seminars in Ttl JIUROSCIENCES, Vol2, 1990: pp 263-275

Towards a neurobiological theory of consciousness

Francis Crick and Christo] Koch
Visual awareness is afavorable form of consciousness to study are undoubtedly important but we doubt whether
neurobiologically. We propose that it takes two forms: a very they will, by themselves, ever be sufficiently
fast form, linked to iconic memory, that may be difficult to compelling to explain consciousness in a convincing
study; and a somewhat slower one involving visual attention manner. Attempting to infer the internal structure
and short-term memory. In the slower form an attentional of a very complex system using a 'black-box'
mechanism transiently binds together all those neurons whose approach (i.e. manipulating the input variables while
activiry relates to the relevant features of a single visual object. observing the output of the system) will never lead
We suggest this is done by generating coherent semi- to unique answers. In short, such methods are not
syru:hronous oscillations, probably in the 40-70Hz range. by themselves powerful enough ever to solve a
These oscillations then activate a transient short-term problem, though they are good enough to suggest
(working) memory. We outline several lines ofexperimental tentative solutions.
work that might advance the understanding of the neural Our basic idea is that consciousness depends
mechanisms involved. The neural basis of very short-term crucially on some form of rather short-term memory
memory espedally needs more experimental study. and also on some form of serial attentional
mechanism. This attentional mechanism helps sets
Key words: consciousness I attention I binding I 40 Hz of the relevant neurons to fire in a coherent semi-
oscillations I short· term memory oscillatory way, probably at a frequency in the 40-70
Hz range, so that a temporary global unity is
imposed on neurons in many different parts of the
IT IS REMARKABLE that most of the work in both brain. These oscillations then activate short-term
cognitive science and the neurosciences makes no (working) memory. In the later parts of the paper
reference to consciousness (or 'awareness'), especially we shall be mainly concerned with visual awareness.
as many would regard consciousness as the major Before approaching the problem in detail, it seems
puzzle confronting the neural view of the mind and sensible to describe our general approach to conscious-
indeed at the present time it appears deeply ness and to decide what aspects of it are best left on
mysterious to many people. This attitude is partly one side.
a legacy of behaviorism and partly because most
workers in these areas cannot see any useful way of
approaching the problem. In the last few years Prolegomenon to the study of consciousness
several books have appearedl-4 that address the
question directly but most of these 1-3 have been We make two basic assumptions. The first is that there
written largely from a functional standpoint and so is something that requires a scientific explanation.
have said rather little about neurons and other There is general agreement that we are not conscious
machinery of the brain. of all the processes going on in our heads, though
We suggest that the time is now ripe for an attack exactly which might be a matter of dispute. While
on the neural basis of consciousness. Moreover, we are aware of many of the results of perceptual
we believe that the problem of consciousness and memory processes, we have only limited access
can, in the long run, be solved only by explanations to the processes that produce this awareness (e.g.
at the neural level. Arguments at the Cognitive level "How did I come up with the first name of my
grandfather?"). In fact, some psychologists5 have
argued that we have only very limited introspective
From the Salk lnstitut8, 10010 N Torrey Pines Road, access to the origins of even higher order cognitive
La Jolla, CA 9203 7, USA, arul Computation and Neural Systemr processes. It seems probable, however, that at any
216-76, California Institute of Technology, Pasadena, CA one moment some active neuronal processes correlate
91125, USA with consciousness, while others do not. What are
©1990 by W. B Saunders Company
1044-5765/90/0204-0003$5. 0010 the differences between them?

263
 264 F. Crick and C. Koch

The second assumption is tentative: that all the pain and so on. We shall assume that self-conscious-
different aspects of consciousness, for example pain ness, that is the self-referential aspect of consciousness,
and visual awareness, employ a basic common is merely a special case of consciousness and is better
mechanism or perhaps a few such mechanisms. If left on one side for the moment. Volition and
we understand the mechanisms for one aspect, we intentionality will also be disregarded and also
will have gone most of the way to understanding various rather unusual states, such as the hypnotic
them all. Paradoxically, consciousness appears to. be state, lucid dreaming and sleep walking, unless they
so odd and, at first sight, so difficult to understand turn out to have special features that make them
that only a rather special explanation is likely to experimentally advantageous.
work. The general nature of consciousness may be 5. No neural theory of consciousness will explain
easier to discover than more mundane operations, everything about consciousness, at least not initially.
such as shape-from-shading, that could, in principle, We will first attempt to construct a rough scaffold,
be explained in many different ways. Whether this explaining some of the dominant features and hope
is really true remains to be seen. that such an attempt will lead to more inclusive and
The following topics will be left on one side or our refined models.
attitude to them stated without further discussion, 6. There is also the problem of qualia. 6 Some
for experience has shown that otherwise much time argue that certain aspects of consciousness (such as
can be frittered away in fruitless argument about whether the red I see is the same as the red you see),
them. being essentially private, cannot in principle be
1. Everyone has a rough idea of what is meant addressed by any objective, scientific study. We feel
by consciousness. We feel that it is better to avoid that this difficult issue is, for the moment, best left
a precise definition of consciousness because of the on one side. Our present belief is that it may prove
dangers of premature defmition. Until we understand possible, in the fullness oftime, to make it plausible
the problem much better, any attempt at a formal that you see red as I do (assuming that psychophysical
definition is likely to be either misleading or overly tests suggest you do). To decide whether one can
restrictive, or both. make a plausible case or not we shall need to know
2. Arguments about what consciousness is for the exact neural correlate in a human brain of seeing
are probably premature, although such an approach red. Whatever the outcome, we believe that any
may give a few hints about its nature. It is, after all, adequate theory of consciousness should explain how
a bit surprising that one should worry too much we see color at all. ·
about the function of something when we are rather This outlines the framework within which we will
vague about what it is. address the problem of consciousness. We ne.?Ct ask
3. We shall assume that some species of what psychology can tell us about the phenomenon.
animals, and in particular the higher mammals,
possess some of the essential features of consciousness,
but not necessarily all. For this reason, appropriate The cognitive approach
experiments on such animals may be relevant to
finding the mechanisms underlying consciousness. The most effective way to approach the problem of
3 .1. From this it follows that a language system consciousness would be to use the descriptions of
(of the type found in humans) is not essential psychologists and cognitive scientists and attempt to
for consciousness. That is, one can have the key map different aspects of their models onto what is
features of consciousness without language. This is known about the neuroanatomy and neurophysiology
not to say that language may not enrich conscious- of the brain. Naturally we have attempted to do this
ness considerably. but have not found it as useful as one might hope,
3. 2. We consider that it is not profitable at this although such models do point to the importance of
stage to argue about whether 'lower' animals, such attention and short-term memory and suggest that
as octopus, Drosophila or nematodes, are conscious. consciousness should have easy access to the higher,
It is probable, though, that consciousness correlates planning levels of the motor system.
to some extent with the degree of complexity of any A major handicap is the pernicious influence of
nervous system. the paradigm of the von Neumann digital computer.
4. There are many forms of consciousness, such It is a painful business to try to translate the various
as those associated with seeing, thinking, emotion, boxes labeled 'files', 'CPU', 'character buffer', and
A theory of consciousness
265
~o on o~curring in psychological models, each with ~utomated procedures, such as typing or swimming,
Its special meth.o~s ofpr~cessing, into the language IS probably largely unconscious.
of neuronal actiVIty and mteraction. This is mainly . Con~ciousness appears to be a process that is fairly
because present-day computers make extensive use 1mmed1ate. In ot~er words, it does not take too long
of precis~ly:detail~d pulse-coded messages. There is to. become conscwus of, at least, straightforward
no convmcmg evtdence that the brain uses such a thmgs. Exa~tly how long is not clear, but figures
system and much to suggest that it does not. For somewhere m the 50-250 ms time range might be
these reasons we will not attempt to comment in reasonable, rather than a few seconds.
detail on these psychological models but only make
rather general remarks.
Johnson-Laird 1 proposes that the brain is a General neurobiological aspects
complex hierarchy of largely parallel processors-
':n idea that is almost certainly along the right When
hnes-and that there is an operating system at the
top of the hierarchy. He considers the conscious mind When is an animal conscious? This turns out not
to correspond to this operating system. Thus the to be a straightforward problem. It seems likely that
mechanism of consciousness expresses the results of the essential features of consciousness are probably
some of the computations the brain makes but not not usually present in slow-wave sleep, nor under
the details of how they were done. If there is an a deep anaesthetic. Rapid Eye Movement (REM)
or dreaming sleep is another matter. It seems to us
operating system of this type it is not easy at the
that a limited form of consciousness often occurs in
moment to see any particular brain area in which
REM sleep. Although cognition is not completely
it is located. 7 Johnson-Laird also lays emphasis on
normal and memories cannot be put into the long-
self-reflection and self-awareness, topics that we have
term store, nevertheless dreams seem to have some
decided to leave on one side.
of the attributes of consciousness. Whether there are
An alternative view, due to J ackendoff2 is that
any experimental advantages in studying REM sleep
consciousness is not associated with the highest levels
to help us understand consciousness remains to be
in the hierarchy but with the intermediate levels. He
seen. At the moment we do not see any.
arrives at this point of view by considering in detail The really difficult case is that of rather light
the language system, the visual system and the music anaesthesia or states induced by modern receptor-
system. His arguments are certainly suggestive, specific anaesthetic agents. 12 It seems likely that in
though not completely compelling. some such states the brain is at least partly conscious.
Both authors emphasize the intimate relationship This would not matter were it not for the large
between consciousness and working memory. amount of experimentation on mammals in these
They also both envisage a serial process (that states. We return to this problem later.
we may loosely identify as attention) on top of
the parallel processes. Thus clearly we shall have
to consider the mechanisms of attention and of 'Where
working memory.
The type of memory that will be of most interest Where in the brain are the neural correlates of
to us is short-term memory, whether it be very short consciousness? One of the traditional answers-that
(a fraction of a second; sometimes called 'iconic' consciousness depends on the reticular activating
memory8 if it is visual) or short-term memory system in the midbrain-is misleading. Certainly the
proper (meaning a few seconds) sometimes called relevant parts of the brain need to be activated in
working memory.9,lO These are so important that various ways, but this is a little like believing that
we shall take up each of these subjects later. It does the characteristic part of a television set is its electrical
not seem essential for consciousness that the brain supply. It is more likely that the operations correspond-
should be able to put anything into the long-term ing to consciousness occur mainly (though not
episodic memory system, 11 since people with certain exclusively) in the neocortex and probably also in
brain damage (to be discussed below) cannot lay the paleocortex, associated with the olfactory system,
down new episodic memories but appear in all other since local damage to the cerebral cortex often
respects to be fully conscious. Procedural memory, 11 removes particular aspects of consciousness (as, for
i.e. that part of memory responsible for highly example, in face agnosia).
266 F. Crick and C. Koch

The hippocampal system (the allocortex) is a little that the neural pathway was subcortical, through the
more complicated. A person with complete bilateral superior colliculus, but this has recently been brought
damage to the hippocampal system and to all the into question. 16 The motor output appears to be
higher order association cortices that are connected voluntary and the subject is certainly aware what
with it, appears to have most aspects of conscious- movements he is making. It is now an urgent matter
ness13 (including short-term conjunctions over at to decide experimentally, by comparative work on
leastthree modalities) so the hippocampal system is humans and monkeys, exactly which neural pathways
unlikely to be essential. But it could be argued that are used in blindsight, since this information may
what goes on in the normal hippocampus does indeed suggest which neural pathways are used for conscious-
reach consciousness, so it may be important to ness and which not.
consider in detail the inputs and outputs of the We shall not discuss here other cases in which
hippocampal system. people respond to a stimulus such as an auditory or
Structures in the midbrain or hindbrain, such as visual cue, but claim to be unaware ofthe stimulus
the cerebellum, are probably not essential for (for example, subliminal perception and priming, 17,18
consciousness. It remains to be seen whether certain and also galvanic skin responses 19). Such stimuli
other structures, such as the thalamus, the basal may affect ongoing mental processes, including
ganglia and the claustrum, all intimately associated higher order processes, without being registered in
with the neocortex, are closely involved in conscious- working or long-term memory.
ness. We shall include these structures together with
the cerebral cortex as 'the cortical system'. Neuronal firing

What is the general character of neural behaviors

Split brains associated, or not, with consciousness? We think it
plausible that consciousness in some sense requires
The study of persons with 'split brains' gives us neuronal activity, in which we include not only
information about some of the pathways that are (or neurons that fire action potentials but also non-
are not) involved in consciousness. It is a well- spiking neurons such as amacrine cells.
established fact14 that for persons whose corpus Our basic hypothesis at the neural level is that it
callosum-the massive fiber bundle connecting the is useful to think of consciousness as being correlated
two cortical hemispheres-has been cut, the left-hand with a special type of activity of perhaps a subset of
side of the brain (for right-handed people) is not neurons in the cortical system. Consciousness can
aware of the activity in the visual system taking place undoubtedly take different fonns, depending on
on the right side, whereas in a normal person it is. which parts of the cortex are involved, but we
(We leave on one side the somewhat controversial hypothesize that there is one basic mechanism (or
matter as to whether the right side, on its own, a few such) underlying them all. At any moment
is as conscious as the left side.) This strongly consciousness corresponds to a particular type of
suggests that some of the information associated with activity in a transient set of neurons that are a subset
consciousness can traverse the normal corpus of a much larger set of potential candidates. The
callosum. It also suggests that such information, with problem at the neural level then becomes:
the exception of some emotional states, cannot be
transmitted from one side of the cortex to the other 1. Where are these neurons in the brain?
by the subcortical pathways that remain intact in this 2. Are they of any particular neuronal type?
operation. 3. What is special (if anything) about their
connections?
4. What is special (if anything) about the way they
Blindsight are firing?

It now seems generally agreed that blindsight is a Visual awareness

genuine phenomenon. 15 Certain people with cortical
blindness can point fairly accurately to the position At this point we propose to make a somewhat arbitrary
of objects in their blind visual field, while denying personal choice. Since we hypothesize that there is
that they see anything. It was previously suspected a basic mechanism for consciousness that is rather
A theory of consciousness
267
similar in different parts of the brain (and in position in space relative to the head or the body.
particular, in different parts of the neocortex): we So far no single area has been found whose neurons correspond
propose that the visual system is the most favorable to everything we see. How is it, then, that we seem to
for an initial experimental approach. The visual have a single coherent visual picture of the scene
system has several well-known advantages as an before us?
experimental system for investigating the neuronal The obvious suggestion is that at any one time the
basis of consciousness. Unlike language, it is fairly relevant neurons in many cortical areas cooperate
similar in man and the higher primates. There is together to form some sort of global activity and that
already much experimental work on it, both by this global activity corresponds to visual awareness.
psychophysicists and by neuroscientists. Moreover This is the hypothesis that we shall try to ft.ll out with
we believe that it will be easier to study than the more psychological and neurobiological details.
complex aspects of consciousness associated with self-
awareness. From now on, then, we shall discuss not
consciousness in all its aspects but visual awareness Convergence zones
and only the awareness that does not involve the
laying down oflong-term episodic memory. Johnson- Which particular cortical areas contribute directly
Laird1 has called this 'bare awareness', which does to visual awareness? For example, is the first visual
indeed convey the idea that it is simpler than self- area, Vl, closely involved? Damasio7 has also
awareness. Although here we mainly consider visual suggested that many cortical areas take part at any
perception, many of the processes we discuss are also moment at several different levels. It is probably
likely to be used in visual inagery. significant that Damasio arrived at this idea largely
from considering the detailed effects of various types
of brain damage, evidence rarely used by other
The mammalian visual system theorists. He suggests that the synchronous activity
in the various cortical areas is coordinated by
The visual system of mammals is complex and we only feedback projections from 'convergence zones'.
give an outline description. In the macaque monkey What exactly convergence zones means in neural
there are many distinct visual areas in the neocortex, terms is somewhat obscure. We suggest that
perhaps as many as two dozen.20 One reason for convergence zones may mainly refer to the neurons
these multiple areas is that to handle all activity in one (or a subset of them) that project 'backwards' (see
single very large neural net, with everything connected Figure 1), such as those that project from the second
to everything else, would make the brain both visual area back to the first one. Thus the term zones
cumbersome and prohibitively large. Loosely speaking may turn out to convey a somewhat misleading
these areas are connected in several hierarchies, with impression, since such neurons may be sprinkled
the processing within any one area being largely fairly uniformly through certain layers of the
parallel (see Figure 1). There are also many back neocortex. Nevertheless the basic idea is a valuable
projections (and some cross connectio~s) between one. It points to the possibility that, in addition,
cortical areas, and also from the cortex back to the convergence zones may also exist in other places,
thalamus, the functions of all of which are unknown. such as the thalamus or the claustrum. (The
The dominant output of the retina projects- claustrum2 1 is a thin sheet of cells located just below
through the lateral geniculate nucleus of the the neocortex of the insula. It receives axons from
thalamus-to the primary visual cortex at the back almost all cortical areas and, in return, projects back
of the brain. This area has neurons responding to to them. Its caudal region is largely visual.)
fairly simple features (such as an oriented edge) One would certainly expect areas near to the hippo-
occupying a tiny part of the visual field. campus, such as the inferotemporal region, as well as
The neurons in the higher cortical areas respond those near to the higher levels of the motor system
to more complex features, e.g. aspects of faces, and to be closely involved in visual awareness. How far
their receptive fields cover much larger areas. The awareness is directly correlated with every neocortical
different cortical areas respond, in general, to level of the visual system remains uncertain, as is
different features. For example, the neurons in area the involvement of the thalamus, although certain
MT are mostly interested in motion and depth, those parts of it, such as its reticular nucleus 22 and the
in area V4 in color and shape, and those in 7a in pulvinar23 may be involved with attention.
268 F. Crick and C. Koch

(FORM) (ATTENTION) (MOTION) (EYE

MOVEMENT)

Figure 1. Some of the major cortical and sub-cortical visual areas in the macaque monkey.
Major pathways are indicated by bold lines, minor pathways by thin lines; dashed lines weak
or uncertain. The shaded ellipsoids correspond to higher-level visual-motor areas. Most cortical
areas project back to the superior colliculus as well as to the different maps in the pulvinar, which
is part of the thalamus. Furthermore, every forward projection has usually an equally strong,
if not stronger, back-projection associated with it (not shown). These feedback pathways usually
terminate outside cortical layer IV, while the forward projection terminates most densely in
this layer. Areas related to the control and expression of visual attention include the superior
colliculus and the pulvinar, as well as areas POa and 7a, part of the posterior parietal cortex.
Abbreviations: FEF-frontal eye fields; IT-inferior temporal lobe; LGN-lateral geniculate
nucleus; MT, MST-motion-processing areas; Vl-primary visual area; V2-4-higher-order
visual processing areas; YIP-ventral interparietal area; VP-ventral posterior. The figure is
from D. Van Essen andj. Maunsell, personal communication. For more information, see ref20.

Binding and selective visual attention specialized neurons responding to very specific and
ecological highly significant objects, such as the
The binding problem neurons responding to aspects of faces in infero-
tem poral cortex of primates. 24
There are an almost infinite number of possible, It seems likely that at any moment any individual
different objects that we are capable of seeing. There object in the visual field is represented by the firing
cannot be a single neuron, often referred to as a of a set of neurons. This would not cause any special
grandmother cell, for each one. The combinatorial problem if the members of the set were in close
possibilities for representing so many objects at all proximity (implying that they probably interact
different values of depth, motion, color, orientation somewhat), received somewhat similar inputs and
and spatial location are simply too staggering. This projected to somewhat similar places. But because
does not preclude the existence of sets of somewhat any object will have different characteristics (form,
 A theory of consciousness 269

color, motion, orientation) that are processed in an almost unlimited potential capacity, although its
several different visual areas, it is highly reasonable capacity at any one time may be limited. If a
to assume that seeing any one object often involves particular stimulus is repeated frequently, this third
neurons in many different visual areas. The problem type of transient binding may eventually build up
of how these neurons temporarily become active as the second, overlearned type of binding.
a unit is often described as 'the binding problem'.
As an object seen is often also heard, smelt or felt,
this binding must also occur across different sensory The role of attention25-28
modalities.
Our experience of perceptual unity thus suggests This form of transient binding probably depends on
that the brain in some way binds together, in a a serial attentional mechanism, sometimes called the
mutually coherent way, all those neurons actively spotlight of attention. 25, 28 It is thought by some to
responding to different aspects of a perceived object. concentrate on one place in the visual field after
In other words, if you are currently paying attention another, possibly moving every 60 ms or so. It is
to a friend discussing some point with you, neurons faster than eye-movements (another, slower, form of
in area MT that respond to the motion of his face, attention), and can work across different spatial scales.
neurons in area V4 that respond to its hue, neurons We suggest (the idea goes back to the last century 29 )
in auditory cortex that respond to the words coming that what reaches visual awareness is usually the
from his face and possibly the memory traces result of this attentional step. In other words, that
associated with recognition of the face all have to be awareness and attention are intimately bound together. Note
'bound' together, to carry a common label identifying that although the results of attention are postulated
them as neurons that jointly generate the perception to reach consciousness, the attentional mechanisms
of that specific face. themselves are probably largely unconscious.

Short-term memory
Types qf binding
It has long been argued that awareness is not only
Binding can be of several types. In a sense a neuron associated with attention but also with some form of
responding to an oriented line can be considered to short-term memory. We have to distinguish two
be binding a set of points. The inputs to such a neuron forms, iconic and working memory. Iconic memory
are probably determined by genes and by develop- is similar to a sensory input buffer, storing information
mental processes that have evolved due to the mainly at a pre-categorical level, i.e. in terms of
experience of our distant ancestors. Other forms of simple visual primitives such as orientation or
binding, such as that required for the recognition movement, although it may also involve more
of familiar objects such as the letters of a well-known complex features, such as familiar words. It has a
alphabet, may be acquired by frequently repeated very large capacity but decays very quickly, perhaps
experience; that is, by being overlearnt. This probably in half a second or less.8 Working memory, in
implies that many of the neurons involved have as a contrast, may last for seconds; 9 • 10 it seems to have
result become strongly connected together. (Recall a rather limited capacity (a figure often quoted is
that most cortical neurons have many thousands of seven items30), is different for different modalities
connections and that initially many of these may be and seems to use a much more abstract
weak.) Both these types of binding are likely to have representation (post-categorical memory). It can be
a large but limited capacity. These are the types of prolonged by rehearsal, as when one rehearses a
binding with which Damasio 7 is mainly concerned. telephone number. We suspect that it is this form
The binding we are especially concerned with is of memory that is strongly activated in the binding
a third type, being neither epigenetically determined process.
nor overlearnt. It applies particularly to objects
whose exact combination of features may be quite
novel to us. The neurons actively involved are Neural basis of short-term memory
unlikely all to be strongly connected together, at least
in most cases. This binding must arise rapidly. By The neural basis of these two forms of memory is
its very nature it is largely transitory and must have much understudied. Three broad types of mechanisms
270 F. Crick and C. Koch

might underlie either iconic or working memory, by activity in the postsynaptic cell. For this reason
either singly or in combination: they are probably all non-Hebbian 36 in character.
('Hebbian' means that the alteration in synaptic
(i) the strength of certain synapses is temporarily strength depends on activity, at about the same time,
increased or decreased; in both the presynaptic and the postsynaptic cell.)
(ii) a neuron keeps on firing for some time due Non-Hebbian implies that the change is not (to a
mainly to its intrinsic biophysical properties; first approximation) just at an individual synapse but
(iii) a neuron keeps on firing mainly due to at all the synapses associated with that particular
extrinsic circuit properties ('reverberatory presynaptic cell, so that the 'unit of change' is the
circuits'). presynaptic cell, not a single synapse. Such changes
Intracellular recordings show that the cell bodies may form part of the neuronal basis of(short-term)
of retinal ganglion cells remain depolarized for priming. 17
200 ms or longer following a very brief flash of light Von der Malsburg3 7 in 1981 postulated on
within their receptive field. 3 1 How neurons further theoretical grounds a fast (fraction of a second)
in the system express a transient memory remains mechanism for increased synaptic efficiency, having
to be seen. In a memory ta~k in which a trained a broadly Hebbian36 character. So far this transient
monkey had to discriminate and retain individual Hebbian alteration to synaptic strength has not been
features of compound stimuli, about 15% of the observed, although ifit existed it would provide very
neurons in inferotemporal cortex continued to frre considerable theoretical advantages.
for 10-20s after the stimulus. 3 2 The mechanism We know almost nothing concerning the anatomical
producing (and terminating) this sustained firing is localization of iconic and working memory, except
unknown. It is not yet known whether this firing is that (as discussed above 13) the hippocampal system
oscillatory or not. It may not have to be oscillatory, need not be present in order to remember for short
since a task can be remembered well enough without times. Both iconic and working memory are likely
one being vividly conscious of it all the time. to be distributed throughout the appropriate cortical
A phenomenon that may correspond to the first areas, with auditory events transiently stored in
mechanism was discovered many years ago at the auditory cortices, visual events in the visual cortices,
neuromuscular junction and called post-tetanic and so on. No case of a person who is conscious but
potentiation. This is not to be confused with the has lost all forms of short-term memory has been
much studied long-term potentiation (LTP)33 that reported.
has a longer time course. The terminology for these
short-lasting changes has now become more compli·
cated, 34·35 with effects variously labelled facilitation Neuronal oscillations
(with two time constants of decay, of 50 ms and
300 ms), augmentation (7 s) and potentiation (30 s It has often been hypothesized, in particular by von
to minutes). There is also depression (5 s). We shall der Malsburg, 37 that in neural terms binding means
refer to all these collectively as 'short-term synaptic the temporarily correlated firing of the neurons
modification'. Note that the decay times are in involved. In other words, neurons in different parts
general in the same range as the times required for of cortex responding to the currently perceived object
the decay ofworking memory. Short-term modification fire action potentials at about the same time. The
has also been observed in the hippocampus and, in main theoretical requirement is for correlated firing.
one case, in the neocortex. These temporary changes It now seems theoretically plausible that this is most
in synaptic strength can be surprisingly large (see easily implemented using oscillations. There are
figures 11 and 13 in ref 33). many oscillations known in the cortex, among them
the a-rhythm and the 8-rhythm of the hippocampus.
Experimental work on '"¥·oscillations' in mammalian
Hebbian or non-Hebbian? olfactory cortex38,39 and in the visual cortex40·46
indicates that an important mechanism may involve
The causes of the various aspects of short-term neurons firing in semisynchrony at a frequency in
modification are not fully understood but many of the 40-70Hz range (at least in the cat) (see Figure 2).
them seem to result from calcium accumulation in In the visual cortex, only a proportion of the active
the presynaptic terminal34,35 and are not influenced neurons ·fire in this way. Most of the neurons that
 A theory of consciousness
271
The activity of the relevant neurons in both the first
and second visual areas has also been shown to
be correlated. 44 We shall refer to these neuronal
oscillations as the 40Hz oscillations, though the
frequency on any one occasion is not very precise
and can vary between 35 and 75Hz.
Most of these experiments have been done on
lightly-anaesthetized cats whose exact state of
awareness is unknown but as similar oscillations have
also been seen in neurons in alert cats,46 they are
unlikely to be merely an artifact of the anaesthetic.

A sketch of a theory

The role of the oscillations in binding may be as

follows. Objects in the visual field give rise to
appropriate responses in the appropriate cortical
areas. Visual attention now has to select one of these
objects at a particular location. Exactly how this
works is not yet known. One possible solution25,28,47
is that the brain has some sort of topographic saliency
map that codes for the conspicuousness oflocations
,. 1200 uV in the visual field in terms of generalized center-
_ J 100 uV surround operations. This map would derive its
500 ms input from the individual feature maps and give a
20 ms very 'biased' view of the visual environment,
Figure 2. The local field potential (upper trace) and emphasizing locations where objects differed in some
multiunit activity (second trace) recorded from an electrode perceptual dimension, i.e. color, motion, depth, from
in the first visual area of an adult cat in response to an objects at neighboring locations. Where this map
optimally oriented light bar moving across the receptive might be located is unclear but parts ·of the
field. The third and fourth traces show a corresponding thalamus,2 2,23 such as the pulvinar, might be in-
part of the first and second ones (respectively) on an
expanded time scale. Note the presence of rhythmic volved. (Note that the pulvinar receives a projection
oscillations (35-45Hz). The multiunit activity occurs near from the superior colliculus, an area that could
the peak negativity of the local field potential. be thought of as having a saliency map for eye
Reproduced, by permission, from ref 42. movements.)
Once a particular salient location has been selected,
do oscillate appear to be complex cells rather than probably by a winner-take-all mechanism, the
simple cells, so that very few are found in cortical information associated with it must be activated by
layer IV.45 referring back to the individual feature maps. The
It has been shown41 in cat primary visual cortex various parts of the visual thalamus are well suited
that in many cases two neurons some distance apart, to influence the behavior of neurons in those parts
but both responding to a particular moving bar, can of the neocortex from which they receive-information
show frequency-locked oscillations with no measurable and to which they project back. Some such feedback
phase shift in their cross-correlations. More tellingly, pathways then activate or synchronize the oscillations
the activity of two relevant neurons, some 7 mm at the corresponding locations in areas such as Vl,
apart, showed 41 more highly correlated oscillations V 4 and MT, so that a coherent set of features is
(at about 50 Hz) to a single moving bar than did the bound. Of course, the system had to decide, using
same two neurons stimulated by two shorter, the categorical knowledge stored in the connections,
separate, moving bars. That is, the correlation was exactly which neurons must oscillate together to
much greater in response to an obviously single produce a veridical representation of the object being
object than to a pair of somewhat distinct objects. attended to. This is probably not an easy task. These
272 F. Crick and C. Koch

are the problems that concern much theoretical The likelihood that only a few simultaneous,
work in cognitive science and computational neuro- distinct oscillatiops can exist happily together might
science. It is well known that making sense of our explain, 49 in a very natural way' the well-known
perceptual inputs is an 'ill-posed' problem and much limited capacity of the attentional system. The 40Hz
'computation • must be done to produce veridical oscillations may also be an important element in
solutions. laying down and improving stored, categorical
It seems likely that, for one reason or another, information, because at any moment they represent
certain neurons in the cortical areas involved tend the brain's consensus as to what it is seeing.
to oscillate at around 40Hz. This could make it
easier to activate such oscillations quickly. The effects Fleeting awareness
of the 40 Hz oscillations is probably greater than the
same amount of random firing for two reasons: first, Can a spotlight of attention, moving over the visual
because spikes arriving at a neuron simultaneously field from one 1 salient' location to the next, explain
will produce a larger effect at the soma; and second, the perceptual richness of our environment? Would
because the 40Hz oscillations may promote 40Hz such a mechanism not lead to a sort of 'tunnel
'resonances' elsewhere. vision', in which the currently attended location
To be effective the phases of the relevant neurons appears in vivid detail with its associated perceptual
must be synchronized. A recent mathematical attributes while everything else is invisible? We
analysis48 shows that a set of oscillating neurons can suggest, very tentatively, that this richness may be
very quickly be frequency- as well as phase-locked mediated by another form of awareness that is very
if there is a centralized feedback unit. On the other transient, being associated with iconic memory and
hand phase-locking is much slower and more difficult having a very large capacity at any one time. This
to achieve if the interactions have to pass over long form, that we propose to call 'fleeting awareness',
chains of interneuronal connections. we expect not to solve the ad hoc binding problem
As there has probably been strong evolutionary (as working awareness does) but to embody 'features'
pressure to produce appropriate phase-locking as that are bound only epigenetically or by overlearning.
quickly as possible, it may be a general principle that Attention can then focus on a subset of relevant items
all phase-locking interactions use only pathways with
within iconic memory for further processing. Because
as few links as possible; these may be within a cortical
fleeting awareness is very transient it may be
area, between cortical areas (in adjacent levels in the
especially difficult to study. Whether it really exists
hierarchy), or to and from one or more global
remains to be seen.
coordinating regions, such as the thalamus or the
claustrum.
We suggest that one of the functions of consciousness Experimental problems
is to present the result of various underlying
computations and that this involves an attentional To understand working awareness it seems clear that
mechanism that temporarily binds the relevant experimental studies of three somewhat different
neurons together by synchronizing their spikes in phenomena are needed. The first is the 40Hz
40Hz oscillations. These oscillations do not oscillations, and indeed all other oscillations and any
themselves encode additional information, except in other kinds of coherent firing. In what cortical areas
so far as they join together some of the existing are oscillations found, especially in alert animals?
information into a coherent percept. We shall call What is their natural history: when do they occur?
this form of awareness 'working awareness'. We How long do they take to set up? How long does
further postulate that obJects for which the binding problem any one of them last? Are there, simultaneously,
has been solved are placed into working memory. In other several different frequencies and/or phases? Are they
words, some or all of the properties associated with indeed mainly correlated with objects? Do they solve
the attended location would automatically be the well-known figure-ground problem? Do they ever
remembered for a short time. One very attractive occur in the thalamus or the basal ganglia? Is the
possibility, with no experimental support, is that the claustrum21 involved-it is ideally placed to help
40 Hz oscillations themselves preferentially activate synchronize neurons in different places?
the mechanisms underlying working memory and The second is attention. Are any of the oscillations
possibly other longer-term forms of memory as well. associated with it? How does the mechanism work
 A theory of consciousness 273
in neural terms? It is already known that neurons in with the monkey's apparent percept. A similar
area V 4'0 and in the parietal region51 in the macaque experiment involving horizontal gratings moving
monkey are influenced by attention. Evidence on either upwards (in one eye) or downwards (in the
humans with brain damage implicate parts of the other) produces a similar variety of responses in area
thalamus in certain aspects of attention, 22 as do MT of the monkey. 52,53
certain experiments on monkeys, 23 but we need to The second approach, 54 using cats, is to let the
know much more about how attention works animal view the same set of visual signals first while
neurobiologically. it is awake and then again when it is in slow-wave
The third is the neural basis of both iconic and sleep. The neurons studied were mainly in the first
working memory. These are neglected subjects that visual area. The general result is that the response
urgently demand an experimental attack. If there of any given neuron is broadly the same in the two
are reverberating circuits, where are they? Under conditions but neurons in the lower layers are often
what circumstances do they oscillate? Are there markedly less active in slow-wave sleep. This was
special mechanisms within certain neurons that make confirmed by experiments in which activity was
it easy for them to continue firing? How widely is monitored histochemically using deoxyglucose. (Note
the short-term synaptic modification mechanism that all this was done before the 40 Hz oscillations
distributed? What are its characteristics in different were discovered.) Both these types of experiments
places? Is there any sign of a Hebbian form of are obviously in their infancy but they certainly
modification? What exactly are the biochemical and suggest that visual awareness can be directly
biophysical mechanisms34,35 underlying it? In approached by suitably designed experiments.
addition we need to relate the neural activity to A third approach to working awareness would be
psychological observations. to study the effect of anaesthetics on awareness and
We plan to discuss these experimental questions recall in humans and on neuronal responses in monkey
more fully in a later paper and to offer some tentative under similar conditions. Surprisingly, no reliable
answers based on more detailed theoretical arguments. means exists today to test whether a patient is uncon-
scious during modern anaesthetic procedures, 12
Some experimental approaches which usually includes muscle relaxants to block
voluntary and involuntary movements. Even the
There are at least three possible methods of directly EEG is no reliable indicator of depth of anaesthesia,
approaching working awareness. One is to exploit as it is often not distinguishable from that of a sleeping
the process known as rivalry, an example of which person. It has been claimed, however, that the
is the two alternative ways of viewing the well-known disappearance of a 40 ~z wave in the auditory evoked
Necker cube drawing. Here the visual input is potential correlates with loss of consciousness under
constant but the percept varies from one interpretation anesthesia. 55 Four to six cycles of such an oscillation,
to the other. A form of rivalry that is easier to study with a period of about 25 ms, can be seen following
neurophysiologically is binocular rivalry, where one stimulation by a single or a series of sharp clicks. 56
eye is given one constant visual input and the other More work on this is obviously very desirable.
a constant but rivalrous one. An example studied by Much useful information can also be obtained by
John Allman and his colleagues (personal communi- careful studies on humans with brain damage who
cation) involved projecting a small horizontal grating have selective impairment of different aspects of visual
into one eye of a macaque monkey and a small perception7,57 and especially of attention and binding.
vertical one into the other eye so that their visual
fields overlap. In the same circumstances people see Conclusion
first one grating and then the other, the two percepts
alternating every second or so. Detailed psychophysical Our tentative theory, most of the elements of which
studies on both monkeys and humans, where the have already been proposed by others, is a program
monkeys signalled the change by pressing one oftwo for research rather than a detailed model.
keys, give fairly similar results, so it is reasonable What are the essential features of visual
to assume that the monkey has an awareness of the awareness? The first requirement is for a form of
changing percepts. Most neurons in the first visual running short-term memory, an idea that is certainly
area fire in response to the constant visual input and more than 100 years old (see ref29)., We postulate
do not change but the firing of others tends to change two distinct forms of this running memory, a very
 274 F. Crick and C. Koch

transient iconic one that records built-in visual Acknowledgements

features and a working memory that lasts for a
somewhat longer time and can also store com- We thank John Allman, Bernard Baars, Patricia
binations of features. We presume that parts Churchland, Paul Churchland, Antonio Damasio,
of iconic memory form the basis of working Charles Gray, BelaJulesz, Dan Kammen, Georg Kreisel,
memory. and Leslie Orgel for helpful comments on earlier versions
of the manuscript and thank Jennifer Altman especially
The information about a single object is distributed
for extensive improvements to the sumitted manuscript.
about the brain. There has, therefore, to be a way
F.C. is supported by the]. W. Kieckhefer Foundation.
of imposing a temporary unity on the activities of C.K. is supported by the Air Force Office of Scientific
all the neurons that are relevant at that moment. Research, a Presidential Young Investigator Award from
(Incidentally we see no reason at all why this global NSF and by the James S. McDonnell Foundation.
unity should require fancy quantum effects.) The
achievement of this unity may be assisted by a fast
attentional mechanism, the exact nature of which is References
not yet understood. This mechanism is postulated
to concentate on one object at a time, choosing by 1. Johnson·Laird PN (1988) The Computer and the Mind.
a winner-take-all process the next object that appears Harvard Univ Press, Cambridge, MA
2. JackendoffR (1987) Consciousness and the Computational
to it to be the most salient. The required unity takes Mind. MIT Press, Cambridge, MA
the form of the relevant neurons firing together in 3. Baars BJ (1988) A Cognitive Theory of Consciousness.
semi-synchrony, probably at a frequency in the Cambridge University Press, Cambridge, UK
4. Edelman GR (1989) The Remembered Present. Basic
40-70Hz range. We tentatively suggest that this Books, New York
activates the appropriate parts of the working- 5. Nisbett RE, Wilson TD (1977) Telling more than we can
memory system. The neural basis of this memory know: verbal reports on mental processes. Psycho! Rev
84:231-259
system is at the moment obscure, although the 6. Churchland PM (1985) Reductionism, qualia and the direct
transient alteration of synaptic strengths is one likely introspection of brain states. J Philos 82:8-28
mechanism. 7. Damasio AR ( 1989) The brain binds entities and events
by multiregional activation from convergence zones. Neural
There is also much neural activity in the visual Computat 1:123-132
system that does not reach full awareness. Much of 8. Coltheart M (1983) Iconic memory. Philos Trans R Soc
this corresponds to the computations needed to arrive Lond B 302:283-294
9. Baddeley A (1986) Working Memory. Oxford University
at the best interpretation of all the incoming Press, Oxford
information that is compatible with the stored, 10. Phillips WA (1983) Short-term visual memory. Philos Trans
categorical information acquired in the past. It R Soc Lond B 302:295-309
11. Tulving E (1985) Memory and consciousness. Can J
is this 'best interpretation' of which we become Psycho! 26:1-12
aware. 12. Kulli J, Koch C (1990) Does anesthesia cause Joss of
Why, then, is consciousness so mysterious? A consciousness? Trends Neurosci, in press
13. Damasio AR, Eslinger PJ, Damasio H, Van HoesenGH,
striking feature of our visual awareness (and of Cornell S (1985) Multimodal amnesic syndrome following
consciousness in general) is that it is very rich in bilateral temporal and basal forebrain damage. Arch Neural
information, even if much of it is retained for only 42:252-259
14. Geschwind N, Galaburda AM (1986) Cerebral Lateraliz-
a rather brief time. Not only can the system switch ation. MIT Press, Cambridge, MA
rapidly from one object to another, but in addition 15. Weiskrantz L (1986) Blindsight. Oxford University Press,
it can handle a very large amount of information in Oxford
16. Stoerig P, Cowey A (1989) Wavelength sensitivity in
a coherent way at a single moment. We believe blindsight. Nature 342:916-918
it is mainly these two abilities, combined with 17. Tulving E, Schacter DL (1990) Priming and human
the very transient memory systems involved, memory systems. Science 247:301-306
that has made it appear so strange. We have no 18. Kihlstrom JF (1987) The cognitive unconscious. Science
237:1445-1452
experience (apart from the very limited view 19. Tranel D, Damasio AR (1985) Knowledge without
provided by our own introspection) of machines awareness: an autonomic index of facial recognition by
prosopagnosics. Science 228:1453-1454
having complex, rapidly changing and highly 20. Van Essen D (1985) Functional organization of primate
parallel activity of this type. When we can both visual cortex, in Cerebral Cortex, vol3, Visual Cortex
construct such machines and understand their detailed (Peters A, Jones EG, eds), pp 259-329. Plenum Press, New
York
behavior, much of the mystery of consciousness may 21. Sherk H (1986) The claustrum and the cerebral cortex, in
disappear. Cerebral Cortex, vol5, Sensory-motor Areas and Aspects
 A theory of conscz"ousness 275

of Cortical Connectivity Qones EG, Peters A, eds), synchronization which reflects global stimulus properties·.
pp 467-499. Plenum Press, New York Nature 338:334-337
22. Rafal RD, Posner MI (1987) Deficits in human visual 42. Gray CM, Singer W (1989) Stimulus-specific neuronal
spatial attention following thalamic lesions. Proc Nat! Acad oscillations in orientation columns of cat visual cortex. Proc
Sci USA 84:7349-7353 Nat! Acad Sci USA 86:1698-1702
23. Petersen SE, Robinson DL, MorrisJD (1987) Contributions 43. Eckhom R, Reitboeck HJ (1989) Stimulus-specific synchron·
of the pulvinar to visual spatial attention. Neuropsychologia izations in cat visual cortex and their possible role in visual
25:97-105 pattern recognition, in Synergetics of Cognition, Springer
24. Desimone R, Unger!eider LG (1989) Neural mechanisms Series in Synergetics, vol43 (Haken H, ed), pp 99-111.
of visual processing in monkeys, in Handbook of Neuro- Springer, Heidelberg
psychology, vol 2 (Damasio AR, ed), pp 267-299. Elsevier, 44. Eckhorn R, Bauer R, Jordan W, Brosch M, Kruse W,
Amsterdam Munk M, Reitboeck HJ (1988) Coherent oscillations: a
25. Treisman A (1988) Features and objects: the fourteenth mechanism of feature linking in the visual cortex? Bioi
Bartlett Memorial Lecture. QJ Exp Psychol40A:201-237 Cybern 60:121-130
26. Julesz B (1981) Textons, the elements of texture perception, 45. Gray CM, Engel AK, Konig P, Singer W (1990) Stimulus·
and their interactions. Nature 290:91-97 dependent neuronal oscillations in cat visual cortex.
27. Posner MI (1986) Chronometric Explorations of Mind. Receptive field properties and feature dependence. Eur J
Oxford University Press, Oxford, UK Neurosci 2:607-619
28. Wolfe JM, Cave KR, Franzel SL (1989) Guided search: 46. Gray CM, Raether A, Singer W (1989) Stimulus-specific
an alternative to the feature integration model for visual intercolumnar interactions of oscillatory neuronal responses
search. J Exp Psycho! 15:419-433 in the visual cortex of alert cats. Soc Neurosci Abstr
29. James W (1981) The Principles of Psychology. Harvard 15:320.4
University Press, Cambridge, MA 47. Koch C, Ullman S (1985) Shifts in selective visual attention:
30. Miller GA (1956) The magical number seven, plus or minus towards the underlying neural circuitry. Human Neurobiol
two: some limits on our capacity for processing information. 4:219-227
Psycho! Rev 63:81-97 48. Kammen DM, Holmes PJ, Koch C (1989) Cortical
31. Baylor DA, Fettiplace R (1979) Synaptic drive and impulse architecture and oscillations in neuronal networks: feedback
generation in ganglion cells of turtle retina. J Physiol Land versus local coupling, in Models of Brain Function
288:107-127 (Cotterill RMJ, ed), pp 273-884. Cambridge University
32. Fuster JM, Jervey JP (1981) Inferotemporal neurons Press, Cambridge
distinguish and retain behaviorally relevant features of visual 49. Stryker MP (1989) Is grandmother an oscillation? Nature
stimuli. Science 212:952-955 338:297-298
33. Brown TH, Ganong AH, Kairiss EW, Keenan CL, Kelso SR 50. Spitzer H, Desimone R, Moran] (1988) Increased attention
(1989) Long-term potentiation in two synaptic systems of enhances both behavior and neuronal performance. Science
the hippocampal brain slice, in Neural Models of Plasticity 24-0:338-240
(ByrneJH, Berry WO eds), pp 266-306. Academic Press, 51. Andersen RA (1987) Inferior parietal lobule function in
New York spatial perception and visuomotor integration, in Handbook
34. Magleby KL (1987) Short-term changes in synaptic efficacy, of Physiology: The Nervous System V (Mountcastle VB,
in Synaptic Function (Edelman GM, Gall WE, Cowan Plum FS, Geiger SR, eds), pp 483-518. American Physiol
WM, eds), pp 21-56. Wiley, New York Society, Bethesda, MD
35. Zucker RS (1989) Short-term synaptic plasticity. Annu Rev 52. Logothetis NK, Schall JD (1989) Neuronal correlates of
Neurosci 1~:13-31 subjective visual perception. Science 245:761-763
36. Brown TH, Kairiss EW, Keenan CL (1990) Hebbian 53. Myerson], Miezin F, Allman] (1981) Binocular rivalry
synapses: biophysical mechanisms and algorithms. Annu in macaque monkeys and humans: a comparative study in
Rev Neurosci 13:475-511 perception. Behav Anal Lett 1:149-159
37. Von der Malsburg C, Schneider W (1986) A neural 54. Livingstone MS, Hubel DH (1981) Effects of sleep and
cocktail-party processor. Bioi Cybern 54:29-40 arousal on the processing of visual information in the cat.
38. Freeman WJ (1978) Spatial properties of an EEG event in Science 291:554-561
the olfactory bulb and cortex. Electroencephalogr and Olin 55. Madler C, Peppel E (1987) Auditory evoked potentials
Neurophysiol 44:586-605 indicate the loss of neuronal oscillations during general
39. Wilson MA, Bower JM (1990) Cortical oscillations and anaesthesia. Naturwiss 74:8.42
temporal interactions in a computer simulation of piriform 56. Galambos R, Makeig S, Talmachoff PJ (1981) A 40-Hz
cortex. J Neurophysiol, in press auditory potential recorded from the human scalp. Proc Natl
40. Freeman WJ, van Dijk BW (1987) Spatial patterns of visual Acad Sci USA 78:2643-2647
cortical fast EEG during conditioned reflex in a rhesus 57. Marcel AJ (1983) Conscious and unconscious perception:
monkey. Brain Res 422:267-276 an approach to the relations between phenomenal
41. Gray CM, Konig P, Engel AK, Singer W (1989) Oscillatory experience and perceptual processes. Cog Psych 15:
responses in cat visual cortex exhibit inter-columnar 238-300