been in use all along in
connectionism, muddled in with other ideas, only recently
has there been some explicit acknowledgment of the need for
it. For example, Kohonen, in a 1993 paper in Neural
Networks, explicitly mentions "chemical agents, which
are formed or released extracellularly at or in the
neighborhood of highly active cells" in order to provide
the physiological justification for his self-organizing map
(SOM) algorithm.
In addition, a control theoretic framework resolves many of
the problems and dilemmas of classical connectionism. Under
such a framework, learning need no longer be instantaneous
and can wait until some information is collected about the
problem at hand. Learning can always be invoked by a
controlling subsystem at a later point in time. This would
also facilitate understanding the complexity of the problem
from the information that has been collected and stored
already. Such a framework would also resolve the network
design dilemma and the problems of algorithmic efficiency
that have plagued the field for so long. So one can
argue very strongly for such a theory of the brain from
both a computational point of view and from the point of
view of being consistent with externally observed human
learning behavior.