Walter J. Freeman Journal Article e–Reprint
Mind/Brain Science:
Neuroscience on Philosophy of Mind
WALTER J. FREEMAN
CHRISTINE A. SKARDA
1. Introduction
With the advent of behaviorism it became fashionable to claim that neuroscience is irrelevant to cognitive science, and since the introduction of the computer, philosophers of a functionalist persuasion have pointed to the symbolic processing model to support this view. Fortunately, this line of thinking did not stop most of the minds/brains in neuroscience from pursuing research that ran directly counter to the functionalist doctrine, Neuroscientists never believed that knowing how the brain functions is irrelevant to cognitive theory: functionalism looked nice on paper to those who knew nothing about how brains work, but it was based on profoundly anti–biological assumptions which, far from encouraging interdisciplinary exchange, prompted most neuroscientists to ignore what was going on in cognitive science. Today, as functionalism gives way to "neurophilosophy" (Churchland, 1986) and "neural network" models replace some forms of symbol–based processing in artificial intelligence, cognitive and neuroscientists should look at what is being said by their colleagues in both disciplines.
John Searle's work in philosophy of mind
has distinguished itself for many reasons, chief among these being that he
never accepted functionalism and, unlike many philosophers, always
enthusiastically embraced neuroscience. We suggest that his contributions
deserve serious consideration, in turn, by neuroscientists. Of special interest
in this respect are Searle's views on the mind/brain. Our comments focus on
four of these: (1) the relation between brains and computers; (2) the issue of
levels of description; (3) Searle's model for interpreting how the mind/brain
functions; and (4) his characterization of perceptual states. Recent findings
in neuroscience corroborate many of Searle's claims and support his view that
there is no mind/body problem except in the minds of some philosophers. But our
research also suggests that some traditional, time–worn and no longer
valid physiological concepts persist in Searle's characterization of how the
mind/brain works which lead to problems that neuroscience may help to solve. We
address these issues in an attempt to contribute to a future mind/brain science
that is being envisaged by many researchers today, including Searle.
2. Brains and Computers
Searle has presented a forceful and convincing
argument against the view that intelligence is formal symbol manipulation
(Searle, 1980). This view, which we also believe to be false, says that the
brain produces intelligent behavior because at the functional level of description
the brain is a device formally manipulating symbols according to rules. This
view characterizes classical cognitivism; it is still not uncommon to read that
a model that does without symbols and rules at the functional level is
"noncognitive" (Earle, 1987).
Evidence from our laboratory indicates that brain functioning
does not resemble the rule–driven symbol–manipulating processes
characteristic of digital computers (Skarda and Freeman, 1987).
Electroencephalogram (EEG) research on preattentive sensory processing taking
place in the olfactory bulb suggests that brains use the functional
architecture of distributed, self–organizing networks similar in many
ways to some types of present–day connectionist models. The essential
feature of connectionist or so–called "neural net" models is
that a distributed system of interacting simple elements can produce
intelligent behavior without the rules and programs that were previously
thought to be required. Learning takes place by strengthening and weakening
connection strengths between units in the network in parallel. When the network
is activated each unit computes its own level of activity in terms of input
from other units and a predetermined threshold value. The global pattern of
activity resulting from these simultaneous, independent, parallel computations
constitutes the state of the system at each moment.
To support Searle's claim with an appeal
to connectionist systems, however, we believe it is important to distinguish
two camps of connectionist models, one typified by so–called PDP models
(Hinton, 1985; Rumelhart et al., 1986), the other characterized by
self–organizing dynamical systems (Amari, 1983; Anderson et al., 1977;
Freeman, 1987; Grossberg, 1981; Hopfield, 1982; Kohonen, 1984). Not every
connectionist network can be said to operate in the absence of symbols or
symbol–like elements. Systems that fall within the PDP class of
connectionist models use globally distributed dynamics, but there is a sense in
which this class of systems still uses internal "representations" in the
production of behavior. Systems like these, which rely on feed–forward
connectivity and back propagation for error correction, have their
"goals" externally imposed: they require a "teacher" or set
of correct answers to be instilled by the system's operator. These answers are
paradigmatic patterns with reference to which the output of the system is
corrected via back propagation and error correction. The teacher used by such
systems may not be contained in a program, but it nonetheless functions in the same
way that an internal representation does in conventional computers.
Self–organizing dynamic systems, because of dense local feedback
connections, do not require or use teachers. No matching or comparison takes
place such as by correlation or completion, and no archetypal set patterns are
placed by an external operator into the system as its goals.
Our data indicate that in the olfactory
bulb, learning consists in a self–organizing process and not in the
manipulation of symbols according to rules: it is the selective strengthening
of excitatory connections among neurons (mitral cells) in the bulb leading to
the constitution of a nerve cell assembly (NCA) and ultimately to a change of
state from local activity in the bulb to a self–organized form of activity
globally distributed over the entire set of bulbar neurons that can be related
with a particular odor and response. Memory for an odor consists in the set of
strengthened excitatory connections of the NCA that, when activated with input,
possesses the tendency to produce a spatial activity pattern characteristic of
a given odor. These are not mechanisms used by conventional digital computers.
No program–specified rule is imposed on olfactory input, the activity is
self–organized, there is no central processor, and learning and memory
are distributed throughout the system.
On the basis of our data we have
suggested that brains do not work in the way in which everyone expected and
that new conceptual apparatus will be required to characterize adequately the
self–organized (see section 4 below) neural processing that we have
observed. With the introduction of alternative models, the identification of
symbol processing with intelligence appears to be an oversimplification at
best, and we, along with Searle, have argued that it is misleading for both
cognitive and neuroscientists to think in these terms when attempting to
understand brain function (Freeman and Skarda, 1985; Skarda, 1986; Skarda and
Freeman, 1987).
3. Levels of Description and the Neuron Doctrine
We agree with Searle that "[pains] and other
mental phenomena just are features of the brain and perhaps the rest of the
central nervous system," and that the important requirement for
understanding this relationship is the distinction between micro– and
macrolevels of neural functioning, The brain produces intelligent behavior, but
the crucial question for neuroscientists has always been which level of
functioning is relevant for explanations of this behavior? Our research has led
us to break with a foundational concept of contemporary research on the nervous
system, the "neuron doctrine," that we once accepted in company with
the majority of our colleagues, but which we now see as mistaken and as a
source of misunderstanding in attempts to comprehend the brain as the organ of
behavior (Freeman, 1984).
The nineteenth–century witnessed two main revolutions in biology. The best known is the introduction of Darwinian doctrine; the other is the cellular doctrine of Rudolf Virchow (187 1), by which we came to understand that the functional basis of all life is the cell. The teaching and practice of physiology and medicine today are entirely based on this doctrine, and it is undoubtedly correct to say that the great majority of janitorial functions of neurons (processes of growth, repair, general chemical maintenance, metabolic fueling, disposal of wastes, and the modification of connections in molecular and membranal processes) take place and are best understood at the cellular level. The crucial question, however, is whether information used in the elaboration of goal–directed behavior and in controlling interactions with the environment is expressed in or operated on by neurons acting as individual summing and switching devices?
For nearly a century the common answer has been
"Of course it is!" The
principle reason for believing this has been the assumption that what counts in
the production of behavior is the action potential, the most prominent signal
generated by individual neurons (Barlow, 1972). Such a cell is called a
"unit" with a "spike train," and it is commonly thought
that the "message" or "information" transmitted by the cell
consists in the temporal patterns of the spikes, in the manner that a message
resides in the pattern of dots and dashes in a telegraph wire.
We note, however, that the great bulk of physiological
studies cited to support this view come from anesthetized or paralyzed animals
in circumstances in which goal–directed behavior is deliberately suppressed.
When studies have been undertaken in normally behaving animals, the variability
of the observed unit activity has been so great as to be uninterpretable, so
that researchers have resorted to Averaging their data in such a way that the
time base is locked to a stimulus or response. This procedure culls from the
data an extremely small fraction of the variance locked to an external event
and discards the overwhelming majority of the data as "noise."
Our studies of the behavior of single
neurons in respect to the tens and hundreds of thousands of their neighbors
indicates that researchers have been searching for neural
"information" at the wrong level. While the activity of single cells
appears to be largely unpredictable and noisy, the mass of cells cooperates to
produce a coherent pattern that can be reliably related with a particular
stimulus. In studies of the olfactory systems of small mammals the results are
unequivocal: the information expressed by neurons that is related with the
behavior of animals exists in the cooperative activity of many millions of
neurons and not in the favored few. In the first stage of the system where odor
discrimination takes place, the olfactory bulb, there is no evidence that the
signaling by the bulb to the rest of the brain is by any small number of
neurons unique for each odor. On the contrary, for every discriminable odor every
neuron in the bulb participates. We
conclude that it is incorrect to say that the behavioral information exists in
the activity of single neurons; it cannot be observed there. This information
exists in patterns of activity that are distributed concomitantly and
continuously over tens and hundred of millions of neurons.
The main reason that the cellular
doctrine holds so well for most cells but fails for neurons is that neurons are
uniquely different from all other types of cells: neurons are involved
constantly in widespread activity with other neurons, some at great distances.
Each neuron transmits to thousands of others in its vicinity, and in some cases
for astonishingly great distances, often thousands of times greater than the
diameter of its own axon. It also receives from many neurons in its surround,
on the order of 1,000 to 10,000. The key to its interaction lies in this: by
virtue of its membrane each neuron has a relatively high degree of automony,
yet all that it does is felt by many other neurons, and those neurons provide
the environment to which it responds. These widespread actions and reactions
lead to the emergence of the constant, ceaseless, ever–fluctuating
activity of masses of nerve cells that start talking and never stop. In fact,
if they do stop, because they are cut off from interaction with their
neighbors, they soon atrophy and die.
Interconnections among neurons are of two
main types: local and long range, corresponding to Golgi II and Golgi I
neurons. Local connections within areas of the cortex and subcortical nuclei
form the basis for interactive masses that create and maintain fluctuating
patterns of activity over the entire spatial extent of each mass. Distant
connections serve to transmit patterns from one local mass to another and back
again, almost always in reciprocal pairs again subserving interaction and not
merely action. These reciprocal connections form. the basis for feedback loops
that are the hallmark of brain structure and function. The overwhelming
majority of past and contemporary studies of the brain have omitted these
feedback pathways, expressing brain dynamics in terms of flow diagrams that
show inferred connections as all being feedforward. But with the introduction
of new mathematics and technology it has become possible to simulate the
dynamics of neural masses in the brain with large sets of differential
equations and to display the solutions as graphic patterns.
These findings have encouraged us to
break with the "reductionist" view that the behavior of a system can
be explained in terms of the properties and relationships between individual
components that constitute the system. Another feature of brain function is
also important in this connection: brains use chaotic dynamics in the
production of behavior (Skarda and Freeman, 1987). There is a simple recipe for
creating this kind of dynamic activity, and it is one that brains use. (1)
Assemble together a large number of distinct elements such as molecules,
neurons, or even people and allow each to transmit to and receive from many
others in the group information, matter, and energy. (2) Specify that the
relationship between input and output for each element be nonlinear. This means
that if the input is increased in small steps from a low level to a high level,
the output changes but not in direct: proportion to the input. For example, in
the brain if a neuron is stimulated weakly it may release currents but not
action potentials, but once a certain threshold of input is reached it will
fire. This sudden "jump" or change of behavioral state reflects
nonlinearity, i.e., a disproportionately large increase in output for a given
input. (3) Make sure that the system is "open", i.e., there is a
ready supply of energy, food, blood, and a good disposal system for heat and
other wastes. (4) Finally, turn up the temperature, e.g., apply heat uniformly,
or infuse an excitatory chemical. Given these conditions, something interesting
is likely to happen, and it is
likely to be unpredictable. With some initial conditions and
arrangements there will emerge new and sometimes fascinating patterns of
behavior. Some may congeal and terminate the experiment; others may move, rotate,
and reform periodically; but others may dissolve into ceaseless activity
without discernable spatial and temporal structure that resembles
"noise." This ceaseless activity is chaos.
"Chaos," in its traditional
meaning, refers to complete disorder, the formless void from which all order in
the universe arose. The new mathematical meaning for the term refers to
activity that appears random, but is not. It is deterministic, in the sense
that it can be reliably simulated by solving sets of coupled nonlinear ordinary
differential equations or generated by building a system to certain
specifications and putting energy into it. Chaos exists in many forms and
degrees (see e.g., Crutchfield et al., 1987). It can be reproduced with high
precision if the initial conditions are identical on repeated runs, but it is
unpredictable if new initial conditions are used. Chaos has precisely definable
qualities and has relatively small degrees of freedom, meaning that its
dynamics can be described by relatively few variables and therefore dimensions
or coordinate axes. Most important, it can be turned on and off virtually
instantaneously as with a switch.
The physiological basis for the view that
brains employ chaotic dynamics involves a hypothesis on the way in which synaptic
strengths change during learning under reinforcement. Essentially, when two
neurons fire together, i.e., when the action potential of the presynaptic
neuron excites the postsynaptic. neuron and generates an action potential, the
synapse that connects them is strengthened. This is a general form of the
so–called Hebb rule of synaptic learning (Hebb, 1949; Viana Di Prisco,
1984). When many interconnected neurons within a mass of neurons fire together
in pairs over repeated stimuli, the selectively co–active neurons are
joined together into a network of strengthened connections. This conjoined set
is called a nerve cell assembly (NCA), and we believe it is the basis for
perception and learning in the nervous system. When perception takes place it
is expressed in a reproducible and identifiable pattern of activity that is
mediated by the NCA.
When a novel stimulus is presented under
reinforcement, a new NCA forms during the first few presentations. However, the
background activity state of neural activity that goes on in the absence of the
new reinforced stimulus should not be patterned, because this would drive the
system into an already existing pattern of activity associated with another
stimulus. The kind of activity that is required in order that a new pattern can
emerge must be unpatterned yet controllable. This is chaos. It is generated by
the nervous system in the presence of a novel stimulus so that the neurons have
activity by which the Hebb rule can operate. Thereby a novel pattern can emerge
from the chaotic state, and the system is not forced into preexisting
"grooves" or patterns of activity. Chaos has been identified in many
areas of the brain (Babloyantz and Destexhe, 1986; Nicolis and Tsuda, 1985;
Freeman and Viana Di Prisco, 1986; Garfinkel, 1983; Skarda and Freeman, 1987)
with the implication that it may provide the basis for flexibility,
adaptiveness, and the trial–and–error coping that make possible the
nervous system's interaction with an unpredictable and ever–changing environment.
The observation that brains employ chaos
to produce behavior is important in the present discussion because phenomena
that are chaotic preclude long–term predictions. Chaotic behavior emerges
from the nonlinear interaction of its parts, and global behavior in the system
cannot be reduced to or deduced from knowledge about the characteristics and
interactions among individual components (Crutchfield et al., 1987). As we have
indicated, it is not individual neurons and their activities that explain or
cause behavior; it is rather the activity produced by masses of neurons that
self–organize to produce new global forms of behavior. As Searle
indicates, "just as we need the micro–macro distinction for any
physical system, so for the same reasons we need the micro–macro
distinction for the brain" (Searle, 1984).
But our agreement with Searle is
predicated on an important caveat, namely, that he means the same thing we do
by micro/macro descriptions, It isn't always clear that he does mean what we
mean. We agree with Searle when he says that while it makes sense to say that a
particular organism with its nervous system is experiencing a given stimulus,
it does not make sense to say for any particular neuron in that brain that it
is experiencing that stimulus (Searle, 1984). The brain, on our view, gives
rise to emergent neural phenomena that are responsible for and explain behavior
and that are not reducible to the features and relations of its component
parts. But sometimes Searle appears to make a further claim, another kind of
level distinction, that is not encompassed by the one we have discussed. This
distinction is between a microlevel that includes all levels of neuronal
processing and a macrolevel of purely mental processes. Searle says, "At
the higher level of description, the intention to raise my arm causes the
movement of the arm. But at the lower level of description, a series of neuron
firings starts a chain of events that results in the contraction of the
muscles" (Searle 1984).
There are two things to say about this claim. First,
it by–passes the role played by global neural activity and conflates the
global neural and mental levels. Second, although Searle argues against
dualism, he seems to argue here for a level of activity that plays a causal role
but is not physiological, a specifically "mental" level. This is
philosophically appealing but lacks biological sense. The reason is that a,,
physiologists we cannot make strict causal inferences from the level of neurons
to that of neural mass actions (see above); a fortiori, we cannot impute cause
and effect between the global neural and mental levels. Quite apart from the
classical problems concerning causality raised by Hume, Kant, Whitehead, and
others, our modern conceptions of feedback necessarily introduce ambiguity and
indeterminacy. These are endlessly compounded in our efforts to comprehend
distributed networks with large numbers of feedback loops. Already we can build
models of brain parts that function in some of the ways that brains do. We can
describe then and largely (not entirely) control them, but we cannot explain
how they work (Hopfield and Tank, 1986). It appears to us likely that during
the next decade or so, machines will be constructed that will display useful
traits heretofore restricted to biologic intelligence, and the irony will be
that we will be unable to understand their processes in causal terms. The
problem this raises for Searle is that where he want a causal explanation there
isn't one to be had.
4. The Reflex Model of Behavior
Our data have forced us to reject yet another
foundational concept implicit in much of contemporary physiological research:
the reflex doctrine. Physiologists an psychologists who work with animals often
have the illusion that they control their behavior by use of reinforcement.
This belief is based on the model of physiological functioning developed to
explain reflex behaviors (Sherrington, 1906) and on thee feedforward models
that experimentalists use to explain how it is that a conditioned stimulus (CS)
that is paired with an unconditioned stimulus (UCS) will elicit conditioned
response (CR) preceding the unconditioned response (UCR). The presumption is
that all behavior can be expressed as a sum of responses to stimuli, a view
that includes and is ultimately derived from such fundamental behaviors as the
slaking of thirst, the satisfaction of hunger, and the titillation of sex.
What experimentalists have failed to note
is the essential fact that in the typical experiment it is the animal that is
controlling their behaviors: researchers spend, as they should have spent,
small fortunes on the care, feeding, and housing of the subjects; they tailor
the equipment and tasks to the capabilities of the species, familiarize and
train them, and then sit waiting for them to deign to stop eating licking,
grooming, or just looking around long enough for the experimenter to get is a
CS for a controlled trial; all this can go on for weeks. What is lost in all
this is the fact that these animals are continually producing behaviors from
within by anticipating external stimuli to guide or pace their actions. These
behaviors express internally generated activity of the nervous system and are
not determinist responses to stimuli.
There is a term to describe such internally generated activity. We say that this type of neural activity is "self–organizing". We see self–organizing activity in many inanimate systems around us: in the emergence of patterns of clouds in a previous clear blue sky, in the bubbles that form at the bottom of a heated pan of water, and in the formation of drops of water from a leaky faucet. In the biological sphere we see self–organized dynamics in the earliest stages of development of structure and function from the fertilized ovum on into the growing embryo. Recent findings indicate that behavior can arise from within the system by self–organized patterns of neural activity in interconnected masses of the brain, and that it is not simply the sum of conditioned or unconditioned responses to stimuli (Freeman and Skarda, 1985; Skarda and Freeman, 1987).
In the olfactory bulb,
self–organized patterned activity is essential for odor recognition and
discrimination, processes that cannot be explained in terms of the reflex
model. Molecules of the odorant fall onto the cilia of olfactory receptor cells
and, following capture, excite a small subset of receptors that are selectively
sensitive to the odorant. This step is called transduction. In the next step of
forward excitation the receptors send action potentials into the olfactory bulb
that excite the large projection neurons known as mitral cells. Those that have
been previously excited by the odorant under reinforcement, and have had their
synaptic interconnections strengthened by the learning process described
earlier, preferentially excite each other. This subset of mitral cells forms a
nerve cell assembly of strongly interconnected cells. In the third stage of
feedback interaction the initial excitation spreads like wildfire through the
assembly. Those other mitral cells that have been excited, although not under
reinforcement, but that receive excitation from background odorants, excite the
bulb further. Both kinds of excitation serve to increase the sensitivity of
bulbar neurons to each other and therefore to increase intrabulbar crosstalk.
When a critical threshold is reached the fourth step occurs. This is a state
change via a self–organized dynamic process in which the entire bulb goes
from chaotic activity to an internally generated burst of oscillatory activity
that we can correlate with a particular odorant. This globally distributed,
self–organized, stereotypical pattern of activity in the bulb is the one
that we hypothesize constitutes odor recognition and memory for the rest of the
system; this is the pattern that is made available to the rest of the brain and
that is behaviorable relevant for the kind of relations that we usually
associate with learning and memory (Skarda and Freeman, 1988).
This feature of nervous system
functioning, i.e., that it creates something that did not exist before its
interaction with the environment, has not gone unnoticed in the past, but
before the introduction of the mathematical theory of nonlinear dynamics and
computers there seemed no viable alternative explanatory framework to stimulus
–response determinism in experimental brain science. Early on,
Sherrington, who developed the synaptic basis for the concept of the reflex
are, uncovered an internally generated neural process that acted on stimulus
input but was not caused by it (Sherrington, 193 1). Sherrington suggested that
this "central excitatory (inhibitory) state" was internally generated
by a mass of neurons that then serve actively to coordinate and integrate
reflexes that, in themselves, were produced passively by a stimulus.
But physiologists were not alone in
recognizing that neural dynamics are more than reactions to external stimuli.
Philosophers, too, have realized that the reflex model of physiological
functioning is inadequate for explaining behavior. Dewey (1896) can be credited
with probably the earliest attack on the notion. As Dewey put it, the stimulus
"must already have one foot over the threshold, if it is ever to gain
admittance" (Dewey, 1896), and this is provided by a internally motivated
activity.
Later Merleau–Ponty (1942) argued that behavior, and hence a
neural state generated from within the organism, is the first cause of all
stimulation citing the very same data used to support reflex theory by Pavlov
(1927) and Sherrington. Perception, on this view, begins within and is not
imposed from without by the stimulus, If the opposite view based on the reflex
theory dominates our thinking, this is not because we have been convinced by
the data but because we interpret all the data to fit the reflex theory we have
about how perception must work. At best, Merleau–Ponty suggested, the
activity described by classical reflex theories is pathological.
The reflex as it is defined in the classical conception does not
represent the normal activity of the animal, but the reaction obtained from an
organism when it is subjected to working as it were by means of detached parts
... [it is] characteristic not of the fundamental activity of the living being
but of the experimental apparatus which we use for studying it.
(Merfeau–Ponty, 1942)
Our data support Dewey's and
Merleau–Ponty's positions. Only under reinforcement, i.e., only in
animals who have been motivated to expect certain odorants, does receptor
activity initiated by the stimulus lead to changes in the patterned activity of
the bulb, the generation of a NCA, etc. In the neural system, if previous,
internally generated activity has not laid the groundwork for interaction with
the environment, receptor stimulation cannot and will not lead to perception
and/or behavior (see also section 5).
We mention this because very often it is
overlooked by contemporary philosophers interested in understanding the
physiological basis of behavior (Skarda, 1986). The tendency is to adopt a
reflex–based model and to generalize it for all behavior. We believe that
this move is not only bad physiology; it is misleading for philosophical
attempts to understand the biological basis of behavior and the mind/ brain
relationship (see section 5 below and Skarda, 1986). We think that Searle falls
into this trap by focusing on the neurophysiological basis of behaviors that
are reflex in nature, e.g., the slaking of thirst or the contraction of
muscles, and then adopting this model for all brain function. If reflexes were all
that were needed to produce behavior as we know it, we would not find brains
generating self–organized behavior and chaos; and certainly, by focusing
attention on reflex behavior Searle misses, along with many others, the most
distinctive feature of brain functioning in the production of
goal–directed behavior, i.e., its self– organizing, creative
dynamics. We encourage philosophers and cognitive theorists to take another
look at what neurophysiologists have discovered about the neuronal processes
responsible for behavior, and we hope that in the future more attention will be
given to the excellent work done by distinguished thinkers like Dewey and
Merleau–Ponty who pointed out long ago that there is more to behaving
than reacting.
5. What's in Perceptual States?
In a recent book on the philosophy of mind Searle
discusses what he terms the "intentional contents" of perceptual
states (Searle, 1983). He describes these contents as follows:
vis exp (that x with certain features is before me and that x is causing this perceptual experience)
Whatever
else is packed into this description, it implies that what counts as far as
perception (in this case visual perception) is concerned, i.e., the essential
features of the perceptual process, are an internal representation of the
object perceived along with its features and the causal impact that the object
has on the system perceiving it. What does neurophysiology have to say about
Searle's description?
The first thing to note about Searle's
description of perceptual states is that it incorporates the reflex model
discussed above in section 4. Studies of physiological functioning, however,
reveal that perceptual processing involves more interesting processes than
those described by traditional reflex–based theory and feedforward
processing have imagined. Evidence indicates that when an organism is trained
to respond to a particular odorant a self–organized process in the bulb
produces a spatially coherent activity state that can be modeled as a
limit–cycle attractor, Such attractors mathematically represent the
qualitative form of behavior exhibited by a system, in this case periodic or
aperiodic (chaotic) behavior. Topologists say that the behavior is governed by
a stable attractor if the system returns to the same form of behavior after it
is perturbed and is allowed to settle. With each inhalation, after learning and
in the presence of this odorant, this more ordered state repeatedly emerges
from the chaotic background state, only to collapse back again with exhalation.
A separate spatial pattern of periodic behavior forms for each learned odor
given under, reinforcement. Each has a latent coexistence in that only one at a
time can find expression or be realized and then only, as far as we have yet
been able to detect, on presentation of its odorant under reinforcement (we
have not sought to explore hallucinations). When the reinforcement contingency
is changed in respect to any one odorant, or if a new odorant is added to the
repertoire under reinforcement, all the spatial patterns undergo small changes
during the process of learning. These changes do not occur in the olfactory
bulb if there is no reinforcement or if the newly learned CS is not olfactory
but visual or auditory. These spatial patterns are manifested in the bulbar EEG
and their information content is measured by assaying the capacity of our
measurements on these patterns to classify correctly the bursts in respect to
what CS was given and what CR occurred on successive conditioning trials
(Freeman, 1987).
Several features of the neural dynamics
underlying perception are important to note. First, only when the odorant is
reinforced leading to formation of a CR, i.e., only when the animal is
motivated and the stimulus input has some meaning for the organism such that it
acts on the stimulus, do odor–specific activity patterns form in the
olfactory bulb. Presentation of odorants to the receptor surface in unmotivated
subjects does not lead to any observable changes in the system. Second, the
odor specific activity patterns are dependent on the behavioral response:
changing the reinforcement contingency changes the activity patterns previously
recorded. Finally, the internally generated odor–specific activity is
context– dependent: introducing new reinforced odorant to the animal's
repertoire leads to changes in the activity patterns of all previously learned
odorants; in other words, adding a new odor under reinforcement introduces not
only a quantitative change in the number of learned activity patterns, it also
qualitatively alters each of the patterns previously learned.
How do these findings bear on Searle's
claims regarding perception? As mentioned above, Searle's characterization of
perceptual contents implies that what is important for the system are the
object with its features and the object's causal impact on the perceiving
subject. He states that '[t]he story begins with the assault of the photons on the photoreceptor
cells of the retina, the familiar rods and cones" (Searle, 1983). Our data
indicate that perception begins with an internally generated neural process
that, by re–afference, lays the ground for processing of future receptor
input. The neural activity patterns that we find related to perception are
indicative of internal states that reflect reliable forms of interaction in a
context. They are sensitive not simply to the presence of an odorant, or to the
response, but to both in interaction and to the context of reinforced odorants
in which this behavior is embedded. For Searle, perception is something that
happens to the system when it is acted upon by an object; for us, perception is a process that
occurs only when the organism initiates interaction with its environment.
In review we suggest that two things are
missing in Searle's description: neural interaction and internal context. With
respect to the first, causal impact on the system should not be confused with
interaction in our sense. As we have indicated, causal stimulation at the
receptor level alone does not lead to the formation of odor–specific
activity patterns in the bulb that are the basis of odor recognition and
discrimination. Only when the odorant is reinforced, when it acquires
behavioral significance in terms of a CR, does the internally generated pattern
form. Thus, the story of perception cannot be told simply in terms of
feedforward causation in which the object initiates neural changes leading to
an internal perceptual state. What is missing here is recognition of the role
played in perception by self–organizing neural processes and by the dense
feedback among subsystems in the brain that allow the organism to initiate
interaction with its environment.
Second, internal context is also part of
the meaning of the activity pattern for the system: when a new odorant is added
to the animal's repertoire all previously
learned patterns undergo a change. Searle is aware that context plays a role in
perception as in all intentional states. He terms this context the
"Network" (Searle, 1983), and he claims that each state is embedded
in this Network of other states, The problem is that nowhere within his description of the intentional content of
perceptual states does he include a reference to this Network. For Searle, each
perceptual state is located in a Network of other states, but the presence or
absence of states in this Network does not effectively alter the content of any
one. We have found, however, that neural dynamics are sensitive to, and hence
the characterization of their content must include information on,
interrelationships within a perceptual system or subsystem, because adding to
this Network of perceptual states in the olfactory system leads to a change in
the patterned activity of each and every odor–specific state. Thus, we
support Searle's insight that perceptual states are always located in a network
of states, but our evidence suggests that this network is internal to each state. We believe that Searle is on the right
track, as usual, but the predominantly reflex–based conception of brain
functioning that he has inherited from classical neuropsychology has stymied
him. We suggest that brain theories couched in self–organizing nonlinear
dynamics will provide the keys he and others need to solve the problem of
explaining how intentionality can function with the physiochemical organism.
6. Conclusions
We are greatly encouraged by the attempts of
contemporary philosophers of mind to include neuroscience in debates about the
mind/body problem and the explanation of behavior. It is high time for a
biologically grounded cooperative effort to understand the mind/brain, and
Searle has done much to contribute to this project. We are eager to
participate, for we believe that there will never be a coherent theory of the
mind/brain and its role in behavior as long as philosophers ignore the brain
and as long as neuroscientists shy away from the mind. We hope that our remarks
on Searle's work illustrate the fruitfulness of cooperative efforts, both with
constructive criticism and support based on our data. Without theories
neuroscience has no way to explain its data, but without facts about how brains
actually function philosophical theories tend to hang in the air. It is only
taken together that theories and facts lead to insights and direct future
research. We hope this is the way of the future.
REFERENCES
Amari, S. (1983) "Field theory of
self–organizing neural nets," IEEE Transactions on Systems, Man,
and Cybernetics, SMC–13, 5, pp.
741–8.
Anderson, J., Silverstein, J., Ritz, S.
and Jones, R. (1977) "Distinctive features, categorical perception, and
probability learning: Some applications of a neural model," Psychological
Review 84, pp. 413–51.,
Babloyantz, A. and Destexhe, A. (1986)
"Low–dimensional chaos in an instance of epilepsy," Proceedings
of the National Academy of Sciences of the United States of America 83, pp. 3513–17.
Barlow, H. (1972) "Single units and
sensation: A neuron doctrine for perceptual psychology?" Perception 1, pp. 371–94.
Churchland, P. (1986) Neurophilosophy:
Toward a Unified Science of the Mind–Brain (MIT Press/ Bradford: Cambridge, Mass.).
Crutchfield, J., Farmer, J. Packard, N.
and Shaw, R. (1987) "Chaos," Scientific American 256, pp. 46–57.
Dewey, J. (1896) "The reflex arc concept in
psychology," Psychological Review 3, pp. 357–70.
Earle, D. (1987) "On the
differences between cognitive and noncognitive systems," Behavioral and
Brain Sciences 10, pp. 177–8,
Freeman, W. (1984) "Premises in
neurophysiological studies of learning," in Neurobiology of Learning
and Memory, ed. G. Lynch, J. MeGaugh,
and N. Weinberger (Guilford Press: New York, NY).
Freeman, W. (1987) "Analytic
techniques used in the search for the physiological basis of the EEG," in Handbook
of Electroencephalography and Clinical Neurophysiology, ed. A. Gevins and A. Remond (Elsevier: New York, NY),
vol. 3A, part 2, chapter 18.
Freeman, W. and Skarda, C. (1985)
"Spatial EEG patterns, nonlinear dynamics and perception: The
neo–Sherringtonian view," Brain Research Reviews 10, pp. 147–75.
Freeman, W. and Viana Di Prisco, G.
(1986) "EEG spatial pattern differences with discriminated odors manifest
chaotic and limit cycle attractors in olfactory bulb of rabbits," in Brain
Theory, ed. G. Palm (Springer Verlag
Berlin).
Garfinkel, A. (1983) "A mathematics
for physiology," American Journal of Physiology 245: Regulatory, Integrative and Comparative
Physiology 14: R455–66.
Grossberg, S. (1981) "Adaptive
resonance in development, perception, and cognition," in Mathematical
Psychology and Psychophysiology, ed.
S. Grossberg (American Mathematical Society: Providence, R.I.).
Hebb, D. (1949) The Organization of Behavior (Wiley: New York, NY).
Hinton, G. (1985) "Learning in parallel
networks," Byte 10, p. 265.
Hopfield, J. (1982) "Neural
networks and physical systems with emergent collective computational
abilities," Proceedings of the National Academy of Sciences of the
United States of America 79, pp. 2554–8.
Hopfield, J. and Tank, D. (1986)
"Computing with neural circuits: A model," Science 233, pp. 625–33.
Kohonen, T. (1984) Self– Organization and
Associative Memory (Springer Verlag:
Berlin).
Merleau–Ponty, M. (1942) La Structure du
Comportement (Presses Universitaries
de France: Paris).
Nicolis, J. and Tsuda, 1. (1985)
"Chaotic dynamics of information processing. The 'magic number seven
plus–minus two' revisited," Bulletin of Mathematical Biology 47, pp. 343–65.
Pavlov, 1. (1927) Conditioned Reflexes (Oxford University Press: Oxford).
Rumelhart, D., McClelland, J. and PDP
Research Group (1986) Parallel Distributed Processing: Explorations in the
Microstructures of Cognition. Vol.
L Foundations. (MIT Press/Bradford:
Cambridge, Mass.).
Searle, J. (1980) "Minds, brains, and
programs," Behavioral and Brain Sciences 3, pp. 417–57.
Searle, J. (1983) Intentionality (Cambridge University Press: Cambridge).
Searle, J. (1984) Minds, Brains, and Science (Harvard University Press: Cambridge, Mass.).
Sherrington, C. (1906) The Integrative Action of
the Nervous System (Yale University
Press: New Haven, Conn.).
Sherrington, C. (1931)
"Quantitative management of contraction in lower level coordination,"
Hughings Jackson Lecture of 193 1, in The Selected Writings of Sir Charles
Shemington, ed. D. Denny–Brown
(P. B. Hoeber: New York, NY, 1940).
Skarda, C. (1986) "Explaining behavior: Bringing
the brain back in," Inquiry 29,
pp, 187–202.
Skarda, C. and Freeman, W. (1987) "How brains
make chaos in order to make sense of the world," Behavioral and Brain
Sciences 10, pp. 161 – 73.
Skarda, C. and Freeman, W. (1988)
"Research of neural dynamics: Implications for models of learning and
memory," in Systems with Learning and Memory Abilities, ed. J. Delacour and J. C. Levy (Elsevier: New York,
NY).
Viana Di Prisco, G. (1984) "Hebb synaptic
plasticity," Progress in Neurobiology 22, pp. 89–102.
Virchow, R. (1871) Die
Cellula?pathologie in ihrer Begriindung auf Physiologie und pathologische
Gewebelehre (A. Hirschwald: Berlin).
END