You are on page 1of 26

Psychological Review Copyright 1991 by the American Psychological Association, Inc.

l991,Vol-98, No. 1.96-121 0033-295X/9I/S3.00

Dreaming: Cognitive Processes During Cortical Activation


and High Afferent Thresholds

John Antrobus
City College, City University of New York

The concepts of nonlocal, or distributed, cortical and cognitive activation are examined for their
usefulness in describing the relations between sleep and waking neurocognitive processes. Changes
in the pattern of distributed activation and inhibition of selected portions of sensory, cognitive, and
motor decision modules account for the differences in imagery and thought across sleep and
waking states in comparable environments. The massive inhibition of sensory and proprioceptive
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

input to perceptual modules in Stage I REM sleep leaves the perceptual and cognitive modules, by
default, with theirown output as their sole input. Given this constraint, the activation of portions of
the cortical structures that execute waking perceptual, cognitive, and motor responses is necessary
and sufficient to produce the imagery and thought of dreaming sleep. Connectionist models are
introduced so that neurophysiological and cognitive concepts of distributed and local activation
and inhibition can be translated into a common language, and in so doing, are used to simulate
several processes fundamental to the production of imaginal thought and dreaming.

This article is predicated on the assumption that the neuro- ful for representing the process of generating dream imagery
cognitive processes of dreaming sleep can be best understood and thought, nor did they suggest a satisfactory way to span the
in the contexts of contemporary cognitive psychology and neu- formidable neuro-cognitive gap.
rophysiology, and it expresses the hope that a better compre- The theoretical approach taken here assumes that imaginal
hension of this puzzling, perceptionlike, but nonsensory cogni- and thought processes of dreaming sleep are modified forms of
tive process may, in turn, contribute to the theoretical advance- waking perceptual and cognitive processes. It rests on the as-
ment of these parent disciplines. sumption that the characteristics that dreaming and perception
The impetus for this research effort comes from the finding have in common must be produced by the activation of the
of Aserinsky and Kleitman (1953) that people are much more same processing modules.
likely to report dreamlike mentation when awakened from elec- The usefulness of the neurocognitive term activation depends
troencephalogram (EEG)-defined Sleep Stage 1 REM (rapid on the extent to which the concepts of neural and cognitive
eye movement) than when awakened from Sleep Stage 2 (often activation are, at the molar level, loosely equivalent. At the mo-
referred to as REM and non-REM [NREM] sleep, respec- lar level, neural activation is, in fact, denned by active percep-
tively). Indeed, 93.5% of REM-NREM report pairs can be tual and cognitive processing, but at the levels of modular and
correctly matched with their sleep states simply by judging micro processes, the terms neural and cognitive activation are
which member is more dreamlike (Antrobus, 1983). The re- equivalent only as working hypotheses. The assumption of
markable strength of this association suggests that it might be equivalence becomes more shaky as the state of the neurocog-
comparatively easy to identify the precise neurophysiological nitive system shifts from waking to sleep. The organization of
configuration that is responsible for the dreamlike qualities of neural processes also shifts dramatically in different sleep
sleep mentation. Published in the Journal of Experimental Psy- states, so that it is increasingly risky to identify the altered
chology by two physiologists at a time when behaviorist canons neural activity in a particular structure with the same percep-
prohibited experimental psychologists from making inferences tual or cognitive process carried out by that structure in the
about private experience, this neurocognitive claim promised a waking state.
new era in which higher cognitive processes could be translated In light of these difficulties, this article reviews the current
into neurological terms. But the promise has gone largely unre- theories and recent research on distributed activation in sleep
alized in the subsequent 38 years. Dreaming, the production of before going on to present a neurocognitive theory of sleep
visual and often bizarre, storylike, hallucinatory sequences of mentation that is based on distributed activation. Several of the
thought and imagery, was regarded as a unique experiential neural and cognitive relationships described by the theory are
phenomenon, the characteristics of which must be the responsi- then mapped onto each other by translating them into connec-
bility of a similarly singular neurophysiological process. The tionist models. This article concludes with several simulations
early models of cognitive psychology were not particularly help- of the models.

The Concept of Activation Within Sleep


Correspondence concerning this article should be addressed to John
Antrobus, Department of Psychology, City College, City University of Aserinsky and Kleitman (1953) located the subjectively ac-
New York, New York, New York 10031. tive experience of dreaming in a state where physical immobil-

96
DREAMING: COGNITIVE AND CORTICAL PROCESSES 97

ity and high sensory thresholds were paradoxically coupled produced in the cortex. Although the activation-synthesis
with a cortical EEC and EM pattern that was remarkably simi- model articulated well the neurophysiological processes that
lar to that of waking. With this discovery they inadvertently distinguish waking, REM, and NREM sleep, its account of
challenged the widely held assumption that the joint occur- cognitive processing was confined to the metaphor, that is, syn-
rence of low sensory thresholds, high motor activity, and de- thesis.
synchronized EEG necessarily defines a state of cortical activa- Mamelak and Hobson (1989) recently proposed that dream
tion. bizarreness might also be the consequence of an increment in
Zimmerman (1967,1970) skirted the challenge when he ex- errors in the output of forebrain neural networks during REM
amined the effect of cerebral "arousal" on dreaming. Activated sleep caused by the "withdrawal of the modulatory influences
sleep, or what he called "light" sleep, was denned by the individ- of norepinephrine and serotonin" (p. 201).
ual-difference variable, the auditory awakening threshold, Central to all of these theories is the assumption that the
which Zimmerman also found to be associated with elevated remarkable similarity of the asynchronous, waking EEG to the
heart and respiration rates, the core body temperature, and REM sleep EEG constitutes neurophysiological evidence that
spontaneous body movements and awakenings. Mentation re- the cortical and cognitive processes of REM sleep are more
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

ports were solicited from intervals of undisturbed REM and "active" than those of NREM sleep, where the EEG is decidedly
This document is copyrighted by the American Psychological Association or one of its allied publishers.

NREM sleep. Highly activated ("low threshold") subjects re- synchronous. Yet there is good reason not to accept this as-
ported more dreaming in Stage 2 NREM than did low activated sumption uncritically. The functional relation between cortical
subjects, but the relation was not seen within REM sleep, ar- EEG asynchrony and cognitive activation ultimately rests on
gued Zimmerman, because of a ceiling in REM sleep activa- the empirical association of asynchrony with information pro-
tion. He concluded that "dreaming is a function of cerebral cessing in the waking state. Unfortunately this association is not
arousal [read high activation] in the absence of reality contact" readily demonstrable in sleep. Johnson and Lubin (1966)
(Zimmerman, 1970, p. 547). The potential contradiction here is showed that the index of activity of some psychophysiological
that low "reality contact" implies high sensory thresholds, processes is sometimes inverted in waking and sleep. Electro-
which in turn define for Zimmerman low cerebral arousal. dermal activity, a measure of emotional responsivity in waking,
The dilemma was solved when Morrison and Pompeiano is most active during Stage 4, the least active stage of sleep. The
(1965) and Pompeiano (1970) showed that within REM sleep, traditional behavioral indices of rate of information processing
sensory thresholds and motor activity were controlled indepen- are not available in sleep to solve the dilemma because the
dently of cortical activation. They identified subcortical nuclei thresholds for various sensory and motor response systems are
that inhibited motoneurons and afferent input differently not only high and erratic but exhibit considerable indepen-
within each sensory modality. This implied that Zimmerman's dence of each other (Pompeiano, 1970; Wills & Trinder, 1978).
use of auditory arousal threshold as an index of cortical activa- The problem becomes even more complicated because the
tion during sleep, particularly within REM sleep, was not term activation, which is central for information processing and
warranted. is a critical concept in this article, has many different referents
The assumption that cortical activation translates into cogni- in neurophysiological and cognitive models. Consider the di-
tive activation next appeared in hemisphere asymmetry the- versity of its referents in models of single neurons, neural net-
ories of dreaming. Ornstein (1972), van Valen (1973), Galin works, neural pathways, and subcortical and cortical subpopu-
(1974), Broughton (1975), and Bakan (1976) proposed that the lations of neurons (see Hobson & Steriade, 1986), in cognitive
right cortical hemisphere is more active than the left during models of response latency (Anderson, 1983), and in parallel
REM sleep and that it is the primary origin of dream imagery. distributed processes (Rumelhart & McClelland, 1986). To fur-
The empirical support for this notion, recently reviewed by ther complicate matters, the criteria for defining molar neural
Antrobus (1987), is weak. Indeed, the hemisphere that is essen- activation are cognitive and behavioral, namely the efficient
tial for the experience of dreaming is the left. processing of information.
In 1977, Hobson and McCarley (1977; Hobson, 1988) pre- Activation is often understood as an index of information
sented the activation-synthesis theory of dreaming that de- processing that is accomplished by the distributed action of
scribed an activated cerebral cortex and the inhibition of affer- massive sets of neurons. Actually, the practical impediments to
ent input as requisite for dreaming. Those two assumptions are such recording have restricted most research to the simulta-
central to the theory described in this article. A major emphasis neous recording of a small number of neurons at most, gener-
of their theory, however, was that the input to the activated ally too few to infer correspondences between neural activity
cortex is eye movement (EM) information which originates in and information processing. Unfortunately, the relation of mo-
the pontine brain stem and reaches the cortex in the form of lar activation to individual neural units or modules is poorly
pontine-lateral geniculate-occipital cortex (PGO) spike bursts. understood. Information processing may actually require as
Note that EMs occur only intermittently within an REM period much inhibition of individual units and modules as activation.
and predominate in the early portion of the period. A critical Furthermore, the activation of one neuron may be defined by
corollary to their assumption is that PGO bursts provide input an increase in neural firing rate, but the activation of another
to the cortex but are not initiated by the cortical processes and neuron may be defined simply by a shift in the membrane
therefore cannot be influenced by cognitive processes. Hobson polarization unaccompanied by spiking. Neurons may be acti-
and McCarley (1977) attributed the bizarreness of dreams to vated or inhibited by interneurons with which they have no
the difficulties encountered by the cortex in "synthesizing" this synaptic contact as well as by the more familiar synaptic spike.
EM information within the context of prior cognitive processes In short, the translation from neural to cognitive and behavioral
98 JOHN ANTROBUS

activation becomes increasingly complex as one moves from and Steriade model describes only one of several mechanisms
molar to microneuronal levels. The next section reviews the of MRF-mediated, distributed, but specific cortical activation
similarities and differences in molar, nonlocal, or distributed (see Figure 1). The term distributed implies the collective activ-
cortical activation in waking and sleep, where the EEG has ity of cortical neurons over an area at least as large as a major
provided the traditional indices of activation. portion of a cortical lobe and possibly as large as the entire
cortex. The term specific indicates that the activation is re-
Distributed Cortical Activation in Waking stricted to an area much smaller than the entire cortex.
and REM Sleep Although the subcortical structures responsible for EEG de-
The frequency spectra of the cortical EEG in sleep and in synchronization in sleep were initially thought to be distinct
waking with eyes open are almost indistinguishable, and their from the ARAS of Moruzzi and Magoun (1949), it is now estab-
common profile of EEG desynchrony is supported by a similar lished (Hobson, 1965; Steriade, Ropert, Kitsikis, & Oakson,
cholinergic activation pattern in the midbrain reticular forma- 1980; Steriade, Sakai, & Jouvet, 1984) that part of the MRF
tion (MRF; see Figures 1 and 2) and by the ascending reticular includes cells that discharge at high tonic rates in both waking
activating system (ARAS) originally described by Moruzzi and and REM sleep in the absence of large motor activity. These
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Magoun (1949). cells are organized to function as a center for distributed corti-
This document is copyrighted by the American Psychological Association or one of its allied publishers.

Desynchronization refers to the diminution or suppression cal activation; they receive multiple synaptic inputs, they pro-
of synchronous, cyclic cortical EEG waves of generally elevated ject to the thalamocortical locations that are thought to control
amplitude. The synchronous waves delta (0.5-2 Hz) and theta cortical EEG, and their activity precedes cortical EEG de-
(4-6 Hz), which typify NREM sleep, and alpha (8-12 Hz), synchronization (see Figures 1 and 2).
which is common in the sleep-waking transition, are generated In a review of the brain stem control of REM sleep, Vertes
in the cortex but are under the control of subcortical structures. (1984) argued that there are strong similarities in the REM
The suppression of the synchronous waves putatively permits sleep and waking brain stem control of PGO spikes and hori-
individual cortical neurons to act independently and therefore zontal saccadic EMs. He further reviewed the evidence that the
frees them to transmit and process information (Paisley & pontine gigantocellular tegmental field (FTG) reticular forma-
Summerlee, 1984). In this sense, desynchronization is assumed tion (RF) neurons have similar discharge patterns in REM
to index both cortical and cognitive activation. Nevertheless, sleep and waking, and likewise, that FTG neurons in the me-
the precise relationship between desynchronization and neural dulla fire at high rates within both REM sleep and waking. He
information processing for each of the major synchronous fre- concluded that "each of the events of REM represents analo-
quency bands is not yet known. Although this relationship is gous or identical types of activity occurring during wakeful-
generally accepted for the waking state, it is certainly question- ness. . . . REM may not be an entirely unique state . . . [I]ts
able with respect to the sleep state. For example, during waking functions may overlap with those of waking" (Vertes, 1984, p.
an increase in EEG alpha waves is generally associated with 279). Sampling of cortical neural unit activity across waking
closure of the eyes, with drowsiness, or even with the transition and sleep states indicates that Stage 1 REM has generally high
to sleep, whereas during sleep an increment in alpha waves is rates of unit activity, though the pattern of specific cell groups
associated with incipient awakening and, by implication, an varies with state (Evarts, 1964).
increase in activation. Increasingly accurate three-dimensional measures of corti-
To understand the functional relation of cortical de- cal activity have recently been developed. Many of the new
synchronization to active cognitive processing it may be helpful techniques are indexed to neural metabolism rather than the
to identify the behavior of single cells in the shift from synchro- transient electrical activity identified by the EEG. Although
nization to desynchronization. Although this process is not some techniques require up to 45 min for reasonable resolution
clearly described for EEG alpha waves and delta waves, Hobson and most require that the subject's head be immobilized during
and Steriade (1986) proposed a model for the desynchroniza- the recording period and be free of electrodes, they have begun
tion of spindles, a prominent feature of NREM sleep in the cat. to yield some evidence concerning sleep stage differences in
They proposed that the rhythmic activity of NREM sleep spin- cortical and subcortical activation. Kanzow, Krause, and Kuh-
dles is produced by a calcium-mediated excitation which elicits nel (1962) reported an increase in cortical blood flow in REM
a potassium-mediated late hyperpolarization, followed by a sleep relative to NREM or slow-wave sleep in cats. Although
postinhibitory rebound. Because no information can be trans- their sample sizes in Stage 1 REM were small, studies of re-
mitted during hyperpolarization, all spindling neurons are es- gional cerebral blood flow in a small sample of human subjects
sentially inactivated with respect to information processing. across states also show that cerebral blood flow may be as high
Krnjevic, Pumain, and Renaud (1971) argued that acetylcho- or higher than in the waking state (Sakai, Meyer, Karacan,
line reduces neural "membrane permeability to potassium, Derman, & Yamamoto, 1979; Townsend, Prinz, & Obrist,
thus causing an increase in overall resistance and facilitating 1973). Vern, Schuette, Leheta, Juel, and Radulovacki (1987) re-
the depolarization effects of synaptic input" (Hobson & Ster- cently verified the association of cortical blood flow and cere-
iade, 1986, p. 767). This acetylcholine-mediated effect is initi- bral metabolism during waking, REM, and NREM. Ramm
ated in mesencephalic reticular formation (mRF) neurons, and Frost (1983,1986), using [14C]2-deoxyglucose as an index
which project to the intralaminar and ventromedial thalamic of regional metabolic activity, found that the reticular core, the
cells and, in turn, to the superficial part of layer one of the substantia nigra, the hippocampal dentate fascia, and the stra-
cortex. From there it extends back down to the deep, neuron- tum molecular, as well as parts of the extrapyramidal system,
rich cortical layers. This simplified summary of the Hobson showed increased activity during REM sleep in rats. Franck,
DREAMING: COGNITIVE AND CORTICAL PROCESSES 99

CEREBRAL CORTEX
Motor Conceptual Perceptual
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

RET1CULAR FORMATION
MIDBRAIN

EXCITATORY

Figure 1. Neural system in the waking state. (Perc. = perceptual; Sens. = sensory; FTG = gigantocellular
tegmental field; FTC = central tegmental field; TRC = tegmental reticular nucleus; RN = raphe nucleus;
PBL = parabrachial zone; LC = locus coeruleus.)

Salmon, Poirrier, Sadzot, Franco, and Maquet (1987) recently ity is too long to permit a close match to the time interval from
reported a positron emission tomography (PET) study of 2 hu- which the thought and imagery of the awakened sleeper are
man subjects. Cerebral metabolism rates in REM sleep were reported. PET also exposes subjects to nuclear radiation. These
about 5% higher than during waking, whereas NREM sleep handicaps notwithstanding, the overall picture is that the corti-
values were 30% lower than waking rates. cal structures that support associative cognitive processing are
Although PET, nuclear magnetic resonance (NMR), and similarly activated in waking and in REM sleep, but markedly
neuromagnetic field techniques are ideally suited to measure reduced in NREM sleep.
local activation of cortical and subcortical activation, they are
difficult to coordinate with the apparatus necessary (i.e., EEG The Subcortical Control of States
electrodes) to identify sleep stages. Furthermore, the recording The activation or desynchronization of the cortical EEG is
time necessary to achieve accurate resolution of cerebral activ- initiated by the discharge of rostrally projecting MRF neurons
100 JOHN ANTROBUS
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

BRACHIUM
CONJUNCITVUM

RETICULAR FORMATION
TONTINE MIDBRAIN
OCULOMOTOR
SYSTEM CHOUNERG1C/
BRAIN STEM
SKELETAL SYSTEM
MUSCLES FTG,FTCTRC
(Er.cept eyes
and respiration)

PONS
off
NORADRENERQC
BRAIN STEM
SYSTEM
RN,PBL,LC

BULBAR/MEDULLA

LEGEND
INHIBITORY
EXOTATORY

Figure 2. Neural system in Stage 1 REM sleep. (Perc. = perceptual; Sens. = sensory; FTG = gigantocellular
tegmental field; FTC = central tegmental field; TRC = tegmental reticular nucleus; RN = raphe nucleus;
PBL = parabrachial zone; LC = locus coeruleus.)

(see Figure 2), which project to the medial thalamus, then to vating system has led Hobson and his colleagues to identify it as
pyramidal and nonpyramidal cortical interneurons, and fi- a cholinergic/cholinoceptive system, although acetylcholine is
nally to the excitatory and inhibitory cortical neurons that pro- not itself present throughout the system. Hobson and McCarley
cess cognitive information (Steriade, Kitsikis, & Oakson, 1978; (1977) and Hobson, Lydic, and Baghdoyan (1986) have argued
Steriade et al., 1984). The prominent role of the neurotransmit- that the control of MRF activation within REM sleep is distrib-
ter acetylcholine and cholinergic agonists throughout this acti- uted among a set of cholinergic-rich nuclei in the pontine and
DREAMING: COGNITIVE AND CORTICAL PROCESSES 101

medulla, or bulbar RF (see Figure 2). In order of decreasing acterizes the NREM EEG. But inhibiting the cholinergic/cho-
concentration, these neurons lie in the FTG, the central teg- linoceptive control center deprives the control center of its own
mental field, and the tegmental reticular nucleus. In contrast to source of activation so that it eventually loses its ability to sus-
the tonic activity of MRF neurons in waking and REM sleep, tain its inhibitory influence. Freed from inhibition, the mRF-
the pontine RF cholinergic neurons "discharge phasically, with based center again activates the thalamus and cortex to produce
highest rates during oculomotor events in desynchronized sleep Stage 1 REM sleep.
but also during some waking movements" (Hobson et al., 1986, The great paradox of REM sleep is that the cortex appears to
p. 743). The contrast is most dramatic (1:40-50) when the wak- be as active as that of a waking person, yet the person is rela-
ing control states are obtained from a restrained animal. Dur- tively unresponsive to external stimuli. The solution to this puz-
ing REM sleep, the highest rate of FTG discharge is during zle is that the cholinergic activation pattern of REM sleep in-
EMs and PGO activity. cludes at least two subsystems not activated in waking, and
The critical issue for a model of sleep state control and for the these subsystems inhibit both sensory input and motor execu-
control of REM imagery and thought is whether the FTG plays tion during REM sleep. Jones and Cuello (1989) recently
a major role in initiating REM sleep and the phasic events of showed that one branch of cholinergic pathways activates sites
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

REM sleep, or whether it merely follows motor activity. Hobson that inhibit the execution of motor commands (Morrison &
This document is copyrighted by the American Psychological Association or one of its allied publishers.

and Steriade (1986) argued that Pompeiano, 1965) and is possibly responsible for the wide-
spread sensory inhibition characteristic of REM sleep. The sub-
the pontine brain stem is necessary for desynchronized sleep. . . system controls the "powerful postsynaptic inhibition of the
[and that]. . . within the brain stem, cells in the pontine reticular brain stem and final-common-path motoneurons" (Hobson et
formation have long phase leads over both the tonic and phasic
phenomena of REM sleep. Pontine reticular cells begin to in-
al., 1986, p. 379), which tonically restrain all skeletal movement
crease discharge rate dramatically during synchronized sleep sev- except eye movements and respiration (Pompeiano, 1967). This
eral minutes before the occurrence of desynchronized-sleep epi- inhibition, mediated by discharges of the MRF and anteriodor-
sodes and long before any discharge-rate changes are seen in other sal pontine tegmentum (Hobson et al., 1986), has the effect in
cell groups. This period of neuronal activation onset occurs with-
REM sleep of canceling the execution of all motor commands
out muscle atonia and without movement. The question thus be-
comes, if the pontine reticular cells are simply follower cells, what initiated by the motor cortex (cf. Figures 1 and 2). The neuro-
are they following, because there is no movement and also no cognitive implications of MRF-cholinergic/cholinoceptive ac-
sensory feedback from movement, (pp. 779-780) tivation in sleep, then, are the activation of those cortical areas
concerned with cognitive processes extending from the later
Opposing the cholinergic/cholinoceptive REM network, stages of perceptual analysis (Perceptual^) to motor com-
Hobson and McCarley (1977) and Hobson et al. (1986) have mands and the inhibition of all sensory input to this system,
proposed for NREM sleep a distributed noradrenergic network including inhibition of proprioceptive feedback from motor
centered in norepinephrine (NE)- and serotonin-rich neurons commands.
concentrated in the dorsal Raphe nucleus, peribrachial region, The relation between sensory and central activation in
and locus coeruleus (LC; see Figure 3). The two states of sleep, NREM sleep appears to be quite the reverse of the REM pat-
REM and NREM, which alternate in 90-min cycles throughout tern. The cortex is less active and sensory thresholds are higher
sleep, are controlled by two sets of brain stem generators. In than in all other sleep states. To explain this paradox one needs
brief, their model describes two intermingled control areas, to examine the behavioral correlates of the NE network in the
each with cells distributed across several nuclei. At sleep onset, waking state. Bloom (1979), Aston-Jones and Bloom (1981),
the noradrenergic network, apparently controlled by the LC Foote, Bloom, and Aston-Jones (1983), and Tucker and Wil-
and DRN, inhibits the second center, concentrated in the mRF. liamson (1984) have proposed that the NE network, which is
The mRF controls the cholinergic/cholinoceptive network that largely controlled by the LC, plays a major role in the early
activates several thalamic nuclei, including the medial and in- stages of selective perception in the waking state. Bloom (1979)
tralaminar thalamic nuclei (Hobson et al., 1986) that in turn demonstrated that NE increases the sensitivity of cortical neu-
desynchronize the cerebral cortex, activating excitatory and in- rons to sensory input and that LC projections phasically aug-
hibitory cortical neurons in both waking and sleep. ment d' (i.e., the signal-to-noise ratio) in sensory-perceptual
Although this effect has generally been regarded as distrib- analysis(Bloom, 1980). Footeetal. (1983) showed that NEselec-
uted, Hobson and Steriade (1986) cautioned that the distinctive- tively augments cells' evoked response to sensory input and
ness of the cortical projections of different thalamic nuclei inhibits the spontaneous background activity of surrounding
imply that the thalamocortical pathways are "a widespread but neurons. Foote (1986) suggested that the release of NE in the
highly specific network" (p. 725). Indeed, "at the level of extra- waking state "facilitates vigilance processes, that is, maximal
cellular recording, there is no visible difference in the mecha- attention to external stimuli and minimal control of various
nisms of cortical activation observed during REM sleep and brain regions by autoactivation, recurrent networks, recruit-
waking" (Hobson et al., 1986, p. 379). Microinjections of the ment, or 'intrinsic' processes" (p. 405). Tucker and Williamson
MRF with cholinergic agonists desynchronize the cortical (1984) made a strong case for a noradrenergic, LC-controlled
EEG (Baghdoyan, Rodrigo-Angulo, McCarley, & Hobson, network, populated with rapidly habituating neurons that bias
1984). By contrast, when the noradrenergic network is domi- the sensitivity of the organism in favor of stimulus change or
nant and the cholinergic/cholinoceptive network is inhibited, novelty.
the medial thalamic nuclei are underactivated and the cerebral In summary, this binary sleep model consists of alternating
cortex displays the slow, high-voltage, rhythmic EEG that char- REM-NREM, cholinoceptive/noradrenergic activation pat-
102 JOHN ANTROBUS

CEREBRAL CORTEX
Motor Conceptual Perceptual
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

RETICUtAR FORMATION
MIDBRAIN

INHIBITORY
EXOTATORY

Figure 3. Neural system in non-REM (NREM) sleep. (FTG = gigantocellular tegmental field; FTC =
central tegmental field; TRC = tegmental reticular nucleus; RN = raphe nucleus; PEL = parabrachial
zone; LC = locus coeruleus.)

terns, the first of which is associated with the wakinglike de- lus selection and precategorical processes. Although the subjec-
synchronized cortical EEG, which earns it the label paradoxical tive or conscious characteristics of these processes may be negli-
sleep. This REM sleep state is characterized by a pattern of gible, they may nonetheless use a substantial portion of cortical
cognitive activation that corresponds to waking processes that activity. The next section attempts to map this neurophysiologi-
extend from Perceptual^ through to motor commands. Al- cal picture onto some of the characteristics of spontaneous
though the precise starting point of Perceptual^ is difficult to nonperceptual imagery and thought.
define, it excludes the retina and optic nerve, the middle ear,
vestibular system, and somatic afferents (Pompeiano, 1970). In Activation, Imagery, and Thought Across
the visual system it may start as early as the LGN or the fourth
layer of the occipital cortex.
Sleep Stages and Waking
The pontine-to-cortical activation pattern of NREM sleep The finding that initially suggested that general cortical and
seems to be confined to Perceptually processes such as stimu- general cognitive activation might be fundamental to the pro-
DREAMING: COGNITIVE AND CORTICAL PROCESSES 103

duction of Stage 1 REM mentation was the REM-NREM re- morning phase of the diurnal rhythm, are additive. The dream-
port-pair study mentioned earlier (Antrobus, 1983). Although like quality of REM reports are intensified to a degree rarely
the global judgment, as measured by the Dreaming Scale, was seen in laboratory reports. Even NREM reports become longer
the best discriminator of the REM-NREM pairs, the length of and the imagery becomes more vivid. The argument that both
the verbal report, as measured by the Total Recall Count (TRC) the REM and the morning rhythm increments in TRC are acti-
Scale, was only 1% behind, correctly discriminating 92.5% of vation effects is strengthened by the fact that both cortical
the 73 pairs, F = 196. But when TRC was partialed out of states are driven by the MRF (Figures 1 and 2).
dreaming, the residual dreaming variance was trivial. In other The NREM increment found by Kondo et al. (1989) echoes
words, the extent to which dreaming discriminated between Zimmerman's (1970) finding that, relative to normal sleepers,
Stage 1 REM and Stage 2 REM was almost totally a function of the NREM reports of light sleepers (i.e, people with low sen-
the total information produced by the sleeper, and TRC is there- sory thresholds during sleep) were more dreamlike. By using a
fore a plausible index of cognitive activation. measure of activation, namely the rising phase of the diurnal
When TRC was partialed out of the Visual Word Count cycle, that is free of the confound of cortical activation and
Scales (nouns, action verbs, modifiers, and spatial preposi- sensory and motor inhibition, however, Kondo et al. were able
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

tions), the REM-NREM visual imagery difference completely to demonstrate an incremental effect of activation on menta-
This document is copyrighted by the American Psychological Association or one of its allied publishers.

disappeared. In a recent analysis of the same report set (Rein- tion in both REM and NREM sleep.
sel, Antrobus, & Wollman, in press), the bizarreness distinction Despite the broad research support for the position that cor-
between REM and NREM reports was similarly erased by par- tical desynchronization indexes cognitive activation, the evi-
tialing out TRC. These results suggest that dreamlike menta- dence is not unanimous. Wollman and Antrobus (1987) failed
tion might be produced throughout the night, but only during to support it when they examined within-state correlations, in
REM sleep is the association cortex sufficiently active to sup- both REM and waking, between TRC and EEC spectral power.
port a lengthy mentation report. Subjects were interrupted three or more times within each state
The pattern of relations between wakingand sleep mentation for a mentation report. EEG power was computed in each of
is altogether congruent with the neurophysiological picture. the standard EEG bands for windows extending for 1 to 5 min
Antrobus, Reinsel, and Wollman (1984) found that REM re- prior to the interruption. After normalizing the power values
ports are surprisingly similar in most respects to mentation within each subject, the values were correlated with TRC values
sampled from the waking state under identical laboratory con- for the interruptions. None were significant, p > .05, N = 30.
ditions. If mentation from three states is judged for its dream- To summarize, dreaming in its strongest form is character-
like quality, waking and REM are virtually indistinguishable, ized by visual imagery that is bright, clearly focused, colored,
and both are much more dreamlike than Stage 2. Waking re- and moving; by a storylike conceptual theme that is almost
ports tend to be somewhat longer than REM reports, but the always perceived as a "real" event (i.e., it is hallucinatory); and
latter tend to be more hallucinatory and bizarre (Reinsel et al, by its sometimes bizarre quality. The magnitude of these quali-
in press). Judges seem to weigh length and storylike quality ties varies over time within sleep as a function of the activated
against the hallucinatory and bizarre qualities to judge waking phases of two biological rhythms, the 90-min REM-NREM
and REM reports equally dreamlike. sleep cycle and the 24-hr diurnal sleep wake rhythm, which
The major cognitive distinction between waking and REM reaches its upper limit when REM sleep is sustained into the
reports is that waking reports have more frequent shifts in the late morning hours when the waking diurnal rhythm of the RF
Topic Scale count (Reinsel, Wollman, & Antrobus, 1986), provides additional activation to the cerebral cortex without a
whereas the primary distinction between REM and Stage 2 corresponding increment in the inhibitory processes of REM
mentation reports is that the former are much longer, that is, sleep.
have higher TRC counts per report. Waking reports are distin- If this increase in cortical activation were accompanied by an
guished from Stage 2 reports both by higher TRC and more equivalent increase in all thequalitiesof waking perception and
topics. thought, the bizarreness of dreaming, to the extent that it repre-
If the traditional global variable, dreaming, had alone been sents a failure of the production process, should completely
used in the studies of Reinsel et al. (in press) and Wollman and disappear. But surprisingly, bizarreness increases with increas-
Antrobus (1987), only one of the state differences would have ing activation, both within REM sleep and waking, under zero
been identified, and it would have been misinterpreted as a illumination conditions. It is proposed here that bizarre events
difference in a complex class of fantasy production, namely, illustrate the partial independence of a low-level visual produc-
dreaming, rather than simply a difference in the amount of tion module from top-down conceptual influence, and the in-
information generated and stored. To identify the possible termittent inability of the conceptual model to provide a plausi-
neurocognitive links in the domain of human sleep mentation, ble interpretation of these out-of-context visual productions.
one must, therefore, relinquish the generic term "dreaming" in
favor of a set of more precisely defined cognitive measures such
The Production of Conceptual Thought and Imagery
as total information reported, brightness of visual imagery, and
so forth. Visual imagery is produced by the activation of some part of
The late morning dream effect recently reported by Kondo the visual perception system when it is dark-adapted and not
(1988; Kondo, Antrobus, & Fein, 1989) shows that TRC in both processing visual sensory stimuli. These conditions are satis-
REM and NREM late morning reports obtained 3 hr later than fied not only in REM sleep, but also in waking, in a light-proof
on control nights is significantly elevated. The two sources of environment. Whether the perceptual and cognitive modules
activation, the REM-NREM increment and that of the rising that produce visual imagery and conceptual thought are
104 JOHN ANTROBUS

equally well activated by increasing cortical activation during The inhibition of motor activation in REM sleep has often
REM is difficult to determine on the basis of existing research. been proposed as the basis for the dreamer's experience of be-
At low levels of activation, denned by TRC, subjects appear to ing unable to move even when the intention to run is the appro-
report images that they identify by name only at the time of priate response to an imaged villain. The motor cortex is able
making the report. There seems to be no recognized context, to issue the command to run, but its execution is inhibited. The
no story line, and no interpretation while asleep. At interme- sensation of being immobilized is created by the absence of the
diate levels of activation subjects seem to be able to recognize or appropriate afferent feedback from the skeletal muscles that
identify their imagery while asleep, but only with difficulty. would normally execute the response.
The names of objects are often prefaced with "sort of like" or At high levels of late morning REM activation the inhibition
"kind of a." This difficulty in categorization suggests that of skeletal muscles and motor feedback may be partly overcome
image production is not necessarily initiated or controlled by a and the dreamer may become more aware of the remaining
higher level conceptual or semantic process. At higher levels of inhibition. Indeed, the dreamer may even manage the spastic
activation, images are identified but sometimes seen as poor or kick of the motor run action and escape from dreaming to
even bizarre exemplars of the person or object they portray, and wakefulness. The clustering of "lucid" dreaming, the experi-
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

sometimes seem out of place in the thematic context of the ence of simultaneously being awake while dreaming (LaBerge,
This document is copyrighted by the American Psychological Association or one of its allied publishers.

dream sequence. In short, the amount of information in the 1985) in the last hour of sleep, supports this conception. Antro-
visual image implies the participation of a top-down process in bus, Antrobus, and Fisher (1965) and LaBerge (1985) have been
its construction. Yet the lack of effective quality control over the able to train some subjects to make a manual signal when they
form of the image suggests that the lower level processes exer- are asleep and dreaming, and the majority of such responses are
cise a considerable degree of independence, implying that the produced in the final hour of REM sleep.
process of image production described in waking studies of Even at a high level of activation, image production by lower
instructed imagery (Kosslyn, 1988) may differ considerably level visual modules is only partly under the control of top-
from that of REM sleep. down cognitive processes, so that waking visual images often
One possible explanation for the shift in the strength of visual fail to match the subjective expectation or subvocal commands
processes relative to conceptual processes with increasing acti- of the imager (Perky, 1910; Segal, 1971). Moreover, the ability of
vation within REM sleep is that conceptual structures may be the waking imager or daydreamer to distinguish his or her own
handicapped by the REM state inhibition of interaction with images from perceptual patterns is a function of the strength of
the sensorimotor components of the speech apparatus. Con- the sensory stimulus. It is not surprising, therefore, that the
ceptual and linguistic processes are intricately related to vocal REM dreamer, deprived of external visual stimuli because of
and acoustic processes. For example, Klapp, Marshburn, and the elevated REM thresholds, has no option but to interpret the
Lester (1983) demonstrated the operation of both acoustic and image as a veridical perception and respond appropriately. Pre-
articulatory loops in working memory Although these loops sented with visual information that is out of context, concep-
have not been examined in sleep, all sensory and proprioceptive tual processes apparently act simultaneously to revise the
afferents that have been studied have shown marked attenua- meaning context to make it more consistent with the image and
tion in REM sleep (Pompeiano, 1970). Furthermore, Orem to modify or reinterpret the image to minimize the mismatch
(1977) has shown that laryngeal muscle tonus in cats is greatly with the conceptual context. Each imposes constraints on the
reduced relative to wakefulness and NREM sleep. Unfortu- other.
nately, the patterns of motor inhibition in REM are somewhat The process of interpreting information in the context of
species specific, and there are no comparable data on humans other information is reminiscent of the synthesis process which
or on a time-of-night effect. Hobson and McCarley (1977) introduced to explain how EM
Nevertheless, there is good reason to suspect that both activ- information transmitted to the cortex by PGO waves is incorpo-
ity and feedback from speech production centers are inhibited rated into the ongoing dream. This cognitive process of finding
in REM sleep and that this inhibition impairs the ability of a congenial solution for a pattern of information, some seg-
conceptual and linguistic modules to generate abstract thought ments of which may be in conflict with each other, may be a
in REM sleep. Only at high levels of cortical activation, namely primary attribute of all perception and thought, and it is a
when the rising phase of the diurnal cortical activation rhythm central focus of the connectionist models that are described
summates with the REM phase of the sleep rhythm, is this later. But in contrast to a synthesis process applied exclusively
inhibition likely to be partly overcome. to PGO inputs, I show later that imagery and conceptual
There are several interesting dream phenomena that may be thought are continuously produced by some explicit form of
controlled by this inhibition of afferent input to the cortex. synthesis or constraint satisfaction (Rumelhart, Smolensky,
Subjects often report that some of their REM experience is McClelland, & Hinton, 1986) among all of the activated partici-
much like passively watching television, whereas they actively pating modules.
participate in the action of other dreams, and in still others they The thematic coherence of most dream reports suggests that
watch themselves participate. Imaged physical participation in the production process, including visual image production, is
the dream may depend on feedback from larger skeletal muscu- under the control of a top-down process. Indeed, this assump-
lature. Conceptual-verbal dreams may be dependent on feed- tion is the basis for the widespread practice of dream interpre-
back from subcortical speech centers. The passive watching of tation. 1 would like to suggest that the contribution of top-
images may well identify a default condition when propriocep- down processes (i.e., the output of a conceptual module) has
tive feedback is strongly inhibited. been overemphasized relative to the influence of bottom-up
DREAMING: COGNITIVE AND CORTICAL PROCESSES 105

processes, particularly the output of the visual module. (The process to the interpretation of external stimuli, but that this
term module will refer hereinafter to a set of interconnected interpretation is the conceptual stuff of the dream. Even when
elementary units dedicated to a particular class of information no external stimuli are imposed, the reported dream event sug-
processing. Modules may pass information to one another, but gests that the cognitive module has a challenging job in provid-
the connections between elementary units of adjacent models ing plausible interpretations for the images produced by the
are comparatively sparse relative to the dense interconnections visual module. For example, a dreamer reports,". . .small per-
within modules. Although module is used as a cognitive term, sons, evidently children in years but with aged and deeply lined
the phrase cortical module is also used to refer to a specific f a c e s . . . . It dawned on me that this. . . was a house of correc-
cortical location where the functional property of that area is tion; the aged children were unfortunates who had grown upon
well established.) the streets.. . ." (Hobson, 1988, p. 274). I am suggesting that
There is good reason to assume that the conceptual and vi- the visual module produced small people and lined faces but
sual modules jointly accommodate each other in order to syn- that the conceptual module was obliged to produce a scenario
thesize the multiple, overlapping schemata into a single coher- or schema that would jointly satisfy the visual constraints, the
ent schema. The impression that the conceptual module con- prior conceptual context of the dream, namely a large public
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

trols this synthesis may be an illusion fostered by the building, and the world knowledge available from the sleeper's
This document is copyrighted by the American Psychological Association or one of its allied publishers.

remarkable facility of the conceptual module to construct in- memory.


terpretations and event scenarios that minimize the discrepan- Although these examples illustrate the facility of the sleeper's
cies between the two modules. conceptual interpretation, they in no way diminish the likeli-
To the extent that this synthesis is an interactive process, and hood that a similar interpretive process is carried out at lower
no single metric for scaling the two processes exists, it is not levels, such as early perceptual but nonsensory, kinaesthetic,
possible at this time to estimate their relative contribution to and visual processing, which are inaccessible to introspection.
the construction of spontaneous imagery. Nevertheless, if one Such early processes might include the construction of visual
assumes that during waking visual perception the conceptual features, such as a lined surface, that are in turn interpreted into
module is biased to accommodate the visual module which, in objects or persons. In short, this interpretive process may be a
turn, is biased to respond to sensory input, and if one further fundamental characteristic of the perceptual cognitive system.
assumes that visual imagery is produced jointly by both mod- In the final section of this article I show how this process might
ules when there is no retinal input to the visual module, then be accomplished by translating the concepts of synthesis and
one may infer that the bias of the conceptual module carries interpretation into the more precise operations of constraint
over to the imagery condition when there is no sensory input. satisfaction.
To the extent that the visual component of mentation is
stronger in REM sleep than in any other state, one may also Bizarre and Anomalous Imagery and Thought
infer that the spontaneous activity of the visual module, or
in REM Sleep
some portion thereof, makes a strong independent contribu-
tion to the dream imagery that the two modules jointly produce Although bizarreness is popularly regarded as one of the
in REM sleep. most salient characteristics of dreaming, it is not particularly
Although this interpretive process may be inferred post hoc prominent in laboratory REM reports except when they are
from REM dream reports, it is demonstrated more convinc- obtained in the late morning hours (Rondo et al, 1989), when
ingly by the presentation of external stimuli during sleep. Be- the cortex reaches its highest level of REM sleep activation.
cause the response to visual stimuli is strongly inhibited, even Furthermore, the fact that the major subscales of the Bizarre-
when the eyelids are taped open and the pupils are dilated, ness Scale, such as Improbable Combinations of visual features,
other sensory modalities must be addressed. After administer- occur with equal frequency in REM (mean = 0.34/report) and
ing a fine spray of cold water on subjects who were in REM waking (0.36/report), under matched conditions of zero lumi-
sleep, Dement and Wolpert (1958) obtained reports such as nance in a sound-attenuated chamber (WakingMC), implies that
a high level of cortical activation is a necessary condition for the
I was walking behind the leading lady, when she suddenly col-
production of bizarre mentation. The same is true for the Dis-
lapsed and water was dripping on her. I ran over to her and felt
water dripping on my back and head. The roof was leaking.... I continuities subscale (low-probability sequences within the
looked up and there was a hole in the roof. I dragged her over to dream). Reported discontinuities are usually prefaced with,
the s i d e . . . . 1 woke up. (p. 550) ". . . and all of a sudden." Although fairly common in REM
sleep (0.48/report), they are twice as frequent in quiet WakingMC
Another subject reported (0.93/report).
. . . children came into the room and came over to me asking for
To the extent that bizarreness is a form of cognitive error and
water. I had a glass of ice water and I tipped the glass to give it to is associated with heightened cortical activation in both REM
them. I was sitting, and I spilled the water on myself. The children sleep and WakingMC, one may ask what condition is necessary
wanted the ice and tried to grab it, but it slipped away. I got mad for its occurrence. The absence of external visual stimuli seems
because they were so greedy and tried to shove them away. . . .
to be the essential requirement. Waking visual perception is a
Then I got out of the chair and was going to change my pants,
(p. 550)
conceptual interpretation of the output of lower level visual
processing modules for form, movement, depth, color, and so
I suggest that not only is the interpretation of the self-ini- forth, whose coherence with one another is imposed in part by
tiated productions of the visual module in REM sleep a similar the temporal and spatial regularities of the external world. Al-
106 JOHN ANTROBUS

though conceptual modules may exercise some control on the assumed, in the waking state, to constrain an individual's own
lower level visual and spatial modules, the control is apparently speech and other motor behavior.
limited. In the absence of sensory input, the visual module or To simplify this discussion, I have described the production
modules are apparently able, when sufficiently activated, to of sleep imagery and thought as Stage 1 REM phenomena. It is
produce images on their own. But these modules are special- more accurate, but still an oversimplification, to say that
ized for spatial rather than temporal analysis. Therefore, in the NREM mentation is produced by the same process but at a
absence of external sensory constraints, they tend to produce much attenuated magnitude. In the next section I compare the
quality visual images but also produce temporal sequences that cognitive and neural differences between the two sleep states in
arc sometimes implausible. more detail.

NREM Versus REM Mentation


Imaginal Sequences
Almost all NREM mentation reports are obtained from
Although the production of sequences and the perceptual Stage 2 because Stage 2 makes up the majority of NREM sleep
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

expectation of sequential events is not well understood, one and is evenly distributed throughout the night. NREM sleep is
This document is copyrighted by the American Psychological Association or one of its allied publishers.

may assume that the information used to produce imagined defined by a relatively high-voltage, slow-frequency, synchro-
sequences includes both parallel distributed information as nized cortical EEC, which lacks the subcortical PRF activation
well as sequential information. One may further assume that in source that activates the association and motor cortex in REM
the absence of sensory input, the input to the system is its own sleep and waking (cf. Figure 3 with Figures 1 and 2; Hobson &
prior output. That is. its state at time ( + 1 is a function of its McCarley, 1977; Hobson & Steriade, 1986). Dreaming is re-
state at time t. Of particular interest here are the ways of model- ported in N R E M sleep, but the probability that the informa-
ing distributed sequential processes both within and between tion in an NREM report, as inferred from the report length,
visual, motor, and conceptual modules. A further concern is to equals or exceeds the information in a REM report, matched by
avoid the assumption that image sequences are produced only subject and time of night, equals only 0.075 (Antrobus, 1983).
by the conscious components of imagination. It seems likely As previously described, NREM dreaming is enhanced among
that the imaginal properties of which one is aware are a small people whose sleep is generally disturbed (Zimmerman, 1970)
but undetermined fraction of the total information that con- and in normal sleepers during the very late morning hours
trols the production of the successive images in an image se- (Rondo et al, 1989). In both conditions the NREM pattern of
quence. activation is no doubt supplemented by the RF-thalamic pat-
The ability of one imaged event j to cue a second image k tern characteristic of REM sleep and waking.
undoubtedly depends on the amount of information carried by The paradox of N R E M sleep is that while it is in one sense
the imaged event and on the strength of the learned connec- the least activated sleep state, it has the lowest sensory thresh-
tions between events j and k. If the image of j carries as much olds, and a subject in N R E M sleep, upon being awakened,
information as the corresponding perceived event j, then the offers the most reports of being awake and thinking. In sum-
probability, p(k), of a particular successive image should be the mary, the cortical EEG and mentation reports indicate a state
same whether event j is an image or a percept and should be oflow activation, but the low sensory thresholds and reported
constrained only by the strength of the learned connections to waking status imply just the opposite.
image k. If, in the course of waking perception, the expectation Sensory thresholds during sleep shift so rapidly following pre-
of event k is not confirmed by external sensory input, the per- sentation of a stimulus, with the modality and personal signifi-
cept at time t + 1 is corrected. But in an imaginal sequence, no cance of the stimulus and with the neural and behavioral crite-
sensory data are available to correct the expectation, so that the ria for the response, that its complex literature cannot be sum-
image at t 4 1 becomes the cue for the image at t + 2, and the marized in this article. Note simply that Stage 2 N R E M is
image at t + 2 becomes the cue for the image at t + 3, and so notorious for its high rate of "I-was-awake" responses, which
forth. often exceed the 50% rate during the first half of the night
Johnson, Foley, Suengas, and Raye (1988) recently showed (Antrobus & Saul, 1980; Bonnet & Moore, 1982; Sewitch, 1984).
that the memory for an imagined event contains less sensory The cues that subjects use for this judgment are unknown, but
information than the memory of a perception. One may de- mentation reports and sensory and perceptual thresholds sug-
duce, therefore, that relative to a waking percept, a dream gest that subjects in Stage 2 often have a vague recognition of
image will impose weaker constraints on the production of a either auditory (Wills & Trinder, 1978) or somatic (Antrobus,
subsequent image. The cumulative effect of this decreased con- 1986) stimuli.
straint over successive images should be that the imagined se- Williams, Hammack, Daly, Dement, and Lubin (1964) and
quences wander or progressively deviate from the path of a Williams, Morlock, and Morlock (1966) found that subjects
real-world sequence that starts from the same point. exhibited lower sensory thresholds in NREM than in REM
In addition to the information attenuation that may be char- sleep, particularly for personally nonsignificant stimuli. Antro-
acteristic of waking images, REM sleep images are impover- bus and Saul (1980) found that subjects reported being awake
ished by an additional factor which has been previously de- on 50% of NREM interruptions, compared to nearly 0% in
scribed. The inhibition at the spinal level of motor commands, REM, a finding that was subsequently confirmed by Sewitch
including speech, and the additional inhibition of kinaesthetic (1984) and Rosekind and Schwartz (1988a, 1988b). Antrobus
feedback effectively el iminate a major source of feedback that is (1983) found that 30% of NREM awakenings included some
DREAMING; COGNITIVE AND CORTICAL PROCESSES 107

sense that the subject's status lay somewhere between waking than good exemplars. Often they are "seen" as novel and im-
and sleep ("I was not really asleep") compared with 2% in REM probable combinations of familiar visual features. Although
sleep. These NREM reports were very brief and unelaborated these attributes are characteristic of REM mentation, and in a
relative to REM reports and waking reports (Wollman & An- diminished form are characteristic of NREM mentation, they
trobus, 1986). can also be produced in the waking state under the identical
However, these early perceptual responses do not seem to conditions of reduced sensory pattern stimulation.
initiate the elaborate conceptual sequences of either the REM A second set of characteristics tends to be exclusive to REM
dream or waking thought. Thus, although the high rate of sleep mentation and may be attributed to the very high and
"awake" responses implies that Stage 2 NREM is the most acti- prolonged afferent and motor inhibition during that stage. The
vated sleep stage, the relative cognitive impoverishment of the inhibition of external and proprioceptive input creates a cogni-
reports and the low rate of recalling anything suggest that the tive situation where the imagined percepts carry the same
conceptual module is comparatively inactive in that state, ex- weight in subsequent cognitive processes as do the perception
cept as already noted among light or activated sleepers, or of external stimuli in the waking state. It is this belief in the
within late morning NREM periods. reality of imagined events that earns the REM experience the
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

The paradox may be resolved if one assumes that the Percep- label hallucinatory.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

tual Early stages of information processing are activated in This belief in turn constrains the course of REM mentation.
NREM sleep, but the output of these stages is not processed Unfamiliar and improbable forms that might be automatically
further by the association and motor modules. Thus the sleeper tagged as fanciful or unknown if seen in the movies must be
may marginally encode kinaesthetic or auditory stimuli and explained when they are "seen" in the dreamer's kitchen talking
conclude that he or she must be not entirely asleep, but can to one's mother. These interpretations in turn modify the subse-
describe little additional ideation. quent sequence of the dream. If, for example, the form is inter-
This hypothesis is supported by several similarities in the preted as unknown and therefore a threat, the dreamer may
neurocognhive patterns of waking perceptual attention and decide to make a run for it.
NREM sleep. As described earlier, the LC stands at the distri- The decision in the motor cortex to execute a motor act,
bution center of a widespread aminergic-rich subcortical and coupled with high motor and proprioceptive inhibition, creates
cortical network (see Figure 3) that, in waking, may modulate the second unique characteristic of REM mentation. The mo-
perceptual attention (see Bloom, 1980; Foote et al., 1983). In tor decision is not followed by the proprioceptive feedback
waking perception, the adrenergic Perceptually and choliner- from the skeletal musculature that accompanies the corre-
gic Perceptual^-association-motor processing phases must sponding waking act, and this mismatch is noticed by the
interact cooperatively. In sleep, however, the phases are func- dreamer and interpreted as the inability to move. This interpre-
tionally disengaged with respect to sensory information pro- tation likewise modifies the subsequent course of the menta-
cessing as the switching comes under subcortical control. In tion.
NREM sleep, the cholinergic system is down, so that little fur- The integration of the imaged sensory, perceptual, cognitive,
ther processing of sensory input takes place. The reduced sen- motivational, and motor information into a unified experience,
sory thresholds of NREM sleep are, therefore, a remnant of the a characteristic of waking perception, is also accomplished in
adrenergic-controlled early Perceptual attention processing sleep but with a difference. In waking perception, the reliability
state of waking perception. of the relationships among perceptual features is constrained
by the sensory patterns in the external world. In sleep, however,
these environmental constraints are absent. The constraints on
Summary of Empirically Supported Relationships and
the possible relationships among features in a scene and on the
Theoretical Conjectures
succession of events in time are dependent on prior perceptual
The imagery and thought of sleep occur periodically when, learning. Yet the self-produced images of REM sleep elicit the
in Stage 1 REM, subcortical processes activate in a distributed same sense of surprise to the dreamer as do external stimuli to
manner a large portion of the cortical processes that, in the the waking observer. The images are based on perceptual ex-
waking state, compute perceptual, cognitive, and motor re- pectations, yet they are experienced as entirely unexpected.
sponses to external stimuli. Sleep imagery and thought, there- This surprise or even bizarreness, which is expressed in the
fore, share many of the characteristics of waking responses to temporal discontinuities and improbable spatial combinations
sensory stimuli. These responses include the creation of percep- of features, represents the joint inability of these perceptual
tual features that are predominantly visual but also include the expectations to adequately constrain the production of percep-
auditory and haptic modalities. Also included are most of the tual features and interfeature relationships. It also represents
cognitive responses of identification or recognition, the inter- the inability of the interpretive portion of the cognitive system
pretation of relationships among the perceptual features, the to come up with a satisfactory account for the unexpected
interpretation of the meaning of objects and event sequences, image relationships.
problem solving, and the cognitive and premotor processes of As the distributed activation of the cortex increases, one
motor response. might expect that the activation of learned constraints would
The visual forms approach the brightness, clarity, and mobil- reduce the frequency of improbable events in the dream. In-
ity of waking percepts. Although the forms are typically identi- stead, paradoxically, the likelihood of such unexpected or bi-
fied or recognized by the sleeper, they are judged frequently by zarre relationships increases with increased cortical activation.
the sleeper as "something like" known objects or persons rather Increased activation increases both the amount of information
108 JOHN ANTROBUS

generated in the dream and also the sophistication of the The Single-Frame Schemata Model
dreamer's interpretations. Apparently the enhanced activation
of the interpretive processes does not keep pace with the activa- A brief examination of the Rumelhart, Smolensky, McClel-
tion of the image production processes. land, and Hinton (1986) schemata model, which I will label
Enough is currently known about sleep imagery and thought SCHEMATA:RSMH, will introduce the reader to connection-
to acknowledge that the activation, or the active inhibition, of ist models and to the merits and limitations of a single-layered
every part of the interactive waking sensory-perceptual-cogni- model that computes schemata for a single time frame. A
tive-motivational-motor system is implicated in its produc- schema is a unit of information created when smaller informa-
tion. Sleep imagery and thought are produced by a periodically tion units interact to form a functional pattern. Several of these
altered form of the waking system. To the extent that these schemata may in turn combine to form a larger schema, and so
sleep alterations and their effects on imagery and thought are on until a single higher order schema is formed. This superor-
well understood, it is the lack of knowledge of the waking sys- dinate schema may represent the coherent unity of a conscious
tem that limits the understanding of sleep imagery and thought experience; the subordinate schemata may represent names
production. In the absence of such knowledge, it may be worth- and objects that are components of that experience. The sub-
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

while to use the current knowledge of both sleep and waking schemata may represent features of the names and objects, and
This document is copyrighted by the American Psychological Association or one of its allied publishers.

processes to model selected characteristics of the system, and the microschemata may represent microfeatures that combine
to utilize the behavior of the model to guide the selection of in various configurations to form a particular feature. They are
future research questions. cognitively impenetrable and therefore not accessible to con-
sciousness. The smallest units of information may be small
micronetworks of neurons, but it is unnecessary to extend this
Connectionist Models and Sleep Mentation model to that level.
A powerful property of this conception of schemata derives
The connectionist models recently introduced to neuropsy- from several assumptions: (a) that the basic units of the schema
chology (Rumelhart & McClelland, 1986) provide several pow- are subfeatural or microcognitive, (b) that these units are distrib-
erful ways to model some of the integrative processes of percep- uted, (c) that the processes by which they interact occur in paral-
tion and cognition and may also be able to simulate the cogni- lel, and (d) that these processes are incremental. Within the
tive integrative processes, as well as the integrative failures, of constraints of cortical architecture and the learning history of
REM sleep. the model, each schema, subschema, and microschema can
The conception of dreaming as a perceptionlike process car- accom modate the context of the other participating units. They
ries with it the assumption that perception originates in the jointly determine not only which schemata will become active,
sense organs and that, with some top-down assistance from but most remarkably, the particular configuration each schema
conceptual and motivational loci, the dominant flow of infor- will take. For example, the "horse" schema not only activates
mation is bottom-up. It is now well established, however, that schemata for four legs, but the four-"leg" schemata are config-
production of dreaming requires no input from the sense or- ured differently from the "leg" schemata for "dogs" and
gans, which are the starting point for waking perception. The "chairs."
perceptual system, when fully activated in early morning REM This accommodation is realized by replacing the feature list
sleep, can do a remarkably good job of simulating the stream of of the traditional list-processing artificial intelligence model
waking perception in real time with no sensory input. This with a distributed set of microfeature units so that different
suggests that a perceptual model that emphasizes a serial pro- variations on a given feature can be constructed from different
cessing order from lower to higher level modules may not be activations of the participating microfeatures. The incremental
entirely inappropriate for dreaming. A better model for dream- manner of building a schema allows each microfeature to grow
ing may be one which integrates the information both within or shrink depending on the current context of the other active
and between modality-specific, motivational, conceptual, and microfeatures. And just as microfeatures combine to form fea-
motor modules in parallel. That is, the information in each tures, so too features combine to form the schemata of objects
module constrains the processes of the others. and persons, and they in turn combine to form the novel sche-
This integrative capability of connectionist models is specifi- mata of the superordinate schema.
cally employed to simulate the following characteristics of Rumelhart, Smolensky, McClelland, and Hinton (1986) il-
dreaming: (a) the holistic, schemalikequality of the dream expe- lustrated the process of creating the schema of an entire room
rience; (b) the coherent flow of perceptionlike imagery in the from units that consist of objects that in turn consist of features.
absence of sensory input; (c) local "interpretive" accommoda- They constructed a network of 40 items or units found in kit-
tion to the inhibition of feedback from motor decisions; (d) the chens, bedrooms, bathrooms, and offices. For each descriptor
discontinuities, novelty, and bizarreness of imagery and they estimated its bias to occur or to be active regardless of the
thought; and (e) the amplifier effects of subcortical distributed activity of its neighbors, and its probability of beingactive in the
activation on local processing. presence of each of its neighbors. Using a Bayesian analysis,
Inasmuch as Rumelhart, Smolensky, McClelland, and Hin- they derived a matrix of weights, wt>, that represented the rela-
ton (1986) have demonstrated that a connectionist model based tion between each unit i and each other unit j in the matrix. To
on Hopfield's (1982) annealing model has some of the proper- satisfy Hopfield's (1982) thermodynamic model of constraint
ties of schemata, this seems like a useful starting point for a satisfaction, they adopted the restrictive assumption that of
model of the formation of sleep imagery and thought. symmetric weights, w^ = ws,. It was assumed that the weights
DREAMING: COGNITIVE AND CORTICAL PROCESSES 109

represented an individual's associative knowledge of the rela- be represented by small changes in the pattern of activation of
tions among the 40 items. the microunits that contribute to the image. This ability to con-
The process of constructing a schema is initiated by turning struct novel forms suggests that connectionist models may be
on or activating one or two units in the net. This activation may well suited to represent the novel, odd, or bizarre combinations
represent sensory input to the network or input from a network that characterize many dream experiences.
in a sensory "pathway" When one or two units are so activated, Although the microcognitive unit may be theoretically re-
the activation spreads to other units to which they are con- ducible to individual neurons, note that Smolensky (1988)
nected by positive weights, and inhibition will spread to other showed that the process of constraint satisfaction can be ap-
units to which they are connected by negative weights. As acti- plied to cognitive units that are as large as features or objects so
vation spreads through the network in successive cycles, the long as the microcognitive units of the model are smaller than
more activated units successively acquire more influence on the the schema or outcome that they collectively construct.
developing pattern of network activation. The less activated There are, however, some limitations on the applicability of
units, by reason of inhibition or receipt ofless activation from SCHEMATA:RSMH to dreaming. Based on the Hopfield
surrounding units, lose influence. Although the pattern may be (1982) model, SCHEMATA:RSMH is handicapped by its as-
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

unstable at first, some subgroup of units eventually reaches a sumption of symmetric weights and, as will be discussed later,
This document is copyrighted by the American Psychological Association or one of its allied publishers.

state in which it dominates the network and inhibits other po- by its inability to represent nonlinear relationships among their
tential subgroups that might have become active if the initial units. SCHEMATA:RSMH starts with the external activation
activation pattern had been different. Rumelhart, Smolensky. of several units and then iterates through several cycles until it
McClelland, and Hinton (1986, p. 13) defined this state as a settles on a solution, at which point it stops unless activated with
kind of goodness-of-fit (G) or constraint satisfaction after Hop- new input. Moreover, once a solution is reached, it takes a very
field (1982) where G(t) reaches an upper limit. large input pattern to disturb the network enough for it to seek a
new solution. The constraints achieved with the initial solution
simply inhibit any local disruption by new input.
where ( = a point in time, i and j = any pair of units in the net,
By contrast, the imagery and thought of a person change
and a = the activation level of a unit or descriptor.
continuously in a sequential fashion and appear to represent
The contribution of a pair of units is given by the product of their previously learned spatial and temporal sequences. If one visu-
activation values times the weights connecting them. Thus, if the
alizes A, a door opening, one expects and therefore may imag-
weight is positive, each unit wants to be as active as possible — that
is, the activation values for these two units should be pushed to- ine B, someone coming into the room. If one sees this person
ward 1 . If the weight is negative, then at least one of the units coming closer, one expects the person to speak C. To model this
should be 0 to maximize the pairwrie_goodness. (Rumelhart, sequence, A must activate B, and B must activate C; in addition,
Smolensky, McClelland, & Hinton, i986,~vbT;2rjr-14J~' B must inhibit A, and C must inhibit B. Clearly sequential
Note that, although the input or sensory term permits the events cannot be represented by symmetric weights. Natural
network to respond to external information, the model will hierarchies in the environment also defy representation by sym-
continue to build even when the input value returns to 0. This metric weights. For example, chairs are present in nearly all
feature permits one to simulate an "imaginal" process by seed- kitchens, but kitchens do not occur in the presence of all chairs.
ing the network with external input and then running it for It is assumed that schemata that represent the characteristics
several cycles without additional external input. of dream mentation must include some of the properties of
As the model iterates through successive cycles, the activa- imagery such as visual form, brightness, and color that are cus-
tion of related pairs of units extends to larger clusters of units, tomarily associated with the concept of consciousness. It is also
and these larger clusters exert increasing influence on other assumed, however, that many of the units that contribute to the
clusters of units to which their units are related by positive or construction of schemata must represent spatial, temporal, and
negative weights. abstract properties that do not have sensory or perceptual prop-
erties. Support for this assumption comes from a recent study
It is these coalitions of tightly interconnected units that corre-
by Intons-Peterson and Roskos-Ewoldsen (1989), who showed
spond most closely to what have been called schemata. The stable
pattern as a whole can be considered as a particular configuration that visual imagery sequences can be under the control of non-
of a number of such overlapping patterns and is determined by the visual information such as the inferred weight of an object to be
dynamic equilibrium of all of these subpatterns interacting with carried. Thus, the imaged act of catching a ball depends not
one another and with the inputs. Thus, the maxima in the good- only on the visual image of the trajectory of the ball but also on
ness-of-fit space corresponds. . . , in the language of schemata,
other factors, such as one's imagined role in the process either as
[to] configurations of instantiated schemata. In short, they are
those states that maximize the particular set of constraints acting an observer or as a participating player, on one's body position,
at the moment. (Rumelhart, Smolensky, McClelland, & Hinton, on motor decisions, and on the proprioceptive feedback from
1986, vol. 2, pp. 20-21) motor behavior, which is greatly attenuated due to the inhibi-
The distributed characteristic of connectionist models al- tion of motor activity and afferent transmission in REM sleep.
lows a small number of microcognitive units to combine to
Simulations
form a large number of possible schemata, just as the small set
of letters or sounds of a language are joined to form an almost Distributed Recurrent Activation Model of Imagery and
infinite number of words. Because new and novel images and Thought (DREAM1TS)
concepts may be produced without increasing the total number To simulate imaginal sequences it is first necessary to con-
of microunits, an infinite number of variations of an image can struct a model that can generate schemata in the absence of the
110 JOHN ANTROBUS

All

UNIT i
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

EXCITATORY
INPUT
SUB UNIT

Input
site

=21= From All From All


Excitatory Inhibitory
Units Units

Distributed to all Afferent Output Subunits

Distributed to

PONTINE RP-to-
THALAHD8,
CORTICAL
INHIBITORY
CONTROL
ONIT
ON only
in REM
Sleep

Figure 4. Organization of DREAMITS units. (RF = reticular formation.)

concurrent "sensory" or "perceptual" input that was simulated abstract concepts such as values and roles. As with SCHE-
in the SCHEMATA:RSMH model with its set of units clamped MATArRSMH each unit has a weight connecting it to each unit
on throughout the construction of the schemata. DREAMIT:S in the matrix. I assigned the initial set of weights as the esti-
avoids the need to clamp on the input units by instead using a mated conditional probability of unit j being active, given the
decay function so that each unit loses only a portion of its acti- activation of unit /. Unlike the Hopfield (1982) model, weights
vation in each cycle, exclusive of any activation it receives from were generally asymmetric even when associated with nonse-
other associated units in the network (see the Appendix and quential relations. Because the weights sending activation to
Figure 4). Units that represent sensory information decay rap- unit J, unit k,.. . unit N were assigned independently of one
idly (40% per cycle). Units that represent contextual informa- another, they represented at best only an approximation of the
tion decay slowly (20% per cycle), and of the remaining units, ideal weights where the activation of unit j is a function of the
the majority decay at an intermediate rate (30% per cycle). joint positive and negative input from all units in the net.
DREAMITrSl is a single-layered network of 90 cognitive A simulation by DREAMITSl is initiated by activating one
units designed to simulate several one- and two-step sequences or two related units at 0.9 (90% of their maximum activation),
that might be engaged in by a male college student. The major- and observing the shifting patterns of activation as the network
ity of the units represent features, objects, persons, places, and moves through successive cycles. The initial activation spreads
DREAMING: COGNITIVE AND CORTICAL PROCESSES 111

across associated units until coherent schematalike patterns of motor decision unit tends to activate the "m: I-throw-ball" mo-
activation are created, as with SCHEMATA:RSMH. For exam- tor feedback unit as well as the subrole unit representing "other-
ple, when the "physics" unit is primed, the network moves over ballplayer recipient." But as the pattern approaches comple-
three or four iterations to a state in which "I-as-student" "re- tion, it already begins to activate units that will form the next
member" to do my physics "homework." The network creates schema, and they begin to inhibit the preceding pattern.
an integrated set of schemata in that it fills in the context of For the schema to represent a visual image, some of these
being at "college," the attitude that physics is "difficult," the units must also activate the lower level visual modality mod-
goal to "do well" and the intention to "try hard." If it runs a few ules. Thus activating "mother" tends to activate some of her
cycles longer, "I-as-student" will "remember" to "buy" a new facial features such as her "v: smile," and activating "ball" tends
"book" to do the homework for the course. "Physics" and the to activate "v: small-white-round."
"I-as-student" role unit inhibit other incompatible roles and the The simulations with DREAMIT:S2 require that a distinc-
persons and events associated with them. Similarly, if the "I-as- tion be made between perceptual and motor responses that are
son" role unit is active, then priming "hungry" will activate closely tied to the initial stimulus input and imaginal responses
"mother," her "cooking" and "smiles," and the "kitchen" and that are remotely related to the input. If the perceptual and
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

kitchen "table." However, priming "hungry" when the "I-as- motor responses cannot be accurately simulated, then the re-
This document is copyrighted by the American Psychological Association or one of its allied publishers.

student" role is active will activate the school "cafeteria," "ham- sponsibility for the failure of the model to produce imaginal
burgers," and lunch "money." response sequences cannot be clearly identified. Accordingly,
A successful schema was denned as the activation of at least the game of catch was divided into eight functional steps. One-
0.85 of three or more associated units and no incompatible and two-step sequences were defined as perceptual or motor;
units. Simulations starting with 30 alternate input pairs demon- longer sequences were called imaginal. A total of 20 different
strated that DREAMITiSl can produce schematalike patterns two-unit input combinations (e.g., "ball-coming" and "I-as-ball-
of activation without clamping on the input units. Although player") were used to test the ability of the model to produce
DREAMITrSI is able to compute schemata responses that are good perceptual-motor schemata through two steps, starting at
one or two steps removed from the input that represents the each of the eight steps in the game.
external stimulus, like SCHEM ATA:RSMH, it always comes to When inappropriate units were activated or appropriate
a stop at that point. A successful model of imaginal processing, units omitted from a schema, some of the weights were modi-
however, must be able to produce a continuous sequence of fied and the perceptual simulation was rerun until an appro-
schemata. Because the restriction of symmetric weights is re- priate schema response was produced. After each response was
laxed in DREAMIT:S1, there is no reason for the model to successfully produced, the prior sequences were retested and
come to a halt provided there is sufficient sequential informa- the weights readjusted where necessary until the 20 different
tion in its weight matrix. To this end I designed DREAMIT:S2 input combinations yielded appropriate responses.
to represent the sequences in a game of catch. The somewhat arbitrary nature of this weight adjustment is a
DREAMIT:S2 differs from DREAMIT:S1 in that it consists tacit acknowledgment that the Hopfield (1982) model is not
of only 62 units and the type of units includes both sensory and altogether suitable for modeling perceptual events. The prob-
motor information. To simulate the sensory threshold and mo- lem clearly lies in the inability of single-layered networks to
tor inhibition shifts that are characteristic of REM sleep, 7 units express nonlinear relations among the units. The conditional
represent low-level visual attributes, such as sizes and colors, probability of unit j becoming active given that unit (' is active is
and 7 units represent feedback from motor decision units, such often dependent on the activation of other units. SCHE-
as "motonl-running," which are represented independently of MATA:RSMH and DREAMIT:S cannot represent such inter-
the motor decisions such as "I-run." The network includes 5 active relations. For example, in a network in which "white,"
distracter units (e.g., "rectilinear," "paper," "for-writing-on") "small," "round," and "for-throwing" activate "ball" and
that are irrelevant to the game of catch but useful for evaluating "white," "small," "rectilinear," and "for-writing-on" activate
the model. "paper," the activation of either "ball" or "paper" activates both
Depending on the status of other units in the network, the "white" and "small," and these two features then feed activation
activation of the "ball-coming-close" unit tends to activate "I- back to both "ball" and "paper." This pattern of weights has the
run" and "I-catch-ball," and both "I-catch-ball" and "I-pick-up- unfortunate consequence that the activation of "ball" in the
ball" activate the "I-throw-ball" motor decision unit, which to- play-catch sequence tends to start a do-your-homework se-
gether activates the motor feedback unit "m: I-throw-ball," quence, or vice versa, and that in turn causes the sequence of
which in turn activates the visual (image) unit, "ball-going." the play-catch schemata to disintegrate. The problem could be
Unlike a schema simulation based on a Hopfield net, the solved by using separate "white" units for "paper" and "ball,"
weights in the sequential simulation are set so that clusters of but this would greatly increase the size of the memory required
units that represent a particular step in a sequence, once fully and violate the basic distributed quality of the network.
active, will then inhibit the units representing the prior step. Nevertheless, further evaluation of DREAMIT:S was carried
Essentially, this progression is similar to any nondistributed, out with the assumption that as long as representation was dis-
a -» b -» c, sequence. The distributed model, however, permits tributed across many units and was able to successfully simulate
many units in the network to participate in each step of a se- short stimulus-response sequences it could provide useful in-
quence and allows some units even to anticipate and prepare formation about imaginal sequences.
the context within which other units in the step subsequently In summary, DREAMIT:S simulations of schematalike se-
become active. For example, activation of the "I-throw-ball" quences were successful but only within certain limits. As de-
112 JOHN ANTROBUS

scribed later, DREAMIT:S1 and DREAMIT:S2 simulated the After several hundred learning trials on each of the 28 pairs,
NREM decrement in dreaming, and DREAMITS2 simulated the learned weights may be tested by presenting each pattern of
the REM sleep inhibition of sensory input, including feedback the input set and comparing the output with the desired output,
from motor units. Above all, DREAMIT:S2 was able to pro- namely the pattern in the training set. For example, the input
duce continuous sequences well beyond the brief perceptual 0011110011 (Table 1, row I) is being trained to produce
sequences that it had been specifically taught. In particular, the
0.000 0.000 1.000 1.000 0.000 0.000 1.000 1.000 0.000 0.000
play-catch sequences could run from 5 to 10 steps, which
amounted to 2 or 3 throws, before the schematalike patterns but may produce
showed the signs of incoherence that are described later. The
simulations demonstrated that a single-layered distributed net- .006 .009 .968 .957 .067 .059 .859 .868 .067 .051,
work can produce a sequence of schematalike patterns, but the
so that the corresponding errors are
shortcomings of the model argue for its replacement by a mul-
tilayered model, DREAMIT:BP .006 .009 .032 .043 .067 .059 .141 .132 .067 .051.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Note that the input and teaching values are binary but the out-
This document is copyrighted by the American Psychological Association or one of its allied publishers.

Distributed Recurrent Activation Model of Imagery and


put values are continuous. Unless specifically noted, training is
Thought: Backpropagation (DREAMITBP)
then continued until each input produces a good output re-
DREAMITBP is designed to overcome the limitations of sponse, denned here as error < 0.1 for each output unit. Note
DREAMITlS, but it is restricted to a smaller number of units that the number of input units equals the number of output
and does not presume to represent real-world episodes such as units.
the student behaviors represented by DREAMITS. It is con- BP is used here to model, in a closed system, the effects of
structed to demonstrate how imaginal sequences may be repre- learned nonlinear contextual relations in temporal sequences.
sented as the activation, in a closed system, of perceptual or For this reason, the model is less concerned with how the rela-
perceptual-motor sequences previously learned in the waking tions are learned than that they are learned. Because it assumes
state. In particular, it shows that nonlinear models are neces- that learning is based on feedback from the environment, that
sary to represent these sequences. is, "supervised," BP is the model of choice. BP has many limita-
A layered network such as BP (Rumelhart, Hinton, & Wil- tions as a general model of learning, but they do not handicap
liams, 1986) can allow the activation of units such as "small" this particular application. BP models are often captured pre-
and "white" to interact with different combinations of the fea- maturely by local minima, that is, solutions that are less than
tures "spherical," "rectilinear," "for-throwing," and "for-writ- optimal and where the above learning criteria cannot be met.
ing-on" to create a unique representation of "ball" and "paper." The recently developed BP/Cauchy algorithm (Wasserman,
This nonlinear representation is accomplished by the interposi- 1989) promises to solve this problem. In thisstudy such training
tion of a single "hidden" layer between an input and output trials were simply aborted and rerun, except where specifically
layer. Each unit of the input layer is connected to each unit in noted.
the hidden layer by a different weight, and each unit in the The primary goal of DREAMIT:BP is to determine whether
hidden layer is connected to each unit in the output layer by perceptual or perceptual-motor sequences that are learned in
another weight, making a total of two weight matrices (see Fig- an open (waking) state can be generated in a closed (sleep) state.
ure 5). An additional important result of the simulations that are de-
The units in the input layer are binary digits 1 or 0 that com- scribed later is serendipitous. The temporal discontinuities and
bine to form a stimulus pattern such as 0011110011. The acti- improbable combinations, the bizarre relations of dreaming
vation from each unit is propagated forward to each unit in the sleep, are shown to be derivative of contingencies learned in the
hidden layer where it is multiplied by the weight connecting the waking state. In particular, well-learned sequences and contin-
two units. This operation is repeated for all input units that gencies can be represented by a closed system with minimum
connect to a given hidden unit, summed over all inputs, and distortion or discontinuity. Imaginal discontinuities and im-
repeated in parallel for all hidden units. The sequence is then probable combinations of features occur where the sequences
repeated between the units of the hidden and output layers to or contingencies have not been well learned in the waking state.
create the activation value for each unit in the output layer. To represent an event sequence in which each digit provides a
The second advantage of a BP model is that it learns its own context for the meaning of the others, let a 5-digit string repre-
weights when given correct feedback in response to its output. sent the following information: the first digit represents "kit-
The value of each output unit is compared to the correct or chen" (1) or "classroom" (0); the second digit represents "door
"teaching" value (see Table 1, column 2), and the "error" or opens" (1) or "not open" (0); the third digit represents "a per-
difference is computed. The error information is propagated son" (1) or "none" (0); the fourth digit represents "speaks to
back (Figure 4, dashed lines) to the output layer, where the you" (1) or "doesn't" (0); and the fifth digit represents "a com-
weights connecting the hidden layer units to that layer are ad- mand" (1) or "none" (0). The sequence of the classroom door
justed up or down according to the generalized delta rule (see opening, the teacher entering (01100) and telling one to pay
Rumelhart, Hinton, & Williams, 1986; Wasserman, 1989, attention (00111) can be represented by 01100 -» 00111. The
chap. 3). Next, error information is propagated back to the sequence of one's mother coming into the kitchen (11100) and
hidden layer, and appropriate adjustments are made to the first telling one to wash one's hands (10111) can be represented by
weight matrix (see Carpenter, 1989). 11100 -» Will. DREAMIT:BP shows how such contextually
DREAMING: COGNITIVE AND CORTICAL PROCESSES 113

OUTPUT TEACHING
LAYER a
:..: 'Or - Or...; -Q——Q- INPUT FROM
ANOTHER
MODULE

ERROR
COMPUTA-
TION: 1
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

LAYER It
INPUT

INPUT
FROM ^
LOWER -( )-
PERCEPTUAL^-'
MODULE
--O--O-

Figure 5. Illustration of standard backpropagation module.

organized learned-sequence pairs behave when, as in the sleep sory input permits the imaginal output to become the new "per-
state, they are cut off from external input. However, no particu- ceptual" input, the output of DREAM1T:BP at time t becomes
lar representations are intended for the digit strings in the fol- the input at time / + 1. Thus, in the imaginal mode, DREAM-
lowing simulations. IT:BP is a recurrent model in which the output of the forward
As with DREAM1T:S2, the simulation of imaginal processes propagation phase is folded back to the input layer (see Figure
must be preceded by a demonstration of perceptual processing, 6). This is equivalent to eliminating the output layer by return-
which in turn must be preceded by the learning of appropriate ing the output of the hidden layer directly to the input layer to
weights. The training set for DREAMIT:BP5-L (number of seg- create a two-layered model. This input-output layer is similar
ments = 5, L = loop, number of units =10, number of hidden to the single-layered schemata model of DREAM1T:S, where
units = 9) consists of 28 input-teaching event pairs arranged in each unit is connected directly to each other by a single weight,
four sets (see Table 1). The 28 out of 32 possible (2s) pairs are except that each intemnit connection is mediated by two sets of
arranged so that the teaching event of 1 pair equals the input to weights, namely the weights that link unit i to each unit in the
the next, thereby forming a single loop. Each event consists of hidden layer and the set of weights that link each hidden unit to
five segments, and each segment is a 00 or 11 pair. The combina- unit j in the original layer. In Figure 6, this arrangement is
tion of the two leftmost segments are unique to each of the four input -» hidden units, and hidden -* output = input units.
sets of 7 pairs, and the three rightmost segments change from Similar recurrent models have recently been proposed for learn-
one event to the next. In this sense, the left segments provide ing serially ordered events (Jordan, 1986; Park, 1989).
stable contexts for the right segments. Simulations of perceptual and imaginal sequences. Each BP-
Following the assumption that in sleep the absence of sen- learned solution for a specific input-teaching set is unique be-
114 JOHN ANTROBUS

Table 1
Training Set for DREAMITBP5-L and DREAMIT:BP5-4S

DREAMIT:BP5-L DREAMIT:BP5-4S

Subset Input Teaching Input Teaching

1 1 0011110011 0011001100 001 1 110011 0011001100


2 0011001100 0011001111 001 1001 100 001 1001 1 1 1
3 0011001111 0011110000 0011001111 0011110000
4 0011110000 0011111111 0011110000 001 1 1 1 1 1 1 1
5 0011111111 0011111100 0011111111 0011111100
6 0011111100 0011000000 0011111100 0011000000
7 0011000000 1100000011 0011000000 0011000011
2 8 1100000011 1100001100 1100000011 1100001100
9 1100001100 1100111111 1100001100 1I001II11I
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

10 11001 1 1 1 1 1 1100000000 1100111111 1 100000000


This document is copyrighted by the American Psychological Association or one of its allied publishers.

11 1100000000 1100110011 1100000000 1100110011


12 11001 10011 1100110000 1100110011 1100110000
13 1100110000 1100111100 11001 10000 11001 1 1 100
14 11001 11 100 minim 1100111100 11001 inn
3 15 i n n urn i i inn 100 llllllllll 1 1 1 1 1 moo
16 1 1 1 1 1 1 1 100 1111110011 m i l l MOO in moon
17 1111110011 1 1 1 1 1 10000 1111110011 1 1 1 1 110000
18 1 1 1 1 1 10000 III100I1I1 1111110000 1 1 1 1001 1 1 1
19 1 1 1 100111 1 11 11001100 1111001111 1111001100
20 I I 11001 100 1111000000 imooiioo 1 1 1 1000000
21 1111000000 000000001 1 1111000000 1111000011

4 22 000000001 1 0000110000 0000000011 0000110000


23 0000110000 0000001111 0000110000 0000001 1 1 1
24 0000001 1 1 1 0000111100 00000011 11 ooooi moo
25 00001 1 1 100 0000001100 0000111100 0000001100
26 0000001100 0000111111 0000001100 0000111111
27 00001 1 1 III 0000000000 0000111111 0000000000
28 0000000000 0011110011 0000000000 0000110011

cause it starts with a different set of randomly distributed closely approximated members of the training set and they oc-
weights. In order to avoid making generalizations from a curred in the order prescribed in the training set. Between 1 and
unique learned solution, a minimum of three independent solu- 3 (mean = 1.4) of the 28 possible imaginal steps, however, were
tions were completed for each of the DREAM 1T:BP models "soft" in that 1 or more of the units fell in the 0.101-0.300 or
described here. 0.700-0.899 range. Some of these soft spots were followed by
DREAM1T-KP-5-L. There is no evidence that human imagi- one or two degraded or noisy responses (some units in the
nation becomes progressively more chaotic in a closed state 0.301-0.699 range). Without exception, however, all of these
such as sleep or sensory deprivation. A major failure of were followed by recovery with good output.
DREAMIT:S was its tendency to become chaotic in runs of On successive imaginal simulations, the soft spots tended to
more than 10 steps. Therefore, the first test of DREAMIT:BP-5- recur at the same points in the loop, which were invariably at
L was that it run indefinitely without showing signs of either transitions from one higher order context to the next. The per-
progressive disorganization or chaos. DREAMIT:BP-5-L was ceptual tests of these steps suggested that the soft spots in the
designed to produce a continuous 28-step loop in the imaginal imaginal sequences were caused by inferior learning of the
mode, that is, with no external input other than an initial transition steps between the four subsets. The probability of a
stimulus. change in the leftmost segment was 0.14 (4 out of 28 sequences
Note that although the "imaginal" test with the continuous in the training set), compared with 0.41 for the next two seg-
loop did not begin until DREAMIT:BP-5-L could produce ments and 0.82 for the rightmost segment.
good output in response to each of the 28 input items, the input To test this learning interpretation, the proportion of transi-
for each imaginal step was the noisy output of the previous tion sequence pairs between subsets in the training set was in-
response (see Figure 6), not the clean binary code of the percep- creased to 0.25 by repeating each transition three additional
tual tests. If this noise was passed along to the output at each times in succession. The soft spots completely disappeared.
step, the cumulative effect would be either chaos or random These points of sequence disruption, followed by resumption
output. at a new point, unexpectedly simulated several characteristics
In the imaginal mode, where the output is folded back to the of the bizarre temporal discontinuities that characterize hu-
input, DREAMIT:BP-5-L produced a continuous sequence of man thought and imagery in REM or WakingMc. The simula-
output indefinitely. Most of the output events in a 28-step se- tion suggests the hypothesis that discontinuities are likely to
quence were good (0.000-0.100 or 0.900-1.000) in that they occur at the ends of sequences that have been well learned in
DREAMING: COGNITIVE AND CORTICAL PROCESSES 115

OUTPUT a
INPUT TO
NEXT
PROCESSING
MODULE

LAYER 2:
HIDDEN

LAYER 1:
a. INPUT
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

b. BUFFER
Equivalent to the
Layer 3 of Figure 4

luccetitv* output
LEyEL
of tlM buffer
PERCEPTUAL
MODULE

Figured. Illustration of the forward-propagation portion of DREAMIT:BP, a backpropagation module for


generating either perceptual or imaginal temporal sequences. (For the sake of simplicity, the back-
propagation of errors is not shown.)

the waking state, or conversely, that continuity is sustained 5-4S simulations suggest how prior learning can determine the
where the components of the sequence have been well learned processes of recovery from a terminated sequence. In particu-
in the waking state. lar, they suggest that contextual features and other well-learned
DREAMITBP-5-4S To further test this simulation, components of the prior sequence strongly influence the point
DREAMIT:BP-5-4S (4S = four subsets) was trained on 28 in- of discontinuity at which the sequence will resume. Finally, the
put-training pairs arranged so that the last teaching event in process by which DREAMITBP recovers from degraded out-
each subset was not the input event for the next pair (Table 1). In put suggests useful hypotheses about the fundamental basis of
the imaginal mode, when primed with 0011110011, it propa- imagery and thought production.
gated forward to a close approximation of the first training The DREAMIT:BP simulations support the hypotheses that
response and then continued through the next six response the continuity of an imaginal sequence is a function of the
events to a close approximation of 0011000011, the end of the strength of the prior learning of sequential constraints and that
first set. It then produced two or three "soft" responses and the onset of a discontinuity is a consequence of too little con-
some output that was noisy. But rather than degrading further, straint. The DREAMITBP-5-4S simulations showed that the
the model recovered within one to three cycles (mean = 2.6) to transition from sequence continuity to degraded output always
produce a "good" response from among the 28 possible re- occurs at the point where, in the learning history of the model,
sponses in the teaching set. Indeed, most of the responses were the sequence segment terminates. The DREAM 1T:BP-L simula-
from the teaching events of the same subset. Similar behavior tions showed that the same effect tended to occur within a
occurred when DREAMIT:BP-5-4S was primed with any event learned sequence at any point where the step to the next point
from the other three subsets. was poorly learned. This empirically testable hypothesis is
Bizarreness in DREAMIT:BP-5: Discontinuities in sequence. unique in that it is suggested by a model of how sequences are
The performance of DREAMIT:BP-5-4S confirms the seren- learned and then played out in a closed system, rather than by
dipitous simulation by DREAM1TBP-L of several characteris- assumptions about unique cognitive processes during REM
tics of the discontinuities that occur in REM and WakingMC sleep.
imagery and thought. Discontinuities occur when the learned Not only do the sequential productions of DREAMIT:BP
constraints that control the production of the next item in a resume, but they tend to recover and resume at a contextually
sequence are insufficiently strong. Second, the DREAMIT:BP- appropriate point in the teaching sequence. This serendipitous
116 JOHN ANTROBUS

result has not been studied in the laboratory, but it is a relation- blocked or impaired feedback can be represented in DREAM-
ship that can be researched. It says that when there is a disconti- IT:BP by observing the effect of degrading one segment in the
nuity in the imaginal sequence, the sequence will resume at a input. The simulation, described earlier in this section, shows
point that is related to the prior sequence rather than at any that inhibited kinaesthetic feedback can produce temporal dis-
random point. For example, DREAMIT:BP-5-L learned the se- continuities.
quences 0011111111 -»• 0011111100 and 0011111100 -»• Bizarreness in DREAMIT:BP: Improbable combinations.
0011000000 (Table 1, rows 5-6). At one point in the imaginal Two potential classes of improbable combinations of
simulation it produced the noisy output .05, .07, .92, .95, .60, DREAMIT:BP output can be defined: improbable within-seg-
.56, .74, .69, .93, and .95. Despite the noisy third and fourth ment and between-segment combinations. Improbable combi-
segments, the greater similarity of the remaining good seg- nations of unit pairs (segments) never occurred in any imaginal
ments to the input values of Step 5 than of Step 6 leads to an simulation. Typically the output values for unit pairs were
output that is much closer to the output of Step 5,0011111100, within 0.05 of each other. To the extent that these digit pairs
than to the output of Step 6,0011000000. This is a straightfor- represent the internal consistency of features that are invariant
ward example of stimulus generalization. A soft input event in the learning environment, this is equivalent to saying that
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

elicits a response that is appropriate to the most similar stimu- bizarre combinations never occur among features that are per-
This document is copyrighted by the American Psychological Association or one of its allied publishers.

lus of the learning set. Note the distinction: It does not produce ceptually invariant. For example, faces are never imaged with
a best approximation to the input value of 5; rather its output their noses inverted.
pattern is the appropriate response or sequel to the stimulus Unlike the within-segment combinations, between-segment
pattern that its previous output most closely resembles. combinations were highly varied. However, not one novel com-
Even the most degraded responses possessed some good seg- bination of the five segments was observed in all the DREAM-
ments, and these segments constrained the alternatives for the IT:BP-5 imaginal simulations just mentioned. The lack of novel
subsequent response. In this sense, each segment of DREAM- between-segment combinations was due, in part, to the fact
IT:BP-5-L behaved as a context for the others. This context ef- that 28 of the 32 possible segment combinations (0.875) were
fect was tested by degrading one of the three rightmost digit used as responses in the training set, so that if the segments
pairs. Degrading was accomplished by substituting 01 o r l O f o r were combined at random, the probability of a novel combina-
11 or 00 in an input item (note that the simulator will not accept tion was only 0.125.
0.5 or 0.5). As with the earlier simulations that started with one In the real world, however, only an infinitesimal proportion
or two spontaneously degraded inputs, it took one to three cy- of all possible feature combinations occur as perceptual events.
cles for DREAMIT:BP-5-L to produce a good output, and the If the probability of novel or "improbable combinations" in the
point of resumption of the sequence represented a temporal imaginal mode isa negative function ofthe proportion of possi-
discontinuity. In 20 such tests, all responses were members of ble combinations in the training sets, then DREAMIT:BP simu-
the same subset, that is, the first segment pairs were un- lations have artificially suppressed the opportunity for novel
changed. segment combinations to occur.
Furthermore, the influence of the good segments seemed to DREAMITBP5-4S-.25 (5 segments) and DREAMIT:BP6-
be nonlinear; that is, output produced in response to input with 4S-.25 (6 segments, 11 hidden units) were designed to use only
a missing or degraded segment was based on the interactive or 0.25 of their possible 32 and 64 output combinations, respec-
nonlinear relations among the segments rather than the linear tively. In all but one of the three BP training phases with each
sum of their effects. Thus the missing parts of the stimulus may model, it was not possible to satisfy the minimum learning
be,- in effect, "filled in" or "figured out." Suppose, for example, criterion on one ofthe perceptual test sequences. In the subse-
that all of the sequences in Table 1 were equally well learned quent steps of the imaginal phase (omitting the soft steps) the
and that DREAMIT:BP-5-L produced a degraded version of chance probability of a novel event was 0.81 for DREAM-
0011000011, a response that does not occur in the input list. It IT:BP5 and 0.78 for DREAMIT:BP6. Disregarding the steps in
will produce as output the correct response to one of the three each sequence that were not adequately learned, DREAM-
stimuli (see Table 1) to which it is most similar, namely the IT:BP5-4S-.025 and DREAMIT:BP6-4S-.25 produced no novel
inputs of Table 1, row 1, 00111100II; row 3, 0011001111; or combinations until they reached the last sequence in the learn-
row 7, 0011000000. ing set. At that point they tended to produce two or three out-
If a stimulus generalization, nonlinear response process such puts in which one or two ofthe segments were degraded, but
as this operates in the production of sleep imagery it could then went on to recover with good output within the next few
account for how the conceptual context of "kitchen" can pro- cycles. Slightly more than 50% of these recoveries included one
duce the image of a "cup of coffee," and the visual image of a novel output event, that is, an improbable combination, which
"cup" can elicit the experience of "kitchen." Thus a degraded typically differed by only one segment from a previously
image of "small" and "white" in the context of playing catch learned output event. For example, after rounding and collaps-
should produce a "baseball" but might produce a scrap of ing over the unit pairs of each segment, DREAMIT:BP6-4S-.25
"paper" on the floor in a "classroom" context. By contrast, the produced an approximation to the novel 111011, where the
linear functions of DREAMIT:S sometimes produced the bi- most similar output in the training set was 111010. Thus,
zarre combination of both simultaneously. DREAMIT:BP can simulate the imaginal production of im-
As stated earlier, the inhibition of appropriate kinaesthetic probable combinations, but these novel events are always the
feedback from motor commands has been proposed as a spe- best possible responses within the constraint of prior learning.
cific antecedent to bizarreness. The consequence of this This section has shown that simulations of the imaginal
DREAMING: COGNITIVE AND CORTICAL PROCESSES 117

mode of various DREAMITBP models have satisfied the goal units much as thalamic neurons control the activation of corti-
of producing continuous sequences of imaginal events, where cal neurons. With A < 0.7, excitation from other units in the
each event is a nonlinear function of the preceding event in the matrix cannot overcome the decay rate of the local units and
sequence. It has also produced sequence discontinuity and DREAMIT:S fails to activate sufficient units to represent a
novel combinations among response segments that simulate the schema of any sort. With A = 0.9, DREAMITrS can simulate
discontinuities and improbable combinations of REM sleep the intermediate level of distributed activation that is charac-
and waking imagery and thought. teristic of late morning NREM sleep. Compared to A = 1, the
Repetitive mentation as looping in DREAMITBP5-4S-.25 and number of items present in schemata are reduced, and the tem-
DREAMITBP6-4S-.25. Every simulation run with DREAM- poral progression is slower and more stereotyped. Because of
IT:BP5-4S-.25 and DREAMIT:BP6-4S-.25 produced a loop of the previously described problems encountered with DREAM-
event sequences even though they were not trained to do so. A ITS, however, extensive tests with intermediate values of A
loop was often initiated when an imaginal sequence ran to the were not pursued.
end of a previously learned sequence subset. Whether the out- A similar diffuse activation function can be applied to the
put at the end of the subset looped back to the beginning of the hidden and output units of DREAMIT:BP5 (Figure 5, layers 1
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

subset or to some subsequent point depended on the similarity and 2), although that simulation has not been carried out.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

of the last output to the various input events in the training set The simulation of motivation, meaning, and metaphor is
as described earlier. The high rate of looping in the DREAM- beyond the scope of this article but not beyond the reach of the
IT:BP5-4S-.25 and DREAMIT:BP6-4S-.25 models may be a DREAMIT models. The linear influences of goal states and
consequence of the small proportion of possible output events motivation can be represented in DREAMITrS by assigning
represented in the training set (i£., 0.25), or these models may positive activation resting-level bias values to particular local
be inappropriate to model imaginal sequences. On the other units in a network. Plans have been represented in backpropa-
hand, sleep mentation sampling procedures, which require the gation models by Jordan (1989). These plans, goals, or motive
termination of the image sequence in order to get a verbal re- states are represented independently of external input, and they
port, may obscure the extent to which looping is actually char- provide input to a portion of the input layer. The external input
acteristic of sleep mentation. can thus be taught to respond differently to the external input
Repetitive thought in humans is regarded as pathological or as a function of different plan states.
controlled by states of extreme motivation. Nevertheless, the As I suggested in earlier papers (Antrobus, 1977,1978), to the
repetitive looping in the waking state of subvocal music phases extent that dream events appear to represent combinations of
is a familiar experience for many persons. Although this char- features of events in one's waking life, they are metaphoric repre-
acteristic of sleep mentation has not been studied systemati- sentations of waking experience. Thus, the imaginal sequence
cally, Rechtschaffen, Vogel, and Shaikun (1963) did describe of 11101 -» 11110 -» 1 111 1, where the digits represent features
one subject who reported dreaming about her examinations or microfeatures, may symbolize a metaphoric representation
over a 6-hr period encompassing two REM and NREM re- of the waking sequence 11001 -» 11010 -» 11011. It is a meta-
ports. phor because the events in the sequence differ from the waking
The effect of slate shifts in activation on imagery and thought in sequence only by one (the middle) feature. For example, a logger
DREAMITS and DREAMIT.BP. Now that local activation and trainman awaiting surgery for vascular obstructions both
characterized by REM and waking processes has been simu- dreamed about unclogging objects in their more occupational
lated, one may reexamine the state shifts in nonlocal, molar, or environments of pipes and rail switches, respectively. Neither
distributed activation that were reviewed at the beginning of dreamed about the vascular obstructions, which were to them
this article. It is assumed that the distributed activation of the much less familiar (Breger, Hunter, & Lane, 1971). In each
association cortex is high in waking and in Stage 1 REM sleep REM fantasy, the feature to unclog, which had been activated by
but much lower in NREM sleep. Just as high distributed activa- the waking anticipation of vascular surgery, apparently acti-
tion is defined by active cognitive and behavioral output, so too vated features that had stronger learned links to unclog in the
the simulation of an activated state should be defined by the occupational environments of the two workers.
output of well-organized schemata. By this criterion all of the Dream reports such as these imply that action biases created
simulations described thus far have been carried out by fully in recent waking state episodes may persist into the sleep state,
activated networks. The problem, then, is to simulate how the but for reasons not understood, the units that represent these
pontine brain stem-to-thalamus control centers reduce their goals may be unable to activate the object units of recent waking
diffuse cholinergic support of cortical activation during perception, particularly the visual units, to which they were
NREM sleep. attached during the waking episode. The goal units may well
This shift in molar activation is simulated in DREAMIT:S by activate some of the features of the objects in the waking epi-
assigning a state-specific constant activation value, A, to a con- sode, but not enough to fully activate a visual representation of
trol unit, PRF-THM, that represents the pontine reticular for- that object.
mation-thalamus activation center, where A = I for waking and
REM sleep and 0.7 for NREM sleep. PRF-THM then assigns A
Conclusion
to all excitatory control subunits (see Figure 4 and Appendix;
"Sigma Pi units" in Rumelhart, Hinton, & McClelland, 1986, This article was premised on the assumption that the study of
pp. 73-74), which in turn multiply all excitatory input to each the neurocognition of thought and imagery within the sleep
local unit by A. The local excitatory units act on the output state can benefit from mapping knowledge of sensory and cog-
118 JOHN ANTROBUS

nitive psychology, sleep neurophysiology, and sleep mentation dreaming and nondreaming sleep. Archives of General Psychiatry, 12,
research onto connectionist models. A review of the sleep litera- 395-401.
ture showed that the salient characteristics of dreaming imag- Aserinsky, E., & Kleitman, N. (1953). Regularly occurring periods of
ery and thought can be attributed to the periodic activation ocular motility and concomitant phenomena during sleep. Science,
118, 361-375.
during sleep of cortical structures that support waking percep-
Aston-Jones, G., & Bloom, F. E. (1981). Norepinephrine-containing
tual, conceptual, and motor processes. Under the very high
locus coeruleus neurons in behaving rats exhibit pronounced re-
sensory and motor inhibition of Stage 1 REM sleep, this por-
sponses to non-noxious environmental stimuli. Journal of Neuro-
tion of the mind/brain runs as a closed system, where its input
science, 1, 887-900.
is its own output. The activation of REM sleep is further en- Baghdoyan, H. A., Rodrigo-Angulo, R. W, McCarley, R. W, & Hobson,
hanced during the rising phase of the diurnal rhythm so that J. A. (1984). Site-specific enhancement and suppression of de-
late morning dreaming is particularly dramatic. The subcorti- synchronized sleep signs following cholinergic stimulation of three
cal sources of excitation and inhibition are both attenuated in brain stem regions. Brain Research, 306, 39-52.
the NREM portion of sleep so that imagery is impoverished Bakan, P. (1976). The right brain is the dreamer. Psychology Today, 9,
and the recognition of external stimuli is restricted. 66-68.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

DREAMIT:S, a modified version of the schemata model of Bloom, F. E. (1979). Chemical and integrative processes in the nervous
This document is copyrighted by the American Psychological Association or one of its allied publishers.

system. In F. O. Schmitt & F G. Worden (Eds), The neurosciences:


Rumelhart, Smolensky, McClelland, and Hinton (1986) simu-
Fourth study program (pp. 51-58). Cambridge, MA: MIT Press.
lated brief sequences of schematalike images and thoughts and
Bloom, F. E. (1980). How neurotransmitters may differentiate a cell's
reduced its output as a consequence of reduced activation from
role in sensorimotor integration and behavioral state control. Neuro-
simulated RF and thalamic control sites. It also modified its
sciences Research Program Bulletin, IS, 19-21.
schemata as a consequence of inhibited afferent feedback from
Bonnet, M. H., & Moore, S. E. (1982). The threshold of sleep: Percep-
the motor commands of prior schemata. tion of sleep as a function of time asleep and auditory threshold.
DREAMIT:BP simulations accept their own output as input Sleep, 5, 267-276.
and are able to produce continuous sequences that include two Breger, L., Hunter, I., & Lane, R. (1971). The effect of stress on dreams.
classes of bizarre mentation, improbable combinations of fea- Psychological Issues, 7 (Monograph 27).
tures and discontinuities in sequence. Broughton, R. (1975). Biorhythmic variation in consciousness and psy-
The simulations establish that the neural and cognitive con- chological functions. Canadian Psychological Review, 16. 217-239.
cepts of distributed activation and inhibition can be repre- Carpenter, G. A. (1989). Neural network models for pattern recogni-
sented successfully in connectionist models. They have pro- tion and associative memory. Neural Networks, 2, 243-257.
Crick, E, & Asanuma, C. (1986). Certain aspects of the anatomy and
vided new insights into cognitive processes, such as dreaming
physiology of the cerebral cortex. In J. L. McClelland & D. E. Rumel-
and even waking imagination, that run for extended intervals
hart (Eds.), Parallel distributed processing: Explorations in the mi-
without sensory input. In several instances these insights have
crostructureafcogmtion(Vol 2, pp. 333-371). Cambridge, MA: MIT
suggested new testable hypotheses. It is hoped that the enforced Press.
discipline of translating the abstract concepts of dreaming sleep Dement, W, & Wolpert, E. A. (1958). The relation of eye movements,
into connectionist models and the theoretical issues raised by body motility, and external stimuli to dream content. Journal of
the model building and simulations will enhance the clarity of Experimental Psychology, 55, 543-553.
future research on dreaming sleep. Evarts, E. V (1964). Temporal pattern of discharge of pyramidal tract
neurons during sleep and waking in the monkey. Journal of Neuro-
physiology, 27,152-171.
Fookson, J., & Antrobus, J. (1988). Executive control in a PDF system:
References
Automatization of task performance and mindwandering. Neural
Anderson, J. (1983). The architecture of cognition. Cambridge, MA: Networks. 1,174.
Harvard University Press. Foote, S. L. (1986). Sleep-cycle generation: Turning on, turning off,
Antrobus, J. (1977). The dream as metaphor. Journal of Mental Imag- and tuning out [Commentary on Hobson, J. A., Lydic, R., &
ery. 2, 327-338. Baghdoyan, H. A.: Evolving concepts of sleep cycle generation: From
Antrobus, J. (1978). Dreaming for cognition. In A. M. Arkin, J. Antro- brain centers to neuronal populations]. Behavioral and Brain
bus, & S. Ellman (Eds.), The mind in sleep (pp. 569-581). Hillsdale, Sciences, 9, 405-406.
NJ: Erlbaum. Foote, S. L., Bloom, F. E., & Aston-Jones, G. (1983). The nucleus locus
Antrobus, J. (1983). REM and NREM sleep reports: Comparison of coeruleus: New evidence of anatomical and physiological specific-
word frequencies by cognitive classes. Psychophysiology, 20, 562- ity. Physiological Review, 63, 844-914.
568. Franck, G., Salmon, E., Poirrier, R., Sadzot, B., Franco, G, & Maquet,
Antrobus, J. (1986). Dreaming: Cortical activation and perceptual P. (1987). Evaluation of human cerebral glucose uptake during wake-
thresholds. The Journal of Mind and Behavior, 7, 193-212. fulness, slow wave sleep and paradoxical sleep by positron emission
Antrobus, J. (1987). Cortical hemisphere asymmetry and sleep menta- tomography. Sleep. 16, 46.
tion. Psychological Review, 94, 359-368. Galin, D. (1974). Implications for psychiatry of left and right cerebral
Antrobus, J., Reinsel, R., & Wollman, M. (1984). Dreaming: Cortical specialization: A neurophysiological context for unconscious pro-
activation and perceptual thresholds: More evidence. Sleep Re- cesses. Archives of General Psychiatry, 31, 572-583.
search, 13, 101. Hobson, J. A. (1965). The effect of chronic brain stem lesions on corti-
Antrobus, J., & Saul, H. (1980). Sleep onset: Subjective, behavioral and cal and muscular activity during sleep and waking in the cat. Electro-
electroencephalographic comparisons. Waking and Sleeping, 4,259- encephalography and Clinical Neurophysiology, 19, 41-62.
270. Hobson, I A. (1988). The dreaming brain. New York: Basic Books.
Antrobus, J. S., Antrobus, J., & Fisher, C. (1965). Discrimination of Hobson, J. A., Lydic, R, &flaghdoyan, H. A. (1986). Evolving concepts
DREAMING: COGNITIVE AND CORTICAL PROCESSES 119

of sleep cycle generation: From brain centers to neuronal popula- behavioural states and activity of the cerebral cortex. Progress in
tions. Behavioral and Brain Sciences, 9, 371-448. Neumbiology, 22,155-184.
Hobson, J. A., & McCarley, R. W (1977). The brain as a dream state Park, (C. (1989). Sequential learning: Observations on the internal
generator: An activation-synthesis hypothesis of the dream process. code generation problem. In Proceedings of the 1988 Connectionist
The American Journal of Psychiatry, 134, 1335-1348. Models Summer School (pp. 85-92). San Mateo, CA: Morgan Kauf-
Hobson, J. A., & Steriade, M. (1986). Neuronal basis of behavioral state mann.
control. In V B. Mountcastle, F. E. Bloom, & S. R. Geiger (Eds.), The Perky, C. W (1910). An experimental study on imagination. American
handbook of physiology (pp. 701-828). Bethesda, MD: American Journal of Psychology, 21, 422-452.
Physiological Society. Pompeiano, O. (1967). The neurophysiological mechanisms of the pos-
Hopfield, J. J. (1982). Neural networks and physical systems with emer- tural and motor events during desynchronizcd sleep. Proceedings of
gent collective computational abilities. Proceedings of the National the Association of Research in Nervous and Mental Disease, 45, 351 -
Academy of Sciences, USA, 79, 2554-2558. 423.
Intons-Peterson, M. J., & Roskos-Ewoldsen, B. B. (1989). Sensory-per- Pompeiano, O. (1970). Mechanisms of sensorimotor integration during
ceptual qualities of images. Journal of Experimental Psychology: sleep. In E. Stellar & J. M. Sprague (Eds.), Progress in Physiological
Learning, Memory, and Cognition, 15,188-199. Psychology (Vol. 3). San Diego, CA: Academic Press.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Johnson, L. C, & Lubin, A. (1966). Spontaneous electrodermal activity Ramm, P., & Frost, B. J. (1983). Regional metabolic activity in the rat
This document is copyrighted by the American Psychological Association or one of its allied publishers.

during waking and sleeping. Psychophysiology, 3, 8-17. brain during sleep-wake activity. Sleep, 6,196-216.
Johnson, M. K., Foley, M. A., Suengas, A. G, & Rave, C. L. (1988). Ramm, P., & Frost, B. J. (1986). Cerebral and local cerebral activity in
Phenomenal characteristics of memories for perceived and imag- the cat during slow wave and REM sleep. Brain Research, 365,112-
ined autobiographical events. Journal of Experimental Psychology: 124.
General, 117. 371-376. Rechtschaflen, A., Vogel, G, & Shaikun, G. (1963). Interrelatedness of
Jones, B. E., & Cuello, A. C. (1989). Afferents to the basal forebrain mental activity during sleep. Archives of General Psychiatry, 9, 536-
cholinergic cells are from pontemesencephalic—catecholamine, se- 547.
rotonin, and acetylcholine—neurons. Neuroscience, 31, 37-61. Reinsel, R., Antrobus, J., & Wollman, M. (in press). Bizarreness in
Jordan, M. (1986). Serial order: A parallel distributed processing ap- sleep and waking mentation. In J. Antrobus & M. Bertini (Eds.), The
proach (ICS Tech. Rep. No. 8604). La Jolla, CA: University of Califor- neuropsychology of dreaming sleep. Hillsdale, NJ: Erlbaum.
nia, Institute for Cognitive Science. Reinsel, R., Wollman, M., & Antrobus, J. (1986). Effectsof environmen-
Jordan, M. (1989). Supervised learning and systems with excess de- tal context and cortical activation. The Journal of Mind and Behav-
grees of freedom. In Proceedings of the 1988 Connectionist Models ior, 7, 259-276.
Summer School (pp. 62-75). San Mateo, CA: Morgan Kaufmann. Rosekind, M. R., & Schwartz, G. E. (1988a). The perception of sleep
Kanzow, E., Krause, D., & Kuhnel, H. (1962). Die Vasomotorik der and wakefulness: 1. Accuracy and certainty of subjective judgments.
Hirnrinde in den Phasen desynchronisierter EEG-Aktivitat im na- Sleep Research, 17, 89.
turlichen Schlaf der Katze [Cortical vasomotor activity in phasic, Rosekind, M. R., & Schwartz, G. E. (1988b). The perception of sleep
desynchronized, activated EEG in the natural sleep of the cat]. and wakefulness: 2. Information used for discrimination and effects
Pfluegers Archiv, 274, 593-607. of personality style. Sleep Research, 17, 131.
Klapp, S. T., Marshburn, E. A., & Lester, P. T. (1983). Short-term mem- Rumelhart, D. E., Hinton, G. E., & McClelland, J. L. (1986). A general
ory does not involve the "working memory" of information process- framework for parallel distributed processing. In D. E. Rumelhart &
ing: The demise of a common assumption. Journal of Experimental J. L. McClelland (Eds.), Parallel distributed processing: Explorations
Psychology: General, 112, 240-264. in the microslructure of cognition (Vol. 1, pp. 45-76). Cambridge,
Rondo, T. (1988). Late REM activation and sleep mentation. Unpub- MA: MIT Press.
lished master's thesis, City College of New York. Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning
Kondo, T, Antrobus, J., & Fein, G. (1989). Later REM activation and internal representations by error propagation. In D. E. Rumelhart &
sleep mentation. Sleep Research, 18,147. J. L. McClelland (Eds.), Parallel distributed processing: Explorations
Kosslyn, S. M. (1988). Aspects of a cognitive neuroscience of mental in the microslructure of cognition (Vol. I, pp. 318-362). Cambridge,
imagery. Science, 240, 1621-1626. MA: MIT Press.
Krnjevic, K., Pumain, R., & Renaud, L. (1971). The mechanism of Rumelhart, D. E., & McClelland. J. L. (Eds.). (1986). Parallel distrib-
excitation by acetylcholine in the cerebral cortex. Journal of Physiol- uted processing: Explorations in the microstructure of cognition (Vol.
ogy, 215, 247-268. 1). Cambridge, MA: MIT Press.
LaBerge, S. (1985). Lucid dreaming. Los Angeles: Tardier. Rumelhart, D. E., Smolensky, P., McClelland, J. L, & Hinton, G. E.
Mamelak, A. N., & Hobson, J. A. (1989). Dream bizarreness as the (1986). Schemata and sequential thought processes in PDF models.
cognitive correlate of altered neuronal behavior in REM sleep. Jour- In D. E. Rumelhart & J. L. McClelland (Eds.), Parallel distributed
nal of Cognitive Neuroscience, 1, 210-222. processing: Explorations in the microstructure of cognition (Vol. 2, pp.
Morrison, A. R., & Pompeiano, O. (1965). An analysis of the supra- 7-57). Cambridge, MA: MIT Press.
spinal influences acting on motoneurons during sleep in the unre- Sakai, F., Meyer, J. S., Karacan, I., Derman, S., & Yamamoto, M. (1979).
strained cat: Responses of the alpha motoneurons to direct electrical Normal human sleep: Regional hemodynamics. Annals of Neurol-
stimulation during sleep. Archives Italiennes de Biologic, 103, 497- ogy, 7, 471-478.
516. Segal, S. J. (1971). Processing of the stimulus in imagery and percep-
Moruzzi, G., & Magoun, H. W (1949). Brain stem reticular formation tion. In S. J. Segal (Ed.), Imagery: Current cognitive approaches (pp.
and activation of the EEG. Eleamencepha/ography and Clinical Ncu- 69-100). San Diego, CA: Academic Press.
rophysiology, 1, 455-473. Sewitch, D. E. (1984). The perceptual uncertainty of having slept: The
Orem, L. (1977). Laryngeal activity during sleep. Sleep Research, 6, 55. inability to discriminate electroencephalographic sleep from wake-
Ornstein, R. E. (1972). The psychology of consciousness. San Francisco: fulness. Psychophysiology, 21, 243-259.
Freeman. Smolensky, P. (1988). On the proper treatment of connectionism. Be-
Paisley, A. C, & Summerlee, A. J. S. (1984). Relationships between havioral and Brain Sciences, 11(\), 1-74.
120 JOHN ANTROBUS

Steriade, M, Kitsikis, A., & Oakson, G. (1978). Thalamic inputs and Vertes, R. P. (1984). Brainstem control of the events of REM sleep.
subcortical targets of cortical neurons in areas 5 and 7 of cat. Experi- Progress in Neurobiology. 22, 241-288.
mental Neurology. 60, 420-442. Wasserman, P. D. (1989). Neural computing: Theory and practice. New
Steriade, M., Ropert, N., Kitsikis, A., & Oakson, G. (1980). Ascending York: Van Nostrand Reinhold.
activating neuronal networks in midbrain reticular core and related Williams, H. L., Hammack, J. R., Daly, R. L., Dement, W C., & Lubin,
rostral systems. In J. A. Hobson&M. A. B. Brazier (Eds.), The reticu- A. (1964). Responses to auditory stimulation, sleep loss and the EEG
lar formation revisited (pp. 125-167). New York: Raven Press. stages of sleep. Electrophysiology and Clinical Neurophysiology, 16,
Steriade, M., Sakai, K.., & Jouvet, M. (1984). Bulbo-thalamic neurons 269-279.
related to thalamocortical activation processes during paradoxical Williams, H. L., Morlock, H. C., & Morlock, J. V(1966). Instrumental
sleep. Experimental Brain Research, 54, 463-475. behavior during sleep. Psychophysiology, 2, 208-216.
Townsend, R. E., Prinz, P. N., & Obrist, W D. (1973). Human cerebral Wills, N., & Trinder, J. (1978). Influence of response criterion on the
blood flow during sleep and waking. Journal of Applied Physiology. awakening thresholds of sleepstages.Waking and Sleeping, 2,57-62.
35, 620-625. Wollman, M, & Antrobus, J. (1986). Sleep and waking thought: Effects
Tucker, D. M., & Williamson, P A. (1984). Asymmetric neural control of external stimulation. Sleep, 9, 438-448.
systems in human self-regulation. Psychological Review, 91, 185- Wollman, M, & Antrobus, J. (1987). Cortical arousal and mentation in
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

215. sleeping and waking subjects. Brain and Cognition, 6. 334-346.


This document is copyrighted by the American Psychological Association or one of its allied publishers.

van Valen, L. fl 973). A note on dreams. Journal of Biological Psychol- Zimmerman, W B. (1967). Psychological and physiological differences
ogy, 15,19. between "light" and "deep" sleepers. Unpublished doctoral disserta-
Vern, B. A., Schuette, W H., Leheta, B, Juel, V C, & Radulovacki, M. tion, University of Chicago.
(1987). Tonic and phasic increases in cortical oxidative metabolism Zimmerman, W B. (1970). Sleep mentation and auditory awakening
during REM sleep in cats. Sleep Research, 16, 59. thresholds. Psychophysiology. 6, 540-549.

Appendix

Some Technical Characteristics of DREAMIT:S

DREAM1T:S is a single-layered network of compound units (see Fig- each output subunit are summed independently, and algebraically in
ure 4) connected to one another by asymmetric weights, M^ . Units can the excitatory and inhibitory input subunits. Therefore, the effects of
take continuous values between I, representing full activation (a) and 0, distributed activation and inhibition on local units is accomplished by
representing maximum inhibition. Each yth unit receives activation or apportioning the functions of each unit among five subunits: (a) excit-
inhibition from the output of each /th unit in the network. The sum of atory input, (b) inhibitory input, (c) inhibitory output, (d) distributed
the incoming activation and inhibition is algebraically summed and excitation, and (e) distributed inhibition (see Figure 4). The activation
then squashed by a sigmoid function. constant, A, for a particular biological state (waking = REM = I ;
The output resting value of a unit is typically 0.5, representing no NREM = 0.7) is assigned by the operator to a pontine reticular forma-
activation. However, motivational or high frequency biases may be tion-to-thalamus excitatory control subunit (PRF-THM-E), which dis-
assigned to selected units by setting their resting values above 0.5. To tributes A to all local distributed excitation subunils. Output from each
bias DREAMIT:S to respond to temporal change, a unit is reset to a excitatory input subunit is passed to a Sigma Pi (Rumelhart, Hinton, &
negative resting value, generally 0.4, for a fixed number of cycles pre- McClelland, 1986) input site of the output subunit, where it is multi-
determined by the operator, if the activation of the unit exceeds a plied by A and the product forwarded to the body of the output subunit
threshold, generally set at 0.95, for three cycles. where it is summed with the input from the inhibitory subunil,
Continuity within shorter time intervals is provided by adecay func- squashed, and then passed to the output of the output subunit. From
tion. Decay of a unit over time is simulated by passing a proportion of there it is distributed to all the units in the network, including the
the output value of each unit back to its input. Units decay to their decay link to itself, to which it is connected by the weight matrix.
resting values, unless supplied with new input. Slower decay, and there- The REM sleep inhibition of local afferent and efferent transmission
fore greater continuity, is assigned to contextual and role units; rapid is accomplished by asimilar inhibitory process controlled by the inhibi-
decay is assigned to perceptual qualities. The former are assigned slow tory control unit (PRF-THM-I), the value /of which is set at 1 during
decay rates, usually returning 0.8 of their value per cycle, and the lat- REM sleep simulations and 0 for the waking perceptual simulations. In
ter, 0.6. DREAMIT:S, this distributed inhibition of noncortical afferents is
Crick and Asanuma (1986) have pointed out that individual cortical restricted only to 11 units, 5 of which are "sensory-perceptual" and 6 of
units carry either activation or inhibition but not both. By giving local which are "kinaesthetic" feedback (K) units.
units both excitatory and inhibitory influence, however, DREAMIT:S All weights, and decay, control, and unit resting values are fixed
is constructed from one third the number of units and one ninth the throughout a given simulation, but activation can be delivered to any
weight matrix required for separate representation of excitatory and set of individual units during the course of a simulation so that the
inhibitory neurons, with no loss of function. network can be observed with different initial primes. While the level
Distributed influence from subcortical control centers to cortical of distributed excitation is fixed throughout a simulation and applies
units is primarily excitatory, and is exclusively so in DREAMIT:S, but to all units in the network, distributed inhibition can be applied phasi-
in REM sleep the distributed subcortical influence on afferent and on cally within a simulation. The remaining input to these units is excit-
motor efferent activity is inhibitory. Because summation and product atory. Local cortical, motor decision units may activate local K feed-
computation cannot both be carried out within a single site, using the back units, but this feedback is inhibited if PRF-THM is in the REM
RochesterConnectionist Simulator, the positive and negative inputs to state.
DREAMING: COGNITIVE AND CORTICAL PROCESSES 121

Furthermore, if "I-run" and associated cortical units activate the weights whose mean absolute value is much higher than that of the
"K: I-am-running" unit, then if, and only if, the "I-run" units remain weights connecting units in different modules.
active but the "K: I-am-running" unit fails to become active, other Simulations were carried out on a Sun 3/60 Workstation using the
portions of the network will generate a dreamer's common response, Rochester Connectionist Simulator, Version 4.
"I-can't-run" and "I-can't-move."
Although the visual, conceptual, motor, and motor feedback units
belong to functionally and neurally distinct processing modules, this Received January 26,1989
modular organization in DREAMITrS is represented in a single weight Revision received March 26,1990
matrix. Units that belong to the same module are interconnected by Accepted April 3,1990 I
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

Butcher, Geen, Hulse, and Salthouse Appointed


New Editors, 1992-1997

The Publications and Communications Board of the American Psychological Association


announces the appointments of James N. Butcher, University of Minnesota; Russell G. Geen,
University of Missouri; Stewart H. Hulse, Johns Hopkins University; and Timothy Salthouse,
Georgia Institute of Technology as editors of Psychological Assessment: A Journal ofConsulting
and Clinical Psychology, the Personality Processes and Individual Differences section of the
Journal of Personality and Social Psychology, the Journal of Experimental Psychology: Animal
Behavior Processes, and Psychology and Aging, respectively. As of January 1,1991, manuscripts
should be directed as follows:

• For Psychological Assessment send manuscripts to James N. Butcher, Department of Psychol-


ogy, Elliott Hall, University of Minnesota, 75 East River Road, Minneapolis, Minnesota
55455.

• For JPSP: Personality send manuscripts to Russell G. Geen, Department of Psychology,


University of Missouri, Columbia, Missouri 65211.

• For JEP: Animal send manuscripts to Stewart H. Hulse, Johns Hopkins University, Depart-
ment of Psychology, Ames Hall, Baltimore, Maryland 21218.

• For Psychology and Aging send manuscripts to Timothy Salthouse, Georgia Institute of
Technology, School of Psychology, Atlanta, Georgia 30332.

Manuscript submission patterns make the precise date of completion of 1991 volumes uncer-
tain. Current editors will receive and consider manuscripts through December 1990. Should
any 1991 volume be completed before that date, manuscripts will be redirected to the newly
appointed editor-elect for consideration in the 1992 volume.

You might also like