Part II. The basic model.
Chapter Two
Model 1: Conscious representations are
îinternally consistentï and îglobally distributedï.
"It seems that the human mind has first
to construct forms independently before
we can find them in things ... Knowledge
cannot spring from experience alone, but
only from a comparison of the inventions
of the intellect with observed fact."
--- Albert Einstein (1949)
2.0 Introduction.
2.1 Contrasting the capabilities of conscious and unconscious
processes.
2.2 The basic model: A global workspace (blackboard) in a
distributed system of intelligent information processors.
2.3 How the theoretical metaphor fits the facts of Table 1.
2.31 When is a Global Workspace System useful?
2.32 Cooperative computation: an illustration.
2.33 Spreading activation and inhibition to communicate
between processors.
2.4 Input properties of the global workspace.
2.41 Conscious experience has a strong perceptual bias.
2.42 Temporal limits on conscious integration of
different stimuli.
2.43 The Threshold Paradox.
2.5 Output properties of the GW: How global is global?
‹j 2.6 Further considerations.
2.61 Functional equivalents to a GW system.
2.62 The Mind's Eye, Mind's Voice, and Mind's Body
as aspects of the global workspace.
2.63 What is the global code?
2.64 Other related models.
2.7 Testable predictions and counterarguments.
2.71 Some testable predictions from Model 1.
2.72 More questions for which the Model 1 suggests answers.
2.73 Some counterarguments.
2.74 Other unanswered questions.
2.8 A summary of the argument so far.
‹"
2.0 Introduction.
Almost everything we do, we do better unconsciously than
consciously. In first learning a new skill we fumble, feel
uncertain, and are conscious of many details of the action. Once
the task is learned, sometimes after only a few repetitions, we
lose consciousness of the details, forget the painful encounter
with uncertainty, and sincerely wonder why beginners seem so slow
and awkward. This pattern holds for everything from walking to
knowledge of social relations, language acquisition, reading, and
the skills involved in understanding this book.
These observations imply that we are unconscious of the
complexity of whatever we know already. This is clearly true for
high-level skills, like the reader's ability to process the
syntax of a sentence. To grasp the meaning of a sentence, at
least part of its syntax must be analyzed. The first small step
in syntactic analysis involves assigning parts of speech to the
words --- nouns, verbs, pronouns, adjectives, and so on. Trying
to do this deliberately, consciously, without paper and pencil,
takes a great deal of time; it is a great burden on our immediate
memory; it is prone to error; and it interferes with other
mental processes that are needed to understand the material.
Conscious sentence parsing is hopelessly inefficient. But
unconsciously we analyze hundreds of sentences every day,
accurately and gracefully. This is true for all the skills that
enable us to navigate through everyday life (La Berge, 1974;
Shiffrin & Schneider, 1977; Langer & Imber, 1979; Pani,
1982). Any task improves with practice, and as it becomes more
efficient it also becomes less consciously available. Thus
anything we do well, we do largely unconsciously. But then what
advantage is there to being conscious at all?
This chapter will focus on these kinds of questions, based
on a contrastive consideration of the îcapabilitiesï of conscious
and unconscious functions. These Capability Contrasts (Table 2.1)
provide the evidence for the core theoretical idea of this book:
that conscious experience is closely associated with a îGlobal
Workspace (GW) Systemï. A global workspace is an information
exchange that allows specialized unconscious processors in the
nervous system to interact with each other. It is analogous to a
blackboard in a classroom, or to a television broadcasting
station in a human community. Many unconscious specialists can
compete or cooperate for access to the global workspace. Once
having gained access, they can broadcast information to all other
specialized processors that can understand the message.
The properties of a Global Workspace System fit the
empirical Capability Contrasts very nicely, resulting in Model 1
(Figure 2.2). In this model conscious events are simply those
that take place in the global workspace; everything else is
unconscious. We will have to modify this idea in future chapters,
but the basic global workspace metaphor will serve us throughout‹j ‹
the book.
Obviously Model 1 is only a first approximation, but a
fruitful one. We explore its implications in some detail, looking
at both input and output functions for the GW. For instance,
does the claim that conscious contents are "globally distributed"
mean literally that it is broadcast all over the central nervous
system? We cite six sources of evidence in favor of this strong
claim. Next, we point out that there may be systems that behave
like Model 1, but which have somewhat different "hardware". Such
functionally equivalent systems should not be excluded from
consideration. Finally, we review similar proposals made in the
cognitive and neurophysiological literature, make some testable
predictions from the model, and point to some of its limitations
--- evidence that Model 1 cannot handle. These limitations
suggest better models, which are developed in later chapters.
2.1 Contrasting the capabilities of conscious and
unconscious processes.
Table 2.1 presents our first major contrastive data set.
Notice that we are comparing conscious processîesï with unconscious
processîorsï. The reasons for that have been given in the previous
chapter: there is good reason to think that unconscious functions
are modular (1.xx). Note also that the contrasts are not
absolute. Conscious symbolic operations are not totally
inefficient. Rather, in general the more efficiently some mental
operation is handled, the more likely it is to be unconscious,
and vice versa. We will now state each contrast formally, and
discuss it in some detail.
===============================================================
=
Table 2.1
Capabilities of comparable conscious and unconscious events.
îCapabilities ofï îCapabilities ofï
îconscious processes.ï îunconscious processorsï
1. Computationally inefficient. Unconscious specialists
are highly efficient
in their own tasks.
(High errors, low speed, and (Low errors, high speed,
mutual interference between and little mutual
conscious computations.) interference.)
2. Great range of different Each specialized processor
contents over time; has limited range over
time, and is relatively
isolated and autonomous.
Great ability to relate
different conscious contents
to each other;
Great ability to relate
conscious events to their
unconscious contexts.
3. Internal consistency, Specialists are diverse,
seriality, and have can operate in parallel, and
limited capacity. together have great capacity.
===============================================================
=
‹` ‹
2.11 îConscious processes are computationally inefficient,
but unconscious processors are highly efficient in their
specialized tasks.ï
Try to calculate (9 x 325)/4, doing each mental operation
completely consciously. Or try to "diagram a sentence"
consciously --- assigning syntactic clause boundaries, word
categories like noun, verb, adjective, etc., and deciding on the
subject and object of the sentence. Probably no one can do even
one of these symbolic operations completely consciously. Even
linguists who have studied syntax for many years cannot parse a
sentence consciously. The rare individuals who are extremely good
at mental arithmetic have probably learned through long practice
to do most computational steps automatically, with minimal
conscious involvement.
Compared to similar unconscious processes, tasks performed
consciously are slow, vulnerable to interference from other
conscious or effortful mental processes, and hence prone to
error. Consider each of these characteristics in turn:
The îspeedï of conscious events is relatively slow. Simple
reaction time (the time needed to give a single known response to
a single known stimulus) is at best about 100 milliseconds. This
is also the time region in which we experience perceptual fusion
between physically different stimuli, and Blumenthal (1977) gives
seven arguments in favor of the idea that the minimum "conscious
moment" is ranges around 100 milliseconds. In contrast,
îunïconscious processes may take place at the speed of neural
firing, ranging from 40 to 1000 times per second. In speech, when
we say "bah", the vocal cords begin to vibrate before the lips
open; when we say "pah" the order is reversed. The difference in
this voice-onset time between "pah" and "bah" is about 20
milliseconds, much faster than conscious reaction time, and
faster than the minimal integration time discussed by Blumenthal
(1977). But of course we do not consciously control the details
of the /pa/-/ba/ difference.
Conscious events are vulnerable to interference. Below we
will make much of the remarkable fact that îanyï conscious event
can interfere with îanyï other (2.xx). Perceptual experiences in
any sense modality interfere with those in any other. Any percept
we experience will interfere with any mental image. Any mental
image interferes with any simultaneous emotional or bodily
feeling. Any of these experiences interfere with any voluntary,
effortful action. And anything said in inner speech interferes
with percepts, feelings, images, or mentally effortful actions.
This fact is fundamental.
Unconscious processes, on the other hand, interfere with
each less predictably. We have previously shown the lack of
interference between automatic and voluntarily controlled skills
(1.x) (Shiffrin, Dumais, & Schneider).‹j ‹å
Finally, conscious events are prone to error. Even simple
mental arithmetic is hard to do without error, much less
conscious syntactic analysis, visual scene analysis, etc. This
vulnerability to error is of great practical importance, since
most airplane crashes, road accidents, and industrial disasters
have a significant component of human error. Not all human errors
are due to the limitations of consciousness --- many are due to
the rather different limitations of îunïconscious events, discussed
in Chapter 1 and below (Reason, 1983, 1984). But conscious
processing limitations are surely part of the problem.
By contrast to conscious limits, of course, unconscious
processing of highly practiced, specialized functions is much
more efficient.
Given this catalogue of woe about conscious processes, we
may be tempted to ask, what good does it do? Should we give up
consciousness if we had a choice? Or does it give the nervous
system some selective advantage not provided by unconscious
processes? The answer, fortunately, is yes. Consider the
following points.
2.12 îConscious processes have a great range of possible
contents, but the range of any single unconscious processor is
limited.ï
We can be conscious of an essentially endless range of
possible contents: sensory and perceptual aspects of the world
around us, internally generated images, dreams, inner speech,
emotional feelings, pleasures and pains. If we include conscious
aspects of beliefs, concepts, and intentions, the range of
possible contents becomes even greater. There is good evidence
that we can gain a degree of conscious control over virtually any
population of neurons, provided that we receive immediate
cosncious feedback from the neural activity (Chase, 1974; 2.5).
Put all these things together, and it becomes clear that
conscious contents can be involved in essentially îanyï aspect of
neural functioning. The range of conscious contents and
involvements is simply enormous.
How do we know that îunïconscious processors tend to have
limited range? One consideration is that specialization in
general seems to lead to limitation. If there is an unconscious
syntax processor, it is unlikely to be much good analyzing visual
scenes. In Chapter 1 we cited several action errors collected by
Reason (1983), as evidence for action schemata that are quite
limited in their own ways, as shown by the stereotyped and
mechanical quality of the errors. We can easily avoid these
errors by remaining îconsciousï of what we are doing. Langer and‹j ‹
Imber (1979) have been able to induce mindless behavior by
over-practicing people on a simple task, and found that once the
task has been practiced to the point of being automatic and
unconscious, the subjects can no longer accurately estimate the
number of steps in the task. Further, subjects are much more
willing than before to accept the false inference that they have
performed poorly on the task, even when they have performed quite
well! Obviously automaticity has its drawbacks.
These examples are revealing because they seem to show the
functioning of unconscious components (specialized processors)
without the intervention of conscious control. In each case, this
functioning seems exceptionally "blind" because it seems to
proceed in ignorance of apparently obvious changes in task and
context. The overall pattern supports our basic contention that
"îunconscious processors have relatively limited rangeï".
The whole pattern makes sense if we consider the advantages
and disadvantages of specialization. Clearly the main advantage
of specialization is that one knows exactly what to do in a
particular, routine situation. In computer language, one has a
well worked-out îalgorithmï for solving a particular problem. This
off-the-shelf algorithm is unexcelled for its particular purpose,
but it is likely to be useless for any other. The main drawback
of specialization for routine tasks is a loss of îflexibilityï in
dealing with new situations.
Thus it seems that unconscious processors are excellent
tools for dealing with whatever is known. Conscious capacity is
called upon to deal with any degree of novelty. This leads
directly to the next point.
2.13 îConscious processes have great relational capacity and
context-sensitivity, but unconscious processors are relatively
isolated and autonomous.ï
The terms "relational capacity" and "context-sensitivity"
are used here with very specific meanings. îRelational capacityï is
used to refer to the ability to relate two conscious events to
each other. Classical conditioning provides a good example. Here
one conscious stimulus serves as a signal for another conscious
stimulus --- a bell may signal the coming of food, a light can
signal an electrical shock, and so on. There is no natural
connection between the bell and food, or between a light and
shock. These relationships are arbitrary. Yet under the proper
circumstances, any conscious stimulus can come to serve as a
signal for the coming of a reinforcing stimulus.
What happens if one of the stimuli in classical conditioning
is not conscious? Soviet researchers claim that Pavlovian
association does not occur if the conditional stimulus has become‹j ‹
habituated through repetition, so that it is no longer conscious
(Razran,1961). There is also good evidence to indicate that conditioning
occurs in humans only when they have some consciously
accessible idea of the relationship beteen the two stimuli.
Dawson and Furedy (1976), in a brilliant series of experiments,
used a tone in auditory noise to signal the coming of a moderate
electrical shock. Ordinarily, people learn very rapidly that the
tone signals shock, so that soon a rise in electrical
skin-conductivity occurs, as soon as the subject hears the
warning tone. But now a different group of subjects was given the
identical series of stimuli, and told a different story about the
relationship between the tone-in-noise and the shock. The purpose
of the experiment, they were told, was to see if people can
detect a tone in background noise, and the function of the shock
was only to mark the beginning of a new trial. (Subjects in
experiments seem willing to believe almost anything.) Thus they
were led to believe that îthe shock may signal the coming of a
toneï, not vice versa. Under these conditions classical
conditioning of tone to shock never occurred, even with many
repeated trials. Even though the subjects were conscious of both
stimuli, they reinterpreted the relationship between tone and
shock, and simply never learned that the tone signaled the coming
of an electrical shock.
These findings suggest that for classical conditioning to
occur, subjects must be conscious of îbothï stimuli, and they must
be conscious of the conditional relationship between the stimuli
as well. If either of these components is lacking or unconscious,
classical conditioning does not seem to take place. Only
conscious functions seem to have the relational capacity to bring
together two arbitrarily related stimuli; unconsciously we cannot
apparently relate two novel, arbitrary stimuli to each other.
But consciousness has more than this kind of relational
capacity; it also facilitates îcontext-sensitivityï.
"Context-sensitivity" is defined here as the way in which
conscious events are shaped by unconscious factors. There are
numerous examples of this (see Chapters 4 and 5). Perhaps the
most obvious ones come from everyday examples of carrying out a
routine action. When driving a car, we may take the same route
every day, so that the predictable actions needed to drive become
less conscious over time. If we something new happens on the
route from home to work, previously unconscious elements must
become more conscious to adapt to the new situation. If we
resolve one day to drive to the grocery store on the way home, we
may suddenly find ourselves already home without having gone to
the store, because we failed be conscious of our goal at a
critical intersection. Similarly, even if we know ahead of time
that the road is blocked along our familiar route, that knowledge
must become conscious in time to make the appropriate decisions
to drive another way. In general, changes in context are not
encoded automatically; they require consciousness. But once‹j ‹
contextual information is encoded, it may control our routine
actions and experiences without again becoming conscious.
Perception textbooks are filled with examples in which our
conscious experiences are profoundly shaped by numerous
unexpected unconscious factors (Gregory, 1966; Hochberg, 1964;
Rock, 1983). For example, we live in a "carpentered" world, a
world of rectangular surfaces and square corners. But we usually
look at the surfaces in this world aslant, so that our eyes
receive îtrapezoidalï projections, not rectangular ones (Figure
2.13). Each of these trapezoidal projections can result from an
infinite set of rectangles or trapezoids, placed at different
angles to the eye. What would happen if we were to look into a
space that was made up of trapezoids, positioned in such a way as
to cast the same retinal projections as a normal carpentered
room?
Adelbert Ames (1953) first tried this experiment some fifty
years ago, and found that people see the distorted space as a
normal, rectangular room. The walls in a trapezoidal room are not
of constant height, even though they seem to be constant, and it
seems likely that the height of other objects is scaled relative
to the nearest wall. What would happen if we observed someone
walk back and forth in the Ames distorted room? The person is not
changing height, while the walls, which seem of constant height,
do change. Hence there is a perceptual conflict between the fact
that human height does not change quickly, and the fact that
walls are assumed not to change at all. The upshot is quite
remarkable: people appear to grow and shrink dramatically as they
walk to and from the observer (Figure 2.13).
------------------------------
Insert Figure 2.13 about here.
------------------------------
As they walk toward the short end of the trapezoidal wall, their
size in comparison to the perceived height of room may double,
and as they walk toward the tall end of the trapezoid, they
shrink in comparison. But why do we not see the room's actual
proportions, and keep the perceived height of the people
constant? For some reason the visual systems seems "committed" to
seeing the room as constant in height, and as a result, its only
option is to interpret the person's height as changing. Clearly
our conscious experience of the person in the Ames room is shaped
by unconscious assumptions about the space in which he or she
appears.
There are numerous other examples of this sensitivity of
conscious contents to unconscious context.
Our ability to comprehend a sentence in a conversation
depends in great part on whether the new information in the
sentence fits into what we take to be given in the conversation
(Clark & Haviland, 1977; Chafe, 1970). But when we hear the new‹j ‹
information, the givens are already unconscious: again, the
unconscious context helps to shape the novel, conscious
information. Our ability to learn any new information is
critically dependent on prior, largely unconscious knowledge
(e.g. Bransford, 1979).
Scholars who study îchanges or differencesï in knowledge are
often acutely aware of the effects of unconscious presupposed
context. An anthropologist studying a new culture is often
forced to confront his or her own unconscious presuppositions,
which may become become conscious only in the encounter with a
social world that violates them. And historians are well aware
that each new age reinterprets the "same" past in accordance with
its own presumptions, most of which are quite unconscious at the
time they have this effect. Chapters 4 and 5 consider these
context effects in detail.
All these examples indicate that unconscious expectations
guide our conscious appreciation of the world. This is quite
different from the "relational capacity" defined above, which
involves relating two îconsciousï events to each other. Context-
sensitivity, as we use the term in this book, implies that all
conscious experiences are constrained by unconscious context.
The contrasting claim about comparable unconscious events is
that "unconscious processors are relatively isolated and
autonomous." It is the unconscious processors that are presumably
responsible for the very smooth and efficient actions cited in
the action errors above, which are carried out perfectly well,
except for the fact that they are wildly inappropriate to the
circumstances. These errors are often amusing because of the
inappropriateness of the isolated action, which may be carried
out perfectly even though its relevance and purpose are utterly
lost.
Action errors all seem to involve either a failure to adjust
to a change in the physical situation, or a loss of the current
task context. Getting up on a holiday and dressing for work is an
error that involves a failure to access a new context. It seems
that routine activities run off automatically, and adjusting to a
new situation demands some conscious thought. Taking a can opener
instead of scissors to cut some flowers seems to involve a loss
of the current task context --- we have "forgotten what we are
doing".
2.14 îConscious experiences have internal consistency, but
unconscious processors may be mutually contradictory.ï
We have already pointed out (1.xx) that selective attention
always involves a densely îcoherentï stream of events. We never mix
up two streams of speech with different contents, or even with‹j ‹
different vocal quality. It is generally true that conscious
experiences are internally consistent. For example, the Necker
Cube shown in Figure 2.14 can only be seen in one way at a time;
each conscious interpretation is internally consistent. We never
see a mix of the two conscious interpretations. For instance, we
never see corner (a) in a different depth plane than corner (b),
because to do so would violate the consistency constraints of a
rigid, square cube.
These phenomena are well©known in perception, but they are
not limited to perception. The same things are true at the
conceptual level. Social psychologists for some decades have
investigated cognitive consistency in value judgments and in
person perception. Here, too, internal consistency is maintained
(e.g. Abelson et al, 1968; Festinger et al, 19xx). We cannot
think of two alternative ideas at the very same instant, though
we can consider two contradictory ideas one after the other. This
becomes very clear when we consider ambiguous words: most words
have at least two different abstract, conceptual interpretations.
It seems impossible for people to entertain two meanings of words
like "turn," "look," or "book," at the same instant.
By contrast to conscious consistency, unconscious processors
working at the same time may be mutually inconsistent. There is a
great deal of evidence, for example, that the unconscious meaning
of an ambiguous word is represented in the nervous system at the
same time as the conscious meaning (Tannenhaus, Carlson &
Seidenberg, 1985).
On to the next claim.
2.15 îConscious processes are serial, but unconscious
processors can operate in parallel.ï
There is much evidence for the seriality of conscious
contents, but it is difficult to prove that the seriality is
absolute. Conscious experience is one thing after another, a
"stream of consciousness" as William James called it.
Psychological theories that are largely confined to conscious
processes, such as Newell and Simon's (1972) theory of human
problem solving, postulate largely serial mechanisms.
Automaticity shows the close relationship between
consciousness and seriality. As a skill becomes more and more
practiced, it becomes less and less conscious; it can then also
begin to operate independently from other processes, just as a‹j ‹
parallel processors does (Shiffrin & Schneider, 1977; Sternberg,
1963; LaBerge, 1981). Conversely, when we interfere with an
automatic skill so that it becomes "de-automatized," it will be
more conscious, and it will be slower and more serial as well.
However, at very fine time resolution, say the level of
milliseconds, the seriality of conscious processes is not so
clear. Just as a serial digital computer can simulate a parallel
system simply by switching rapidly back and forth between
different processes, so it is possible that some apparently
parallel events are really controlled by a serial system
(Anderson, 1983). For these reasons it is difficult to be
absolutely sure about the seriality of consciousness. But it is
clear that over a period of seconds and longer, conscious events
appear to be serial, while unconscious ones seem to work in
parallel.
The claim here is that unconscious processors îcanï operate in
parallel, not that they must always do so. (E.g. Sternberg, 1963;
Banks & White, 1982). Indeed, if unconscious processors are
required for a contingent series of decisions, it is hard to
conceive how they could work in parallel: if A leads to B which
leads to C, then A,B, and C must become available in that order.
Thus the linguistic hierarchy discussed in a previous section may
operate serially when there is no "top-down" information, even
though the hierarchy is largely unconscious.
Further evidence for parallel unconscious processing comes
From neurophysiology. As Thompson (1967) remarks, t
organization and functioning of the brain "is suggestive of
parallel processing". Many areas of the brain are active at the
same time. Within the past few years, mathematical models of
parallel processing have become available that cast light on the
ways in which many of these neural systems could work, and
several systems have been modeled in some detail (Grossberg,
1982; Rumelhart, McClelland & the PDP Group, 1986).
Finally, there are some very important cases, like language
comprehension, where evidence exists that unconscious language
processors act in a "parallel-interactive" fashion (Marslen-
Wilson & Welsh, 19xx). Obviously when parallel processors
interact with each other, they are no longer acting exactly in
parallel (i.e., independently from each other). We suggest below
that consciousness facilitates exactly this kind of parallel-
interactive kind of processing.
But the simplest summary of the evidence is still the claim
that conscious processes are serial, while unconscious processors
can operate in parallel.
2.16 îConscious processes have limited capacity, but
unconscious processors, taken together, have very great capacity.ï
‹j ‹å
We have previously discussed limited capacity in terms of
three phenomena: (1) selective attention, in which one is
conscious of only one of two demanding streams of information to
the exclusion of the other (1.xx). (2) Dual-task paradigms, in
which two conscious or voluntary tasks degrade each other; and
(3) immediate memory studies, in which only a very limited amount
of novel or unorganized information can be retained. All three of
these phenomena are associated with consciousness, though they
are not identical to it.
There is one interesting counter-argument to the notion of
conscious limited capacity, and that is the case of a very rich
perceptual scene. In looking at a football game with a stadium
full of cheering sports fans, we seem to have an extremely
complex visual experience, apparently full of detail, but
apparently completely conscious. The key here is the internal
organization of the football scene, the fact that each part of it
helps to predict the rest. If instead we present people with an
arbitrary number of small unrelated visual objects, and ask them
to estimate the number in a single glance, visual perceptual
capacity drops down again to about four to six items (ref,
subitizing). In addition, we scan even a coherent scene with
serial eye-movements, picking up a relatively small information
with each fixation. Thus the complex scene is not necessarily in
perceptual consciousness at any one time: we accumulate it over
many serial fixations.
Thus conscious capacity does appear to be quite limited, as
shown both by the selective attention experiments and by the
limitations of short-term memory. What about the idea that
unconscious processors "taken together have very great capacity"?
This is obvious just from considering the size of the central
nervous system. The cerebral cortex alone, taking up about half
the volume of the cranium, contains on the order of 55 billion
neurons, according to recent estimates (Mountcastle, 1983). Each
neuron may have as many as 10,000 connections to other neurons.
The interconnections between neurons are extremely dense --- one
can reach any neuron from any other neuron by passing through no
more than six or seven intervening neurons. Each neuron fires on
the average forty impulses per second, up to 1,000 when
activated, and this activity continues in all parts of the brain,
including those that are not currently conscious (Shevrin &
Dickman, 1982).
This is by any standards a very large system. Viewed as an
information processor, it is orders of magnitude larger than
anything built so far by human beings. And clearly, most of its
activities at any one moment are unconscious. Further, Long Term
Memory, which has enormous capacity, is unconscious. The
information processing capacity of all the automatic skills
learned over a lifetime is similarly great. And
neurophysiologically, it is clear that the great bulk of brain
activity at any single time is unconscious. ‹j ‹å
Why does this awesome system have such remarkable
limitations of conscious capacity? There is something very
paradoxical about these differences between conscious limitations
and the huge unconscious processing capacity. Is this paradox a
functional property of the nervous system, or is it somehow a
mistake made by evolution? Later in this book we will suggest
that humans have gained something valuable in return for our
apparently limited conscious capacity (2.x, 10.x).
2.17 A summary of the evidence of Table 2.1.
Before we begin to interpret the contrasts discussed so far,
we will take a glance backwards. If one is willing to accept the
vocabulary of information-processing we apply here, speaking of
conscious and unconscious "representations" and "processes", some
facts can be established very clearly.
Conscious processes are computationally inefficient; they
are relatively slow, awkward, and prone to error. But they
involve an unlimited range of possible contents; any two
conscious contents can be related to each other; and conscious
contents are also profoundly shaped by unconscious contextual
factors. Conscious experiences appear to be internally
consistent; different ones appear serially; and there are rather
narrow limits on our capacity to perform tasks that have
conscious components.
On the other hand, unconscious processors seem to be highly
efficient in their special tasks. Each unconscious processor
seems to have a limited range, and it behaves relatively
autonomously from the others. Unconscious processors are highly
diverse and capable of mutual contradiction; can operate in
parallel; and together have very great processing capacity.
In the following section we will suggest a theoretical
metaphor to explain these observations. This metaphor greatly
simplifies the diverse facts described above, combining them into
only a few basic theoretical properties. Further, it suggests a
functional interpretation for these facts, a selective advantage
for having this kind of nervous system.
2.2 The basic model: A global workspace (blackboard) in a
distributed system of intelligent information processors.
In recent years computer scientists, psychologists and some
neuroscientists have become increasingly interested in‹j ‹
distributed information processing systems --- systems that are
really collections of intelligent, specialized processors. These
systems are a hot topic of research in artificial intelligence
(e.g. Reddy & Newell, 19xx; Erman & Lesser, 19xx), cognitive
psychology (Rumelhart, McClelland & the PDP Group, 1986), and
neuroscience (Arbib, 1979; Grossberg, 1982; Mountcastle, 1978).
They have been used to model the visual system, human memory,
control of action, and speech perception and production. In a
distributed system, numerous intelligent specialists can
cooperate or compete in an effort to solve some common problem.
Together, several specialists may perform better than any single
processor can. This is especially true if the problem faced by
the distributed system has no precedent, so that it must be
handled in a novel way.
In a true distributed system there is no central executive
--- no single system assigns problems to the proper specialists,
or commands them to carry out some task. For different jobs,
different processors may behave as executives, sometimes handing
off executive control to each other in a very flexible way.
Control is essentially decentralized. The intelligent processors
themselves retain the processing initiative --- they decide what
to take on and what to ignore. (In a later chapter we will argue
that the nervous system does have components that act as
executives. But these executives operate in a fundamentally
decentralized environment, much as a government may operate to
regulate a market economy, which is still a fundamentally
decentralized sort of thing.)
But even without a true executive, a distributed collection
of processors still needs some central facility through which the
specialists can communicate with each other. This kind of central
information exchange has been called a "global workspace",
"blackboard", or "bulletin board" (Reddy & Newell, 1974; Erman &
Lesser, 1974; Hayes-Roth, 1984, etc.). A "workspace" is just a
memory in which different systems can perform operations, and the
word "global" implies that symbols in this memory are distributed
across a variety of processors. Each processor could have local
variables and operations, but it can also be responsive to global
symbols. (This is discussed in more detail below.)
-----------------------------
Insert Figure 2.2 about here.
-----------------------------
Analogies will be used throughout this book to make things a
bit more comprehensible. For instance we may speak of the global
workspace as a television station, broadcasting information to a
whole country. There is one especially apt analogy: a large
committee of experts, enough to fill an auditorium. Suppose this
assembly were called upon to solve a series of problems which
could not be handled by any one expert alone. Various experts‹j ‹
could agree or disagree on different parts of the problem, but
there would be a problem of communication: each expert can best
understand and express what he or she means to say, by using a
technical jargon that may not be fully understood by all the
other experts. One helpful step to solve this communication
problem is to make public a îglobalï message on a large blackboard
in front of the auditorium, so that in principle anyone can read
the message and react. In fact, it would only be read by experts
who could understand it or parts of it, but one cannot know ahead
of time who those experts are, so that it is necessary to make it
potentially available to anyone in the audience.
At any time a number of experts may be trying to broadcast
global messages, but the blackboard cannot accomodate all of the
messages at the same time --- different messages will often be
mutually contradictory. So some of the experts may compete for
access to the blackboard, and some of them may be cooperating in
an effort to broadcast a global message.(Indeed, one effect of a
global message may be to elicit cooperation from experts who
would not otherwise know about it. Coalitions of experts can be
established through the use of the blackboard.)
This sort of situation is common in human society. It
describes fairly well the case of a legislature or a committee,
or even a large scientific conference. Clearly this "system
architecture" has both advantages and disadvantages. No one is
likely to use it when the problem to be solved is simple and
well-understood, or when quick action is required. But it is not
a bad way to do things when cooperation between otherwise
separate knowledge sources is required, so that all viewpoints
must be heard, when there is time to agree or disagree over
possible solutions, and when the cost of making a mistake is
greater than the benefits of a quick, "make-shift" action.
Given this brief description, we can now go back to the
facts about conscious and unconscious processes shown in Table
2.1 to see if we have a theoretical metaphor that can simplify
and make sense of those facts.
2.3 How the theoretical metaphor fits the facts of Table
2.1.
If we assume, as a first approximation, that messages on the
blackboard are conscious and that the experts in the assembly
hall correspond to unconscious processors, the fit between the
model and the contrastive analyses in this chapter is quite
close.
Take the first point of Table 2.1: "Conscious processes are
computationally inefficient." Committees and legislatures are not
notoriously efficient in getting things done, because every‹j ‹
action requires at least the tacit consent of many separate
individuals. If something is to be done efficiently it is better
done by a hierarchical organization like a bureaucracy, an army,
or a police force. Committees and legislatures have some virtues,
but speed and efficiency are not among them.
This point applies also to global processes in a large,
distributed nervous system. Any global message is likely to
involve a set of cooperating processors, and at least tacit
cooperation from other processors that could interrupt the first
set. This is very useful in dealing with a novel problem, one
that does not have a known algorithm for its solution.
Information from many knowledge-sources may be combined to reach
a solution. For example, Reddy and Newell (1974), and Erman and
Lesser (1975) developed a distributed system called Hearsay to
deal with the very difficult problem of speech recognition. A
good deal is known about the ways in which sound waves can
represent English phonemes --- but not enough to determine the
right phoneme for every sound. Indeed, as we have suggested
above, it is may be that speech is simply locally ambiguous, so
that there is no unique phonetic solution for every acoustical
waveform. For this reason, Hearsay used a number of distributed
specialists called "knowledge sources" cooperating and competing
by means of a global workspace to arrive jointly at the best
phonetic description of the sound.
In the rapidly-developing field of machine speech
recognition, the Hearsay system was quite good for its time: it
was able to understand almost 1,000 words spoken by any male
speaker in a normally noisy room, using an ordinary microphone
(ref). This was quite a bit better than most comparable systems
were doing at the time.
The subsequent history of the Hearsay project is rather
interesting. In working with the various expert knowledge sources
used by Hearsay, the researchers discovered a way to improve the
acoustic processor so that it could do predictive tracking of
acoustical formants, the regions of the highest acoustical energy
in the frequency spectrum. In other words, they discovered a
successful îalgorithmï which made it possible for the acoustical
processor to solve problems which previously required cooperation
From other processors, like syntax and semantics. Once th
became clear, the Hearsay team was able to dispense with the
distributed architecture of Hearsay, since cooperative
computation was less necessary. They developed a new system
called Harpy based on the improved acoustical processor, which
could do the same job in a more specialized way (ref). But from
our point of view, Hearsay is actually more interesting as a
psychological model than the specialized Harpy. Hearsay did not
fail; rather, it succeeded as a development system, a stepping
stone to a specialized algorithm for translating sounds into
phonetic code.
There is a nice analogy between this history and the‹j ‹
development of new human skills. When people start learning some
new task, doing it takes a great deal of conscious processing.
Apparently many functionally separate processors need to
cooperate in new ways in order to perform the task. Over time,
however, simpler means are found for reaching the same goal, and
control over the task is relegated more and more to a single
specialized processor (which may take components from existing
processors). Thus the distributed "committee system" îshould beï
surpassed in the normal course of events by the development of a
new expert system. This is certainly what we would expect to
happen as a new skill becomes automatic and unconscious.
Thus the first point in Table 2.1, the computational
inefficiency of consciousness, fits the model we are considering.
Computations carried out entirely through the medium of the
global workspace demand the tacit or active cooperation of all
relevant processors. Naturally such a process takes much more
time than a comparable process that is done exclusively by an
expert system prepared to solve the problem by itself. But what
about the contrasting point about unconscious processors?
According to Table 2.1, "unconscious processors are highly
efficient in their specialized tasks". This is already assumed in
the model we are discussing here, so this point also fits the
model.
What about the second point in Table 2.1? "Conscious
processes have great range, but unconscious processors have
relatively limited range." If blackboard messages correspond to
conscious contents, then they must range as widely as do the
distributed processors which are able to place a message on the
blackboard. Thus the range of messages in the global workspace is
very great, while the range of information processed locally by
any individual processor must be more restricted.
Further, "conscious processes have great relational capacity
and context-sensitivity ..." Relational capacity is defined as
the ability to relate different conscious contents to each other.
Obviously, several blackboard messages could be related to each
other, especially if some expert were alert to such
relationships, and if several messages occurred close together in
time. (We will defer discussion of context-sensitivity, the
shaping of conscious contents by unconscious factors, until
Chapter 4.) Contrastively, on this point, "... unconscious
processors are relatively isolated and autonomous." This is
assumed, of course, in the very nature of a distributed system.
So far, the fit between the model and the data to be explained is
very close.
What about the "internal consistency" of conscious contents?
This fits well also, because blackboard messages require at least
tacit cooperation from the audience of experts. If some global
message immediately ran into powerful competition, it could not
stay on the blackboard. And what about the contrastive point that
"... unconscious processors are highly diverse"? This, too, is‹j ‹
already inherent in the idea of a distributed system of expert
systems. So far, so good.
Table 2.1 further claims that "conscious processes are
serial". This follows directly from the requirement that they be
internally consistent --- different messages, those which cannot
be unified into a single message, can only be shown one after the
other. Thus we cannot see two objects occupying the same location
in space at the same time, as we would have to, to interpret the
Necker Cube in two different ways simultaneously. The blackboard
portion of the system is therefore forced into seriality. But
"... unconscious processors can operate in parallel". This, too,
is already inherent in our model.
Finally, "conscious processes have limited capacity..." .
This feature also flows from the "internal consistency"
requirement. If any global message must be internally consistent,
one must exclude irrelevant or contradictory messages that may
come up at the same time. Such irrelevant or contradictory
messages are likely to exist somewhere in some of the distributed
processors, and are therefore a part of the system. But they
cannot gain access to the blackboard unless they can drive off
the current message, or unless it leaves the blackboard of its
own accord. Hence, "... unconscious processors, taken together,
have very great capacity", and can be doing many things locally
at the same time, provided these local processes do not require
access to the global workspace.
In conclusion, we can now replace all of the facts described
in Table 2.1 with a rather simple model: the idea of a set of
specialized processors, each well-equipped to handle its own
special job; all the specialists can communicate with the others
through a global workspace. In this way they can cooperate and
compete with each other, to strengthen or weaken a global
message.
Like consciousness itself, this system works best when
routine tasks are directly delegated to the best expert that is
ready to solve it, and the use of the blackboard is reserved for
just those problems that cannot be solved by any expert acting
alone. When the cooperating processors discover a single
algorithm able to solve the problem, that algorithm can again be
handled by a single expert, freeing up limited global capacity
for other unsolved problems.
2.31 When is a Global Workspace System useful?
The main use of a GW system is to solve problems which any
single expert cannot solve by itself --- problems whose
solutions are îunderïdetermined. Human beings encounter such‹j ‹
problems in any domain that is novel, degraded, or ambiguous.
This is obvious for novelty: if we are just learning to ride a
bicycle, or to understand a new language, we have inadequate
information by definition. Further, if the information we
normally use to solve a known problem becomes degraded,
deteriminate solutions again become indeterminate. So much is
clear. What may not be so obvious is that there are problems
that are inherently ambiguous, in which all the local pieces of
information can be interpreted in more than one way, so that we
need to unify different interpretations to arrive at a single,
coherent understanding of the information.
This kind of inherent ambiguity is ofen found in language
processing and even in visual perception (e.g. Marr, 1980; Rock,
1981). We discuss the prevalence of local ambiguity in the world
of perception, action, language, and thought in Chapter 4.
Briefly, the argument is that any restricted amount of
information tends to have more than a single interpretation.
Since we often must deal with a restricted information,
ambiguities must be resolved by reference to new and
unpredictable information. The Global Workspace architecture is
designed precisely to allow resolution of ambiguity by
unpredictable knowledge sources.
A further use of a global workspace is to îupdateï many
specialized processors at the same time. Updating is necessary
not merely to remember where one's car is parked, but also to
track changes in social relations, perceptual conditions, and the
like. There is good evidence that social perception can be
changed by a single conscious experience (Lewicki, 1986), and
similarly, phoneme perception is known to be changed by recent
experiences (Eimas & Corbit, 1972).
2.32 Cooperative computation: an illustration.
We can illustrate this process in a slightly fanciful way by
imagining the different language experts communicating to
analyze a children's language game, based on a junctional
ambiguity, "I scream, you scream, we all scream for ice cream."
Suppose that the sound is consciously perceived, and globally
displayed, as the following string:
AYSKREEMYOOSKREEMWEYALSKREEMFRAYSKREEM
This is quite realistic, because there are no physical gaps
between most words in ordinary speech. (Most of the word
boundaries we îseemï to hear are perceptual, not physical.) Now let
us imagine a conversation between several experts trying
cooperatively to understand this string.
‹j ‹åLexical processor: Looking at the first syllable, I can make out
a couple of possibilities. The phoneme A could be the exclamation
"ah!"; AY could be "I"; AYS could by "ice"; AYSK is not in my
lexicon, nor is AYSKR.
Syntactic processor: If the first syllable is the start of an
exclamation "ah!", I predict another exclamatory word, or the
beginning of a new sentence. If it is "I", I predict that a verb
phrase with an active first-person verb will come next, or much
less probably a subordinate clause. If the first syllable is
"ice" it would be an unusual way to start a sentence, though it
could be a generic noun. Highest probability: "I".
Lexicon: OK, let's assume it's "I". Next syllable, SKREE or
SKREEM. "Scree" is very rare, "scream" is possible.
Syntax: "Scream" is an active, first-person verb, as predicted.
Most likely, the next word will be the preposition "at", which
will start a noun phrase with the object of the verb.
Semantics: "I scream" is indicative of excitement, usually due to
fear.
Lexicon: Assuming the analysis is on the right track so far, the
next syllable is YOO (you) or YOOS (use), but not YOOSK or
YOOSKR.
Syntax: Surprise! Neither "you" nor "use" is the expected
preposition "at"; therefore I'm changing my prediction about a
coming noun phrase. "I scream" must be treated as a separate
clause, or we're doing this all wrong. If so, then "you" could
start the next clause, with the optional copula "and" being
understood. If so, we should get a verb phrase next, with an
active first-person verb.
Lexicon: OK, let's assume it's "you"; then the next syllable is
SKREE or SKREEM as before. "Scree" is rare, "scream" is
predicted.
Syntax: "You scream" is a well-formed formula.
Semantics: "I scream, you scream" indicates a lot of excitement,
probably due to fear.
Pragmatics: What is this speaker trying to tell me? I am not
screaming. Counterfactual assertion: Look out for possible joke
or language game.
Lexicon: Next syllable is WE, WEY, WEYA, or WEYAL --- only "we"
is in my vocabulary. WE and WEY could be alternative strings for
the same word, since WE is followed by a vowel, and Y is often
used as a transition before a vowel.
Syntax: I'm beginning to see a pattern here. We could be getting‹j ‹
a series of clauses consisting of subject-predicate pairs, strung
together without copulas. If so, I predict the next word will be
another active first-person verb.
Semantics: If so, the next verb should have something to do with
excitement, possibly fear. "I scream, you scream, we ... what?"
Lexicon: Assuming we are OK so far, next syllable is YAL or AL.
"All" is in my vocabulary.
Semantics: OK, "all" makes sense after the plural "we".
Syntax: Not quite what I expected, but it could be part of the
subject noun phrase. Next, I predict a verb phrase with an active
first-person verb.
Semantics: OK, that fits "we".
Lexicon: Looking for parallels as suggested, the next word SCREEM
is "scream" again. This is getting boring.
Syntax: Just as I expected.
Semantics: Just as I expected. "I scream, you scream, we all
scream." More excitement, possibly fear.
Pragmatics: Don't trust the predictions too much. Watch out for a
tricky language game!
Lexicon: Next syllable could be FRA, FRAY, or if the vowel is
reduced, FOR. "Fry" and "for" are in my vocabulary.
Syntax: I vote for "for", now expecting an object noun phrase.
Semantics: Finally, we find out what this is all about. "We all
scream for ... what?"
Lexicon: Same string as before, A, AY, or AYS. Predict AY on the
basis of previous pattern.
Syntax: Sorry, "I" is not an object noun.
Lexicon: Well, going back a bit, "for" could mean "because", but
that is archaic and unlikely. How about "ice"?
Syntax: OK, let's try it. If there is any more to this sentence,
I predict it will be the remainder of the object noun phrase.
Lexicon: Assuming we are OK so far, the next syllable is KREE, or
KREEM; only KREEM is in my vocabulary as "cream".
Semantics: "I scream, you scream, we all scream for ice cream."
"Ice cream" is a food-stuff much appreciated by young humans, who
stuff it in a hole in the middle of their faces. Previous use of‹j ‹
"scream" three times could be indicative of excitement about ice
cream, though the style is distinctly hyperbolic. Note repetition
of the same pattern three times.
Lexicon: No, four times, given the phonetic identity of KREEM,
and the fact that the string AYSKREEM has two distinct readings.
Pragmatics: See, told you, it's a language game. Presumably
humorous.
Facial processor: Please initiate a small smile.
Notice how the cooperating and competitive hypotheses
generated by these very different expert systems help to solve
quite a complex problem. The sameness of "I scream" and "ice
cream" never presented any real problem to this system, because
syntax predicted "I scream" for the first occurrence of AYSKREEM;
similarly, syntax predicted a noun phrase like "ice cream" for
the second occurrence. The Hearsay system used a global workspace
to communicate hypotheses back and forth, but more direct
channels might also be used. The advantage of a global workspace
is that it permits rule-systems îwhose relevance cannot be known
ahead of timeï to participate in solving the problem. The more
novelty and ambiguity there is to be resolved, the more useful it
is to have a global workspace.
2.33 Spreading activation and inhibition to carry out
cooperative and competitive processing.
There are several ways to carry out this notion of
cooperative processing in detail. One method that is currently
very promising involves the spread of activation in a network
(e.g. Rumelhart & McClelland, 1982). Processing in the GW model
can also use activation; in practice, this means assigning number
to different potentially conscious messages to show the
likelihood of their becoming conscious. In the example above, the
acoustic processor can display its hypothesis on the global
workspace. Syntax, semantics, the lexicon, etc. can then add or
subtract activation from this hypothesis. If the activation falls
below a certain level, or if some alternative gathers more
activation, the current acoustic hypothesis fades from the GW and
is replaced by a more popular one. This is processing as a
popularity voting contest.
We have previously suggested that high activation may be a
necessary but not sufficient condition for conscious experience
(1.xx). We can now add the idea that activation may be
contributed by many cooperating processors, adding to the vote
for the current content, to keep it on the blackboard.
‹j ‹å
2.4 Input properties of the global workspace.
We can suggest a few input properties of the conscious
global workspace. First, we emphasize, as in Chapter 1, that
relatively long-lasting conscious contents seem heavily biased
toward perception and imagery (and imagery resembles perception).
Further, it seems that the minimal conscious integration time is
approximately 100 milliseconds. Here are the details.
2.41 Conscious experience has a strong perceptual bias.
Consciousness is not identical to perception, but perception
is certainly the premier domain of detailed conscious experience.
In later chapters we will argue that conscious access to abstract
concepts, and even conscious control of action, may be mediated
through rapid, quasi-perceptual events (7.0). If that is true,
then the "language" of GW input systems may be perceptual (1.xx).
2.42 The conscious moment: Temporal limits on conscious
stimulus integration.
Several temporal parameters are associated with conscious
experience. Short Term Memory seems to involve maximum retrieval
times on the order of ten seconds without rehearsal (Simon,
1969), and there is evidence for a .5 second lag time between
sensory input and conscious appreciation of a stimulus (Libet,
1978, 1981; Figure 2.42). Here we are mainly concerned with the
"cycle time" of the conscious component, presumably the Global
Workspace. Most of the evidence for such a cycle time comes from
studies of perceptual integration. Blumenthal (1977) presents a
remarkable synthesis of the vast perceptual literature on "the
psychological moment." He argues that "Rapid attentional
integrations form immediate experience; the integration
intervals vary from approximately 50 to 250 milliseconds, with
the most common observation being about 100 milliseconds."
Blumenthal's excellent summary of the evidence for this
integration interval is worth quoting in full:
î1. Time-Intensity Relations.ï Within the integration interval
there is a reciprocity of time and experience. A mental
impression is integrated and formed over this duration. Several‹j ‹
faint stimuli occurring within the interval may be summed to the
mental impression of one intense stimulus. If events should
somehow be cut off midway through the integration, our impression
of them will be only partially or incompletely formed.
î2. Central Masking.ï When two or more events that cannot be
easily integrated occur within an integration interval, the
process may develop or form impressions for some events and
reject others.
î3. Apparent Movement.ï Two spatially as well as temporally
separate stimuli that fall within an integration interval may
again be fused, or integrated, to a unitary impression. Because
of their spatial separation, however, they will be experienced as
one object in motion between the two locations.
î4. Temporal Numerosity and Simultaneity.ï In any sequence of
rapidly intermittent events, intermittency can be experienced at
rates no faster than approximately 10 events per second. This is
a limit on the rate of human cognitive performances in general.
î5. Refractory Period and Prior Entry.ï Sometimes when two
events occur in the same integration interval and are neither
fused nor masked, one event will be delayed and integrated by a
succeeding pulse of attention. It will thus appear to be
displaced in time away from its true time of occurrence. If two
responses must be made, one to each of two rapidly successive
events, the second response is delayed for the duration of a
rapid integration interval.
î6. Memory Scanning.ï Impressions that are held in short-term
memory can be scanned no faster than the rate determined by the
attentional integration process. In searches of logically
structured information held in long-term memory, the scan through
chains of associations proceeds at the rate of the attentional
integration process --- about 75 - 100 msec for each node in the
chain.
î7. Stroboscopic Enhancement.ï In an otherwise unstructured
stimulus environment, an intermittent stimulus (such as a
flashing light) that pulses at a rate of about 10 per second can
drive the rapid attentional integration process to exaggerated
levels of constructive activity so as to produce hallucinatory
phenomena."
The most straightforward interpretation of these findings is
that perceptual specialists can cooperate and compete within the
rough 100 millisecond period, but that longer intervals between
them do not allow them to interact to create a single, integrated
experience.
There is a problem with the 100 millisecond period: some
competition for access to limited capacity mechanisms takes much
longer than that. For example, we may have conscious indecision,‹j ‹
considering this side and that of a difficult question. Such
indecision may take seconds, minutes, or hours; it does seem to
involve a kind of slow competition for access to conscious
experience (7.0). Usually after a decision is made, only one
perspective will continue to have access to consciousness. One
hundred milliseconds is an absurdly short time to allow two
different thoughts to compete for access to consciousness, or to
decide between two different courses of action. We will suggest
in Chapters 4 and 5 that these kinds of competition for limited
capacity involve competition between the îcontextsï of immediate
experience, rather than between instantaneous qualitative
conscious events.
It is good to put this in a larger perspective. Figure 2.42
presents a number of time parameters associated with conscious
experience.
v **vÉ
v * v------------------------------É
v * vùúInsert Figure 2.42 about here. É
v * v------------------------------É
2.43 The Threshold Paradox.
The reader may already have noticed a problem with our
approach so far. Namely, in order to recruit a coalition of
specialized processors to work on some global message, we must
broadcast it. But it needs the help of other systems to become a
global message in the first place. How then can a message gain
access to the global workspace if it does not already have a set
of other systems supporting it? How does it cross the threshold
of conscious access? This is not unlike the problem faced by a
budding young artist and writer: in order to get public interest,
one needs to show that the work has appeal. But one cannot show
that unless the work has been shown to the public.
This Threshold Paradox may be quite serious for our model,
and may demand some change (x.xx). In general, there are two
solutions. First, it may be that there is a hierarchy of
workspaces of increasing global reach. At lower levels there may
be broadcasting across some systems, but not all; at higher
levels, there may be truly global broadcasting. This would allow
new messages to recruit an increasing number of supportive
systems, until ultimately it is broadcast globally. We can call
this the Waiting Room option, as if there is a series of waiting
rooms, each of which is closer to the global broadcasting
facility. In the same way the budding artist can show his or her
work to increasingly large and influential groups of people, who
then may make it possible to come closer and closer to true
public acceptance.‹j ‹å
There is another option. It is conceivable that all systems
clamoring for conscious access may receive momentary global
access, but for too short a time to be reported consciously. Each
instant of access may allow the recruitment of additional
supportive systems. The more systems are recruited, the more
likely it is that the message will remain on the global workspace
long enough to be perceived and recalled as a conscious event. In
this way a new message may gain more and more support, and
increasingly likelihood of being held long enough to be
reportable. This may be called the Momentary Access option.
The Threshold Paradox leads to a theoretical choice©point
that we cannot entirely resolve at the present time. Indeed, both
options may be true. They are both consistent with the finding
that conscious access may take as long as .5 seconds. We will
suggest in Chapter 3 that the neurophysiological evidence
supports a "snowballing" development of access to consciousness,
rather than an instantaneous process. The temporal factor may
seem to support the Momentary Access option, but in fact the
Waiting Room option presumably takes time as well. We do not ave
the evidence at this point to choose between them.
2.5 Output properties of the global workspace:
How global is global?
Once a specialized systems gains access to the global
workspace, what does it gain? In everyday language the word
"global" means "world-wide" or "relating to the whole of
something," but in computational parlance its meaning is more
specific. A îglobal variableï is one that is defined for all
subsystems of a larger system, as opposed to îlocalï variables,
which are only defined for a single subsystem. If the entire
system has three parts, a global variable can be recognized in
all three. Thus access to a global workspace implies access to
the larger system as a whole.
Things get a lot more interesting if we consider that the
nervous system has a very large number of processors, many of
which can themselves be decomposed into sub-processors. A truly
global variable in the nervous system might be distributed in
principle to all levels of all processors, perhaps down to the
level of single cells. Thus one question we can ask is, "How
widely is global information distributed"? Is it made available
to just a few specialists? Or is there evidence that in the
nervous system, "global information" is really broadcast
throughout?‹j ‹å
Several source of evidence suggest that conscious events
have very wide distribution in the nervous system. Consider:
1. îAnyï conscious or effortful task competes with îany
other.ï
We can call this the "any" argument. Perceiving a single
star on a dark night interferes with the voluntary control of a
single motor unit in one's thumb; the consciousness of the
letters in this sentence will interfere with conscious access to
the meaning of this chapter. Indeed, being conscious of îanyï
stimulus in îanyï sense modality interferes with consciousness of
îanyï other stimulus, and also with conscious access to îanyï
voluntary act or conceptual process. When these same events are
unconscious and involuntary they usually do not interfere at all.
If we believe that the nervous system consists of many
specialized systems that decide by their own criteria what
information to process, it follows that even relatively small
systems --- like those needed to see a single star on a dark
night, or to control a single muscle fiber --- must somehow have
access to the conscious system.
One can extend the "any" argument to other cases.
Psychophysicists have used the technique of "cross-modality
matching" for decades now, showing that any stimulus in any
modality can be compared in intensity with any other stimulus in
any other modality --- and the result is not chaos, but some of
the most elegant and mathematically regular data found in
psychological research (Stevens, 1966). Research on classical
conditioning indicates that within wide biological limits
(Garcia & Koelling, 1966) very many different stimuli can come to
signal the appearance of a great variety of other stimuli, even
if there is no natural connection between them. The strength of
classical conditioning is greatly increased when there is a
natural, biological connection between the conditioned and
unconditioned stimulus. But the fact that classical conditioning
can occur at all between a tone and shock, which are not
biologically related, suggests a capacity for some arbitrary
connections. Below we will explore in more detail the related
finding that one can apparently gain novel voluntary control, at
least temporarily, of "any" neural system with the help of
biofeedback training (Chase, 1974). In humans, "any" stimulus
can serve as a signal to perform "any" voluntary act. In language
comprehension, when one encounters ambiguities --- which are rife
at every level of language --- influences from "almost any"
contextual factor can serve to resolve "any" ambiguity (4.0). And
so on. The "any" argument applies in a number of domains. It
always implies the existence of some integrative capability, one
that allows very different specialized systems to interact.
If the brain equivalent of a "global workspace" is truly‹j ‹
global, then it should be true that any brain event that is
conscious or under conscious control can interact with any other
event, no matter how different. It seems difficult to explain
this without something like a truly global workspace. Consider
now the case of biofeedback training.
2. Conscious feedback can be used to gain a degree of
voluntary control over essentially any neural event.
It is not emphasized often enough that biofeedback training
îalwaysï involves conscious information. To gain control over
alpha-waves in the EEG, we sound a tone or turn on a light
whenever the alpha-waves appear; to gain control over muscular
activity we may play back to the subject the sound of the muscle
units firing over a loudspeaker; and so on (Chase, 1974;
Buchwald, 1974). This is not to say, of course, that we are
conscious of the details of action --- rather, we must be
conscious of some feedback from the action to establish voluntary
control. In terms of the global-workspace theory, establishing
biofeedback control requires that we "broadcast" the conscious
feedback in some way.
With conscious feedback people can gain at least temporary
control over an extremely wide range of physiological activities
with surprising speed. In animals, biofeedback control has been
established for single neurons in the hippocampus, thalamus,
hypothalamus, ventral reticular formation, and preoptic nuclei
(Olds and Hirano, 1969). In humans large populations of neurons
can also be controlled, including alpha-waves in the EEG,
activity in the sensory-motor cortex, evoked potentials, and the
lateral geniculate nucleus (Chase, 1974). In the human voluntary
muscle system, single motor units --- which involve only two
neurons --- can come under conscious control with half an hour or
so of training, and with further biofeedback training subjects
can learn to play drumrolls on single spinal motor units!
(Basmajian, 19xx). Autonomic functions like blood pressure, heart
rate, skin conductivity, and peristalsis can come under temporary
control; more permanent changes in autonomic functions are
unlikely, because these functions are typically controlled by
interlocking negative feedback loops, producing a system that
tends to resist change. But in the Central Nervous System (CNS),
as Buchwald (1974) has written, "There is no question that
operant conditioning of CNS activity occurs --- in fact, it is so
ubiquitous a phenomenon that there seems to be no form of CNS
activity (single-unit, evoked potential, or EEG) or part of the
brain that is immune to it." (Footnote 2.)
This is what we would expect if conscious feedback were made
available throughout the brain, and local distributed processes
"decided" whether or not to respond to it. We may draw an analogy
between biofeedback training and trying to locate a child lost in
a very large city. It makes sense initially to search for the‹j ‹
lost child around home or school, in a îlocalï and îsystematicï
fashion. But if the child cannot be found, it may help to
broadcast a message to all the inhabitants of the city, to which
only those who recognize it as personally relevant would respond.
The message is global, but only the appropriate experts respond
to it. Indeed, it is difficult to imagine an account of the power
and generality of biofeedback training without some notion of
global broadcasting.
3. Event-Related Potential (ERP) studies show that conscious
perceptual input is distributed everywhere in the brain, until
stimulus habituation takes place. After stimulus habituation
(i.e. after consciousness of the stimulus is lost) there is only
local activity.
There is direct neurophysiological evidence for global
broadcasting associated with consciousness. E.R. John has
published a series of experiments using Event-Related Potentials
(ERPs) to trace the neural activity evoked by a repeated visual
or auditory train of stimulation --- that is, a series of bright
flashes or loud clicks (Thatcher & John, 1977). Thus a cat may
have a number of recording electrodes implanted throughout its
brain, and a series of light- flashes are presented to the cat
(which is awake during the experiment). Electrical activity is
monitored by the implanted electrodes, and averaged in a way that
is time-locked to the stimuli, to remove essentially random
activity. In this way, remarkably simple and "clean" averaged
electrical traces are found amid the noise and complexity of
ordinary EEG.
John's major finding of interest to us was that electrical
activity due to the visual flashes can initially be found
everywhere in the brain, far beyond the specialized visual
pathways. At this point we can assume that the cat is conscious
of the light flashes, since the stimulus is new. But as the same
stimuli are repeated, habituation takes place. The electrical
activity never disappears completely, as long as the stimuli are
presented, but it becomes more and more localized --- until
finally it is limited only to the classical visual pathways.
These results are strikingly in accord with our expectations.
According to Model 1, prior to habituation, the information is
conscious and globally distributed. But after habituation, it
ceases to be conscious and becomes limited only to those parts of
the brain that are limited to visual functions. Only the
specialized input processor is now involved in analyzing the
stimulus. (Other refs, DAVID G.?)
4. The Orienting Response, closely associated with conscious
surprise at novelty, is known to involve every major division of
the nervous system.
First, we know that any novel stimulus is likely to be‹j ‹
conscious, and that it will elicit an Orienting Response (OR)
(e.g. Sokolov, 1963). The OR is probably the most widespread
"reflexive" response of all. It interrupts alpha-activity in the
EEG, it dilates or contracts blood vessels all over the head and
body, it changes the conductivity of the skin, causes orienting
of eyes, ears, and nose to the source of stimulation, triggers
changes in autonomic functions such as heart-rate and
peristalsis, it causes very rapid pupillary dilation, and so on.
In recent years it has been found to have major impact on the
cortical evoked potential (P300 wave, ref). All these changes
need not be produced by a globally broadcast message, but the
fact that they are so widespread, both anatomically and
functionally, suggests that something of that sort may be going
on.
5. The reticular-thalamic system of the brain stem and
midbrain is closely associated with conscious functions. It is
known to receive information from all input and output systems,
connects to virtually all subcortical structures, and it projects
diffusely from the thalamus to all parts of the cortex. Thus it
can broadcast information to the cortex.
This system is explored in detail in the next chapter.
6. All aspects of a conscious event seem to be monitored by
unconscious rule-systems, as suggested by the fact that errors at
îanyï level of analysis can be caught if we become conscious of the
erroneous event.
This may seem obvious until we try to explain it. Take a
single sentence, for example, spoken by a normal speaker. We very
quickly detect errors or anomalies in pronunciation, voice-
quality, location, room acoustics, vocabulary, syllable stress,
intonation, emotional quality, phonology, morphology, syntax,
semantics, stylistics, discourse relations, conversational norms,
communicative effectiveness, or pragmatic intentions of the
speaker. Each of these aspects of speech corresponds to a very
complex, highly developed rule-system, which we as skilled
speakers of the language have learned to a high level of
proficiency (Clark & Clark, 1977). The complexity of this
capacity is simply enormous. Yet as long as we are conscious of
the spoken sentence we bring all these different knowledge
sources to bear on it, so that we can automatically detect
violations in îanyï of these rule-system. This implies that the
sentence is somehow available to all of them. But if we are not
conscious of the sentence, we do not even detect our own errors
(MacKay, 1980). Again, there is a natural role for "global
broadcasting" in this kind of situation.
In sum, how global is global? The previous six arguments
suggest that conscious events can be very global indeed.
‹j ‹å
2.6 Further considerations.
Below we explore further ramifications of Model 1. There are
several models that behave much like the Global Workspace system,
and these must be considered. We derive some testable predictions
From Model 1, suggest some further questions it may answer, a
some that it fails to handle. Some of these puzzling questions
may be answered by more advanced models developed in later
chapters.
2.61 Functional equivalents of a GW system.
While we will continue to speak of a "global workspace" in
this book, there are other systems that behave in much the same
way: They are îfunctionally equivalentï. Take our previous analogy
of an assembly of experts, each one able to publicize his or her
ideas by writing a message on a blackboard for all to see. This
is much like a global workspace in a distributed system of
specialized processors. But take away the blackboard, and instead
give each expert a megaphone, loud enough to be heard by all the
others. The megaphones are wired together in such a way that
turning on any megaphone turns off all the others. Thus only one
expert can broadcast a message at any moment. îFunctionallyï this
is equivalent to the blackboard system (see Figure 2.62).
We suggest in this book that consciousness is associated
with a global workspace îor its functional equivalentï. How this
system is realized in detail remains to be seen. One way to
emphasize this is to say, following the title of this chapter,
that consciousness is characterized by at least two primary
properties --- conscious contents are îcoherentï and îglobally
distributedï. If we state things at this level of abstraction, we
can avoid becoming committed to any particular "hardware"
instantiation.
‹^ ‹
2.62 The Mind's Eye, Mind's Voice, and Mind's Body as
aspects of a global workspace.
------------------------------------------------
FIGURE 2.62: THE MIND'S SENSES AS A GW EQUIVALENT
------------------------------------------------
Figure 2.62 shows one kind of functional equivalent to the
Global Workspace system, in which the Mind's Senses are global
workspaces, wired so that only one can work at a time. As we have
noted, the Mind's Senses can be treated as workspaces of a kind
(x.x). Inner speech has of course long been associated with Short
Term Memory or "working memory" (Baddeley, 1976).
Note however that we have some voluntary control over visual
imagery, and especially over inner speech. Voluntary control is
something which a theory of conscious experience should tell us
about (Chapter 7.0). Certainly we cannot take voluntary control
for granted, or presuppose it in a theory of mental imagery.
Further, current models of mental imagery have little to say
about consciousness as such (but see Pani, 1982). They typically
do not account for habituation and automatization. Nevertheless,
Figure 2.62 suggests one attractive instantiation of the Global
Workspace notion.
2.63 What is the global code?
We have previously raised the question of a mental îlingua
francaï, a common language of mental functioning that may operate
across many systems. Such a common code seems plausible to make
a GW system work, though it may not be absolutely necessary,
since one could broadcast local codes through a global workspace.
One possibility is that input into the GW may be perceptual
or quasi-perceptual, as we suppose in Model 1, and that
processors in the "audience" respond only to the most general
aspects of these global message, namely their spatio-temporal
properties. Thus a motor control program may be able to recognize
at least the spatio-temporal aspects of a rich perceptual scene,
enough to know that "something of interest is happening at 12
o'clock high at this very moment". The motor program could then
cause other receptors to orient to the stimulus, thereby helping
to make better information available to all relevant parts of the
system. The idea that the îlingua francaï may be a spatio-temporal
code is consistent with the fact that many brain structures are
sensitive to spatio-temporal information. Further, we know that
biofeedback training, which can be done with any specialized
system, always involves temporal near-simultaneity between the
biofeedback event and the conscious feedback signal.
This is consistent with the notion of global temporal coding.
‹j ‹å 2.64 Other related models.
Ideas related to the GW system have been discussed for some
time. In Chapter 1 we pointed out that Aristotle's "common sense"
has much in common with the global workspace. More recently,
Lindsay and Norman (1976) have pointed to the global workspace
architecture as a psychological model, as have others. Recent
work on formal models of distributed systems also has explored a
global workspace architecture (McClelland, Rumelhart, & the PDP
Group, 1986, Chapter 10). Others refer to the "spotlight of
consciousness" (e.g. Crick, 19xx).
A GW is a natural domain for interaction between otherwise
separate capacities. The relationship between conscious
experience and integration between separable aspects of
experience has been noted by Treisman & Gelade (1982) (above), as
well as by Mandler (1983), Marcel (1983 ab), La Berge (1981), and
others.
There are fewer sources for the somewhat surprising notion
that conscious experiences may be broadcast everywhere in the
nervous system. E.R. John's "statistical model of learning"
seems to be closely related to this (1976), and Gazzaniga (19xx)
suggests a specific connection between conscious experience an a
publicity device. Neurophysiologists have long known about
diffuse and non-specific anatomical areas, and some
neuroanatomists have explicitly related the brain stem reticular
formation to Aristotle's common sense (see next chapter).
Gazzaniga has recently proposed that consciousness serves as a
publicity organ in the brain. Curiously enough, he also suggests
that its primary function is post-hoc rationalization of past
events. This seems an unduly limited view of the functions of
consciousness (see Ch. 10).
The closest models are those which relate conscious
experience to limited capacity mechanisms in parallel distributed
systems. Thus Reason (1984) and Norman & Shallice (1980) have
proposed systems along these lines. Similarly, John Anderson
(1983) suggests that the "working memory" in his ACT* system,
currently the most detailed architectural model, is closely
related to conscious experience.
2.7 Testable predictions and counterarguments.
2.71 Testable predictions from Model 1.
Many of the empirical studies we have cited can be further‹j ‹
developed to test aspects of Model 1. We will focus here on a few
possibilities. Note that many of these are phrased in terms of
measures of limited capacity, rather than of conscious experience
as such. This is because Model 1 states some but not all of the
necessary conditions for conscious experience; but it is a model
of limited capacity. Thus we will phrase predictions in terms of
limited capacity îorï conscious experience at this point.
îThe GW as a domain for novel interactions.ï
One core concept is that novel interactions between routine
processors require the global workspace. Sophisticated processing
may go on much of the time without conscious or limited-capacity
processes, but înovelï processes are presumably unable to proceed
in this way. One experimental prediction, then, is that novel
word-combinations cannot be processed unconsciously, while
routine ones can. We will discuss the issue of novelty and
informativeness in Chapter 5, and suggest some experimental
predictions there.
îTesting global interaction and broadcasting.ï
Biofeedback training may provide an excellent experimental
paradigm for investigating the claims about global interaction
and broadcasting. It is quite a strong claim that any part of the
nervous system can in principle interact with any other, given
the global workspace as a mediator. But this result is predicted
by Model 1. Because any global message should interfere with any
other, and because biofeedback allows us to control a repetitive
neural event, perturbations in the control of a biofeedback-
controlled tracking task may be used to monitor global-workspace
activity. It is well-established that people can learn to control
a single muscle fiber, controlled by two spinal neurons,
separately from all the others with a brief period of biofeedback
training. Typically, the feedback is provided by a "click,"
consisting of the amplified electrical signal from the muscle
unit, played over a loudspeaker. Indeed, Basmajian () has shown
that one can learn to play drum-rolls on a single motor unit! If
subjects can be trained to emit a steady stream of motor pulses,
at a rate of perhaps 5 Hz, one could investigate the interaction
of this marker stream with other conscious or voluntary events,
such as the detection of a perceptual signal, the comprehension
of a sentence, and the like. Any conscious or limited-capacity
event should interfere with the control over the motor unit.
îThe minimum integration time of conscious experience.ï‹j ‹å
Evidence for perceptual fusion has been cited above to
support the possibility of a rather brief 100 millisecond cycle
time for the global workspace. One might use biofeedback training
to investigate this temporal interaction window. For example, in
motor unit training, a discrete and covert neural event (the
motor spike) is amplified and fed back through loudspeakers to
create a discrete conscious event (the auditory click). It would
be easy to delay the click for 50, 100, 200, and 300
milliseconds, to measure the allowable lag time between the two.
An approximate 100 millisecond upper limit would be very
interesting. If the 100 millisecond cycle time is approximately
correct, biofeedback training should not be possible past this
conscious exposure time.
The same suggestion may be made on the input side. The work
of Treisman and her colleagues (Treisman & Gelade, 1980; Treisman
& Schmidt, 1982) and Julesz (Sagi & Julesz, 1985) suggests that
one can easily specify separable features in visual perception.
What would happen if one were to delay one feature 50 milliseonds
before others became available? If the notion of a minimal cycle
time were valid, there should be integration with short temporal
disparities, with a rapid loss of integration beyond some "magic
number" around 100 milliseconds.
îDoes composition, decomposition, and reorganization of
processors take up limited capacity?ï
We claimed in Chapter 1 that many slips of speech and action
show a momentary separation of otherwise integrated action
schamata. Does such separation take up limited capacity? And if
one re-integrates the fragmented schema, does that require
limited capacity? Model 1 would certainly suggest that
reintegration between otherwise separate systems makes use of the
global workspace. In recent years it has become possible to
trigger a variety of predictable slips of speech and action in
the laboratory (e.g. Baars, 1980). We have not at this point
investigated the question of limited-capacity loading in errors,
but it may provide a fertile domain for testing GW theory.
There are numerous other cases of re-organization of
coalitions of processors. Speaking and listening probably involve
different configurations of overlapping sets of processors ---
both involve lexical access, syntax, and semantics. Presumably,
switching from speaking to listening should require at least a
momentary load on limited capacity.
îParallel error detection by many unconscious systems.ï‹j ‹å
A related question is whether error detection in performance
involves parallel monitoring systems (7.xx). Notice that
monitoring systems are not conscious ordinarily, and that, like
other specialized processors, they should be able to operate in
parallel unless they are contingent on each other. If monitoring
systems operate in parallel, "looking at" a globally displayed
event, then the time needed to detect two simultaneous errors in
some conscious material should take no longer than detecting only
one. The work of Langer and Imber (1979) indicates that error
detection becomes quite poor when some skill becomes automatic:
the less conscious it is, the more difficult it is to monitor. Of
course, automatic skills can "de-automatize" when they encounter
difficulties (x.x). That is, aspects of these skills can become
conscious once again. Any experiment on error-detection in
automatic skills must deal with this complication. However, de-
automatization presumably takes more time, and the skill should
degrade when it becomes more conscious. Thus one could monitor
whether some automatic skill continues to be automatic and
unconscious during the experiment (e.g. Marslen-Wilson & Welsh,
19xx).
Along similar lines, the creation of new specialized
modules, perhaps from previously available separate automatic
systems, should take up limited capacity and may have testably
conscious aspects (e.g. Case, 19xx).
2.72 More questions for which Model 1 suggests answers.
îA functional explanation for limited capacity.ï
Limited capacity is a prominent and surprising feature of
human psychology, but we seldom ask îwhyï it exists. Would it not
be wonderful to be able to do half a dozen things at the same
time? Why has evolution not resulted in nervous systems that can
do this? Model 1 suggests an answer. If it is important for
information to be available at one time îto the system as a wholeï,
global information must necessarily be limited to a single
message at a time. There is only one "system as a whole", and if
all of its components must be able to receive the same message,
then only one message at a time can be broadcast. There are many
reasons for making information available to the system as a
whole, notably the case where a problem cannot be solved by any
single known specialist. The knowledge required to solve the
problem may reside somewhere in the system, but in order to reach
it, the problem must be made available very widely. Notice that
this suggests a purely îfunctionalï explanation of limited
capacity.
‹j ‹å Of course, global broadcasting is expensive. If some problem
can be assigned to a specialized processor, it is efficient to do
this and not take up the limited resources of the global
workspace.
îOrganization vs. flexibility.ï
Other facts about human psychology also fall into place with
Model 1. For example, cognitive psychologists over the past few
decades have become convinced of the importance of organization
in perception and memory. There are numerous powerful
demonstrations of the effects of organization (e.g Bransford,
1979; Rock, 1981; Mandler, 1962, 1967). It is easier to remember
something if we learn a set of regularities that apply to it;
indeed, we cannot remember or even perceive utterly disorganized
information. Even "random" noise has specifiable statistical
properties.
The trouble with this is that organization tends to commit
us to a particular way of doing and viewing things. Organization
often creates rigidity. Most of the time it is appropriate for
adults to analyze language in terms of meaning, but there are
times (in proofreading for example) when we must switch from a
meaning analysis to a spelling analysis; this switch often leads
to problems. The famous "proofreader illusion" shows that we
often miss errors of spelling and vocabulary when we focus on
meaning. What kind of a system architecture is needed to
reconcile the value of organization with the need for
flexibility? In terms of Model 1, it should be a system in which
specialized processors can be decomposed and re-arranged when the
demands of the task change. This is very difficult to do with
other conceptions of organization in memory.
Some of the best demonstrations of flexibility in the
nervous system come from the area of conditioning. Originally
conditioning theorists believed that any arbitrary relationship
between stimuli and responses could be connected, and they proved
that under surprisingly many circumstances a tone can indeed come
to signal a shock, and the like. This is the very opposite of the
powerful organizational effects found by cognitive psychologists:
there is no natural connection between tones and shocks, or many
of the other standard stimuli used routinely in conditioning
studies. Indeed, when conditioning occurs between ecologically
related stimuli and responses, the effects found are far stronger
than when biologically arbitrary stimuli are used (Garcia &
Koeller, 1967). Nevertheless, it is striking that biologically
arbitrary connections can be made at all in a system that is so
strongly affected by non-arbitrary, organized, and biologically
significant relationships.
Biofeedback training provides an excellent example. When it
was first discovered, physiologists and psychologists were‹j ‹
surprised that autonomic functions such as heart rate, skin
conductivity, blood vessel dilation and contraction and the like
were affected by biofeedback (at least in humans). As the word
"autonomic" suggests, these activities were thought to be free
From conscious control. It now appears that just about any neur
system can be responsive to conscious biofeedback control,
although autonomic functions seem to resist permanent retraining
().
To account for this high degree of flexibility we favor
something like Model 1, in which routine organization of
information and control can be accessed quickly, but which also
allows for movement between different levels of organization, for
reorganization of modules in different ways, and for the
creation of entirely new, organized coalitions of processors.
îHow can people talk about their conscious experiences?ï
Finally, how is it that people can talk about the things
they experience consciously? And how can they act upon conscious
information? This is after all, our first operational definition
of conscious experience, and at some point our model should be
able to connect to it (see Chapters 7 and 9). We can already
suggest part of the answer. Speech requires a coalition of
specialized processors. Since all such processors can receive
information from the global workspace, we can explain in general
terms how it is that speech processors can describe and act upon
conscious contents. Speech systems in the global "audience" can
presumably receive the relevant information; but this does not
explain how these linguistic system organize a coherent speech
act to describe the global information. Nevertheless it is a step
in the right direction (x.xx).
Presumably the same point applies to other voluntarily
controlled systems. Instead of asking people to say "There is a
banana" when we present a banana, we can ask people to raise
their fingers or blink their eyes for bananas, and not for
anything else. All of these voluntarily controlled systems must
presumably have access to global information provided by the
conscious stimulus.
îSurprise as a momentary erasure of the global workspace.ï
We know that surprise triggers all the measures of the
Orienting Response, that it loads limited capacity, creates
massive neural activity, and tends to cause a loss of current
conscious contents. One obvious explanation is that surprise
serves to erase the Global Workspace, thereby allowing the new‹j ‹
and surprising information to be distributed for widespread
cooperative analysis (Grossberg, 1982). This is indeed part of
the story that seems to follow from the theory developed so far,
though we will have more to say about this in later chapters
(Chapters 4 and 5).
îConsciousness and executive control.ï
We are not claiming, of course, that consciousness is an
executive; in the society metaphor, it resembles a broadcasting
station rather than a government. However, governments can use
broadcasting facilities to exercise control, and presumably
executive processors may use consciousness to try to control
other processors. In this connection Shallice (1978) suggests
that consciousness has to do with the selection of a "Dominant
Action System", an idea that has obvious similarities with our
Models 2 and 3 (see Chapters 4 and 5). However, action is not the
only thing that is selected in consciousness --- conscious
experience is as selective for perception as it is for action
--- and Shallice still leaves unexplained why a Dominant Action
System would bother to dominate conscious capacity. What is the
pay-off for actions and goals to become conscious (x.x)?
Nevertheless, the general concept of a Dominant Action System is
extremely valuable, and we will propose a generalization from
Shallice's idea in Chapters 4 and 5, called a Dominant Goal
Context. This is where we will begin to introduce executive
control systems as goal contexts that shape and control access to
conscious contents, though they are not themselves conscious.
îConsciousness and repression.ï
Some readers will no doubt wonder how we can possibly
discuss our topic in any depth without dealing with the Freudian
unconscious, surely the most influential idea of this kind in
this century. The general answer is that Freud's work îpresupposedï
a cognitive theory of conscious and unconscious processes, one
which we need to work out explicitly (Erdelyi, 1985). Like most
19th century thinkers, Freud tended to take the existence of
conscious experience for granted. He treated it as equivalent to
perception, and did not discuss it in much detail. The great
surprise at the end of the 19th century was the extraordinary
power of îunïconscious processes, as shown, for example, in
post-hypnotic suggestion and the relief of hysterical symptoms
after emotional expression of traumatic memories (Ellenberger,
1970; Baars, 1986a). Freud has nothing to say about
unconsciousness that is due to habituation, distraction, or
hypnotic dissociation --- those phenomena are all quite obvious
to him, and require no explanation. He is really concerned with‹j ‹
the îdynamicï unconscious, the domain in which wishes and fears are
purposefully kept unconscious, because their becoming conscious
would lead to intolerable anxiety. The dynamic unconscious is a
conspiratorial unconscious, one that aims to keep things from us.
It is closely associated with primary process thinking, the
magical thinking displayed by young children, in dreams, and in
some mental disturbances. But these phenomena presuppose a more
general understanding of consciousness and its functions.
Our aim in this book, therefore, is to try to build a solid
cognitive foundation from which such phenomena can be understood.
We will make some suggestions later for specific ways in which
psychodynamic phenomena can be explored empirically, and how they
may be modeled in a general cognitive framework (7.xx; 8.xx).
However, there is an interesting relationship between our
basic metaphor and the repression concept of psychodynamic
theory. The global workspace is a publicity device in the society
of processors --- after all, global messages become available to
potentially îanyï processor, just as published information becomes
available to potentially any reader. Freud originally used the
opposite metaphor to explain repression, i.e., motivated
îunïconsciousness: the idea of newspaper censorship. As he wrote in
îThe Interpretation of Dreamsï (1900, p. 223):
"The political writer who has unpleasant truths to tell to
those in power ... stands in fear of the censorship; he therefore
moderates and disguises the expression of his opinion. ... The
stricter the domination of the censorship, the more thorough
becomes the disguise, and, often enough, the more ingenious the
means employed to put the reader on the track of the actual
meaning."
For Freud, the dynamic unconscious exists because of
censorship. Would it follow then that making things conscious is
the opposite of censorship, namely publicity? Repression is
presumed to be a censoring of anxiety-provoking information, but
Freud apparently did not pursue the question, what is the
censored information hidden îfromï? We might speculate that it is
sometimes desirable to conceal information from global publicity,
because some processors in the system might react to it in an
unpredictable way, challenging established control mechanisms.
For someone on a diet, it may be useful to exclude from
consciousness seductive advertisements for delicious food;
conscious contemplation of the food may lead to a loss of
control. In the same sense a politician might wish to hide a
scandal from publicity, because some political forces might react
to this information in an uncontrollable way. In both cases,
limiting publicity is a useful device for maintaining control.
There are many ways for information to become unconscious.
These mechanisms are not inherently purposeful. Habituation,
forgetting, and distraction are not conspiratorial devices to
hide a stimulus from conscious experience. However, mechanisms‹j ‹
like distraction may be îused byï some specialized systems in a
purposeful way to help control the system as a whole (Chapters 7,
8 and 9).
Experimental psychologists have had great empirical
difficulties in assessing the existence repression (Erdelyi,
1974, 1985; Baars, 1986). The points we are making here do not
solve these empirical problems. But it is pleasing to find that
this very influential conception of the psychological unconscious
may fit our analysis quite readily.
2.73 Some counterarguments.
Model 1 is clearly incomplete. Worse than that, it seems to
contradict some empirical findings, and certain powerful
intuitions about conscious experience. It clearly needs more
development. Consider the following four counterarguments.
î1. The model does not distinguish between conscious
experience and other events that load limited capacity.ï
So far, Model 1 suggests a way of thinking of the limited
capacity part of the nervous system, the part that presumably
underlies conscious experience. But in fact there are events that
load limited capacity which are înotï consciously experienced (see
Chapters 4 - 7). One counterargument to Model 1 is simply that it
does not distinguish between conscious experience and other
limited capacity-loading events. Later models will correct this
deficiency.
îï î2. The idea that we are conscious of only a single
internally consistent event at any time seems counterintuitiveï to
some people.
In reading this sentence, the reader is presumably conscious
of the printed words as well as inner speech. Most experiences,
at least in retrospect, seem to combine many separable internal
and external events. But of course at any single instance, or in
any single 100-millisecond cycle of the global workspace, we may
only have one internally consistent object of consciousness;
multiple events may involve rapid switching between different
conscious contents, just as a visual scene is known to be
integrated over many rapid fixations of the eyes. We can call
this the îbandwidth questionï: In any single integration period of
the global workspace, can more than one internally consistent
message gain access? Again, this is a difficult question to‹j ‹
decide with certainty at this point, so we will call this another
îtheoretical choice-pointï: we will assume for the sake of
simplicity that only one global message exists in any
psychological instant, and that the sense we have of multiple
events is a îretrospectiveï view of our conscious contents. Normal
conscious experience may be much like watching the countryside
flash by while sitting in a train; when we reflect
metacognitively on our own experience, we can see parts of the
train that have just gone by, as if it has gone around a curve so
that we can view it from the outside. Presumably in retrospect we
can see much more than we experience at any instant.
î3. The 100 millisecond global integration time is much too
short for many integrative processes involving consciousness.
ï
A single coherent conscious content is presumably at least
100 milliseconds long. Though it may last as long as a second or
two, longer than the minimum conscious integration time, even
that is not long enough to think through a difficult problem, to
integrate information from two domains in memory, or to do many
other things that people plainly do consciously. Even if we
assume that people can voluntarily "refresh" a conscious content
(by rehearsal, for example), there are surely structures that can
gain access to consciousness that last longer than we are likely
to voluntarily rehearse a thought. Attitudes, for example, may
last an adult lifetime, and attitudes surely must affect one's
conscious thoughts, images, and feelings. We need something else
to bridge the gap between evanescent conscious contents and
long-term knowledge structures. In Chapter 4 we fill this need
with a new construct called a "context," defined as a
representation that shapes and evokes conscious experiences, but
that is not itself conscious.
4. îThe Threshold Paradox: at what point does a global
message become conscious?ï
If it takes global broadcasting to become conscious, and if
newly global systems need to broadcast in order to recruit
support for staying on the global workspace, there is a paradox,
a kind of "Catch©22": in order to be global long enough to be
perceived as conscious, a system must first be globally
broadcast. We have suggested above (2.43) that this problem may
be fixed in two ways: by having an increasingly global hierarchy
of workspaces,or by allowing momentary access to all potential
contents, long enough to broadcast a recruiting message to other
systems, but not long enough to be recalled and reported as
conscious. This Threshold Paradox is a counterargument, but one‹j ‹
that does seem to have a possible solution.
ùú
2.74 Other unanswered questions.
Here are some questions that we have not yet touched on,
which a complete theory must address.
Some obvious ones:
(1) Why do we lose consciousness of habituated stimuli and
automatic skills? (Chapter 5)
(2) Why are we unconscious of local perceptual contexts that
help to shape our conscious percepts and images?(Chapter 4)
(3) Why are we unconscious most of the time of the
îconceptualï context, the presuppositions that form the framework
of our thoughts about the world ? (Chapters 4 and 5).
(4) Does the common idea that we have "conscious control"
over our actions have any psychological reality? Is there a
relationship between consciousness and volition? (Chapter 7.)
(5) What if anything is the difference between consciousness
and attention? (Chapter 8.)
For some questions we have no ready answers. For example:
(6) How does the item limit of Short-Term Memory fit in with
a globalist conception of consciousness? We know that with mental
rehearsal people can keep in mind 7 +/- 2 unrelated items ---
words, numbers, or judgment categories. But that fact does not
"fall out of" the GW framework in any obvious way.
(7) Why do perceptual and imaginal processes have a unique
relationship to consciousness? What is the difference between
these "qualitative" and other "non-qualitative" mental contents?
(Chapters 4 and 7).
(8) We are never conscious of only single features or
dimensions, such as specialized processors presumably provide,
but only of entire objects and events --- i.e., internally
consistent and complete îcombinationsï of dimensional features. Why
is that?
Finally:‹j ‹å
(9) When we say that "I" am conscious of something, what is
the nature of the "I" to which we refer? Or is it just a
meaningless common-sense expression? (Chapter 9.)
Obviously there is much still to do.
2.8 A summary of the argument so far.
We have explored the first detailed contrastive analysis of
conscious and unconscious phenomena. Conscious processes were
said to be computationally inefficient, but to have great range,
relational capacity and context-sensitivity. Further, conscious
events have apparent internal consistency, seriality, and limited
capacity. In contrast to all these aspects of conscious
functioning, unconscious processîorsï are highly efficient in their
specialized tasks, have relatively limited domains, are
relatively isolated and autonomous, highly diverse, and capable
of contradicting each other; they can operate in parallel, and,
taken together, unconscious processors have very great capacity.
There is a remarkable match between these contrasts and a
system architecture used in some artificial intelligence
applications, called a "global workspace in a distributed system
of specialized processors". This organization can be compared to
a very large committee of experts, each speaking his or her
specialized jargon, who can communicate with each other through
some "global broadcasting" device, a large blackboard in front of
the committee of experts, for example. If we pretend for the time
being that global messages are conscious and specialized experts
are unconscious, the whole contrastive analysis of Table 1 can be
seen to flow from this model. Model 1 yields a great
îsimplificationï of the evidence.
Encouraged by this apparently helpful model, we considered
some issues in more detail. How global is global broadcasting? We
presented six arguments in favor of the idea that conscious
(global) information is truly distributed throughout the nervous
system. We pointed out that there may be several different ways
to implement Model 1, ways that are "functionally equivalent".
Indeed several authors have made claims similar to ours in some
respects and we briefly acknowledged this literature. We ended
the chapter by setting forth predictions that flow from Model 1,
and pointing out some of its deficiencies.
The upshot of this chapter is a major simplification of the
evidence in terms of a straightforward model (Figure 2.2). As
the title of this chapter indicates, conscious experience seems
to involve mental representations that are globally distributed
throughout the nervous system, and that are internally
consistent. This is clearly not the whole story. In later‹j ‹
chapters we will discover that Model 1 needs additional features
to accomodate further contrastive analyses. Notably in Chapter 3
suggests the existence of feedback from the input and output
processors, to create a stable coalition of processors that tends
to support one conscious content over another. In Chapter 4 we
are compelled to create a role for îcontextï in our model, in order
to represent the fact that conscious experience is always shaped
and directed by an extensive unconscious framework. Later in the
book we will find that the theory leads in a natural way to a
perspective on intentions, voluntary control, attention, and
self-control.
‹ ‹
Footnotes.
(1) The wording of some of these examples has been changed
to American English.
(2) Some researchers treat biofeedback as a type of operant
conditioning. In GW theory, operant conditioning is the
acquisition of novel voluntary actions (Chapter 7).