p&mJRJz VOL.30 I4
JULY 1983
COMMENTARY
III
The
function of dream sleep
Francis
Crick* & Graemi Mitchison*
We propose that the function
of
dream sleep (more proper& rapid-&ye movement or REM sleep) LT to
remove certain undesirable modes
of
interaction in networks
of cells
in the cerebral cortex.
We
postulate that this is done in REMsleep by a revem learning mechanism (see also p. 158), so that the
trace
in the brain
of
the unconscipus dream is weakened, rather than strengthened, by the dream.
MANKIND
has always been fascinated by
dreams. As might be expected, there have
heen many attempts to assign a purpose Or
significance to them. Although we dream
for one or two hours every night, we do not
remember most of our dreams. Earlier
thinkers, such as Freud, did not know this.
Modern theories (not reviewed here in
detail) have usually proposed that sleep
and dreams save energy or have various
restorative functions, either to replenish
the brain biochemically in some way, or to
reclassify or reorder the information stored
in it.
Sleep is of several kinds. Dream sleep, or
rapid eye movement (REM) sleep, is
predominantly found in viviparous mam-
mals and birds. It seems to be associated
with homeothermy (a constant internal
temperature) and the possession of an
appreciable neocortex or its equivalent.
It is not unimportant because of the ap-
preciable amount of time we spend in this
Peculiar state.
We propose here a new explanation for
the function of REM sleep. The basis of
our theory is the assumption that in
viviparous mammals the cortical system
(the cerebral cortex and some of its’
associated subcortical structures) can be
regarded as a network of interconnected
Cells which can support a great variety of
modes of mutual excitation. Such a system
is likely to be subject to unwanted or
‘Parasitic’ modes of behaviour, which arise
as it is disturbed either by the growth of the
brain or by the modifications produced by
experience. We propose that such modes
are detected and suppressed by a special
mechanism which operates during REM
sleep and has the character of an active pro-
cess which is, loosely speaking, the op-
posite of learning. We call this ‘reverse
learning* or ‘unlearning’. This mechanism,
which is not the same as normal forgetting,
*TheSalk
Institute
10010NorthTorrey PinesRoad, La
JOlkX California 92037. USA. Present address (GM.):
MRC Laboratory of Molecular Biology and the
KmWh Craik Laboratory, Cambridge CBZ 3EG. UK.
is explained in more detail below. Without
it we believe that the mammalian cortex
could not perform so well.
We first describe our ideas about the cor-
tex followed by a brief account of neural
networks. Next we outline what is known
about REM sleep. (For general accounts,
see refs 1,2.) We then describe our
postulated mechanism and how it might be
tested. Finally.we discuss various implica-
tions of our ideas.
.
The cortex
The cortex consists of two separate sheets
of neural tissue, one on each side of the
head. The neocortex, which has a
characteristic layered structure, is found
only in mammals (see ref. 3 for recent
survey), although a somewhat analogous
structure, the wulst, is found in birds. If
allowance is made for body weight, it is
larger in primates than in most other mam-
mals and larger in man than in other
primates. It makes up a substantial fraction
of the human brain.
Different areas of the cortex perform
different functions, some being mainly
associated with vision, touch and so on,
while others appear to process more com-
plex information not associated with a
single sensory mode. The exact function of
the neocortex is unknown but it appears to
be closely associated with higher mental ac-
tivities. It seems likely that it has evolved to
perform in a rather special way.
In examining the neuroanatomy of the
neocortex one is struck by the very large
number of axon collaterals (this is not true,
for instance, of the thalamus). In any area
of the cortex the great majority of synapses
come from axons originating locally and
running within it. There is also evidence
that the majority of the synapses in the cor-
tex are excitatory in their action. This sug-
gests a capacity for self-excitatory modes
of behaviour in the cortex. And indeed, in
various conditions, such as epilepsy,
migraine and certain kinds of drug-induced
hallucination*, parts of the cortex appear
to go into large-amplitude instabilities5 .
Neuronal networks
Now, if one asks what functions such richly
interconnected assemblies of cells could
serve, one attractive possibility is that they
could store associations6-R. To see this,
suppose an ‘event’ is represented by the ac-
tivity of a subset of cells in a cell assembly.
If all the cells involved in that event form
mutual synapses, then when part of that
event is encountered again these synapses
can cause the regeneration of the activity in
the entire subset.
Much exploratory theoretical work has
been done on such networks of cells (for an
introduction see refs 6-8). In these models,
information is stored in the strengths of the
many synapses and sometimes in the firing
thresholds of cells as well. Although the ex-
act behaviour naturally depends on the
details of the particular model, certain
general properties can emerge even from
relatively simple models. The associations
which are stored are not assigned specific
locations for each item, as in a digital com-
puter. Instead the information is: (1)
Distributed: this implies that a particular
piece of information is distributed over
very many synapses. (2) Robust: this im-
plies that the information will not be totally
lost if a few synapses are added or remov-
ed. (3) Superimposed: this implies that one
synapse is involved in storing several
distinct pieces of information.
A properly designed net can be trained
(meaning that the strengths of the synapses
can be adjusted) so that given an input (a
pattern of axonal firings) it can produce the
appropriate output (another pattern of ax-
onal firings). It is found that certain
general properties will often emerge. (1)
Completion: given only part of the input
(as a clue) it can produce fairly exactly the
whole of the appropriate output (examples
aregiven in ref. 7). In computer jargon, the
memory is ‘content addressable’. (2)
Classification: given an input which is
related toseveral of its associations, it may