A Critical Review of Neural Net Theories of REM Sleep

MARKBLAGROVE

Department of Human Sciences, Brunei University, Oxbridge, Middlesex UB8 3PH, England.

CONTENTS

Page

Abstract 228 1. Introduction 228 2. Two Neural Network Theories of REM Sleep 229 3. Relevant Evidence from Waking Psychology 238 4. Relevant Evidence from the Physiology of REM Sleep 240 5. Relevant Evidence from the Psychology of REM Sleep 243 6. Conclusion 248 References 250

227 Vol. I, No. 3,1991 Λ Critical Review of Neural Net Theories of REM Sleep

Abstract

This paper discusses two recent speculations within the field of neural networks about the purpose of Rapid Eye Movement sleep, the time of sleep during which most vivid dreams occur. These theories are part of the information-processing view of REM sleep and are based on the finding that the mnemonic efficiency of neural networks is increased if the networks are programmed to periodically enter an altered state of functioning. One theory, that of Hopfield, Feinstein and Palmer /54/, emphasises the necessity for the network to enter this state throughout its life, whereas the other theory, that of Clark, Rafelski and Winston /19/, views the altered state as a preparation for learning by the newly formed network. Evidence is presented that the former algorithm is only relevant to systems which have orthogonal, separate memories; that the phenomenology of REM dreams is different from that predicted by both theories, and that the information-processing paradigm of REM sleep is widely contested.

1. Introduction

In 1957 Dement and Kleitman reported /28,29/ periods of jerky rapid eye-movements (REMs) in sleeping subjects, during which there would be short pauses with the eyes still. They also found periodic changes in the sleeping subjects' electroencephalograms, with a desynchronised pattern of brain waves, like those present in awake subjects, during rapid eye movement sleep alternating with large slow waves during non-rapid eye movement sleep. They also found that subjects woken from REM sleep were much more likely to report that they had been dreaming than were subjects woken during non-rapid eye movement (NREM) sleep. Due to this finding the use of the word "dream" throughout this paper should be taken to mean "REM dream", unless prefixed with NREM. The idea that REM sleep is involved with the processing of information from waking life has been actively investigated for over 20 years. Many researchers have used an analogy with digital computers to propose a function for REM sleep, from Evans and Newman's /31/ idea of the, practising of waking programs and the elimination of wasteful subroutines to Schatzman's /90/ anecdotes of solving problems during sleep and Cartwright's /14,15/ work on changes of REM time and REM dream

228 Mark Blagrove Journal of Intelligent Systems content in divorcees. This information-processing approach accorded with work claiming: the beneficial effect that REM sleep has on learning and memory (e.g. /35,71/); the increased total REM time which follows periods of intense learning (e.g. /99,108/); the relation between REMS and protein synthesis in the brain /30/; and also the finding that the brain is as active during REM sleep as it is in waking life, rather than being merely passive or resting /101/. Recently workers in the field of associative neural networks have also speculated about the function of both REM and NREM sleep. Many of these researchers had discovered that the efficiency of memory recall in some nets was increased by causing the network to enter temporary states of changed activity, during which the rules of operation of the neurones were different from those normally followed. In order to evaluate these extrapolations from artificial memories I will describe the two most elaborate theories, compare these with each other, and then review relevant evidence from the psychology of memory and the physiology and psychology of sleep mentation, some of which contests the basic assumption of the information-processing function of REMS.

2. Two Neural Network Theories of REM Sleep

The Hopfield net /53/ is a collection of neurones, each of which has a firing state of +1 or -1. The neurones have many interconnections, each having a variable synaptic strength, or connection weight. The connection weights are changed during learning, when each neurone is clamped at the value to be learned as part of an overall memory matrix. Each weight is then adjusted by the product of the firing states of the two neurones involved (the Hebb rule) so that if two neurones are positively correlated then their connection weight is increased, and if they are negatively correlated then the connection weight is decreased. The result is that in the former case if one neurone is "on" at a future time, there will be a greater likelihood that the connected neurone will turn "on" also. In this way a memory is formed which is: a) robust, in the sense that the loss of a few neurones does not lead to complete failure of the net, and b) distributed, in the sense that information is held in the synapses of many neurones.

229 Vol. 1, No. 3,1991 Λ Critical Review of Neural Net Theories of REM Sleep

Unlike the memories of digital computers these memories are content addressable, since fragments of a particular memory can be used as an address to call up the remainder. After memories have been learnt a partial input can be provided and then:

"the neural state of the system changes in time under the following algorithm. Each neurone i interrogates itself at random in time, but at a mean rate W, and readjusts its state, setting [itself at +1 or -1] according to whether the input to i at that moment is greater or less than zero. The neurones act asynchronously.... there are stable states of the network of neurones, in which each neurone is either "on" and has an input > or equal to 0 or "off and has an input < 0. These stable states will not change in time. ... This network now functions as an associative memory. If started from an initial state which resembles somewhat state [A] and which resembles other [state B] very little, the state will evolve to the state [A]." /54, p. 327/

Hopfield /53/ and Tank and Hopfield /100/ described how this evolution of the state of the system can be envisaged as the movement of a ball bearing across a bumpy surface, memories being analogous to troughs in the surface. Each point in the space corresponds to a certain pattern of active and inactive units. However, it was found that for a net which had learnt many memories some of the superimposed memories were easier to recall than others. There was also a class of "parasitic" traces of the type below, defined as a mixture of previously inputted traces /54/: inputted memory 1 + + + +- -- - + +- + - + -- inputted memory 2+ + + +------+ - + - + + inputted memory 3 + +- - + +- - + -- + +- _ + parasitic memory + + + +- -- - + -- + +- - + which Hopfield et al /54/ state would have a human counterpart of: inputted memory 1 Walter white inputted memory 2 Walter black inputted memory 3 Harold grey parasitic memory Walter grey

230 Mark Blagrove Journal of Intelligent Systems

Crick and Mitchison /23/ note that this mixing:

"is especially likely to happen if the patterns are not totally distinct but have some parts in common." (p. 234 /104/.)

In order to eliminate these two difficulties Hopfield et al /54/ propose the application of an "unlearning algorithm" to the net. This involves giving the network a random input and allowing it to settle into the nearest minima (a small local trough) or minimum (global) energy trough. This process evokes memories which are present, so that it is possible to decide which particular inter-neural connections need to be changed. If the net were learning, the connection weights would then be made to follow the rule that each connection weight changes by an amount equal to the product of the firing states of the neurones it connects. Instead, when unlearning is required, the connection weights are changed by:

-€ [product of firing states], where e is a small positive constant, much less than 1.

In doing this the unlearning mechanism "separates the memories"; they are decorrelated, so that (a) they interfere less with each other, and (b) their accessibilities equalise. When the algorithm was applied to a net with 5 memories the following result was obtained /54/:

As required, the real memories become more equal in accessibility, and many of the spurious memories are weakened. Hopfield et al /54/ claim that "...only a detailed analysis shows why the spurious states should be so sensitive to [unlearning] (p. 328)" and warn that "too much unlearning will ultimately destroy the stored memories" (p. 328). Crick and Mitchison state that "both during cortical growth (when we may say that certain broadly predetermined "associations" are laid down), and also in facing the experiences of adult life, such parasitic modes will be unavoidably generated" 1221. They claim that if these are not eradicated "certain patterns of behaviour" are likely to emerge:

"(1) The net may produce many far-fetched or bizarre associations ("fantasy"). (2) The net may tend to produce the same state, or one

231 Vol. 1, No. 3,1991 A Critical Review of Neural Net Theories of REM Sleep

Accessibility

No. of unlearning trials

Fig. 1: The fraction of random starting states which leads to particle memories (accessibility). The five dashed lines are the five nominal memories. The solid line is the total accessibility of all spurious memories. In these trials « was set at 0.01. [Reprinted by permission from Nature Vol. 304 p. 158. Copyright © 1983 Macmillan Magazines Ltd.]

of a small set of states, whatever the input ("obsession"). (3) Certain kinds of nets, particularly those which feed back on themselves, may respond to inappropriate input signals which would normally evoke no response from the net ("hallucination")." /22/

232 Mark Blagrove Journal of Intelligent Systems

Crick and Mitchison proposed that such unwanted waking behaviour can be reduced by:

"a special mechanism which operates during REM sleep and has the character of an active process which is, loosely speaking, the opposite of learning ... [but] which is not the same as normal forgetting" 1221.

It is this algorithm which they propose explains one of the characteristic features of REM dreams, that of "condensations" or "bizarre intrusions"; for example: In a dream recorded by Ferenczi, a composite image occurred which was made up from the figure of a doctor and of a horse and was also dressed in a night-shirt /45, p. 438/. Crick and Mitchison /23/ write that:

"one rarely dreams of a previous event in correct detail. More typically, the intrusion consists of a mixture of features, all or most of which can be related to events which have occurred recently. These bizarre intrusions seem to follow at short intervals, perhaps every second or so."

This is, of course, an overestimate of the amount of bizarreness and mixed memories in REM dreams. In a phenomenological study Hunt /57, p. 568/ notes that only 9% or so of REM dreams are "intrinsically novel- bizarre", and that 5% are rated highly bizarre. This makes that feature a dubious one on which to base a theory of dreaming. Before unlearning is applied the memories have to be evoked. Crick and Mitchison 1221 put it thus:

"The major inputs and outputs of the system should be turned off, so that the system is largely isolated. It should then be given successive "random" activations, from internal sources, so that any incipient parasitic modes would be excited..." 122, p. 112/.

They claim that in mammalian systems this evocation is caused by the intermittent Pontine Geniculate Occipital stimulation from the brain-stem, which is one of the physiological indicators of REM sleep. This stimulation originates in the pons, a part of the brain stem, goes thence to the lateral geniculate nucleus (simultaneously stimulating the REMs), and finishes in

233 Vol. 1. No. 3.1991 A Critical Review of Neural Net Theories of REM Sleep the occipital cortex. Whichever memory traces react to this stimulation are supposedly damped down, leaving a greater proportion of non-parasitic traces.

"The mechanism we propose is based on the more or less random stimulation of the forebrain by the brain stein that will tend to excite the inappropriate modes of brain activity referred to earlier, and especially those which are too prone to be set off by random noise rather than by highly structured specific signals." 1221

Any resulting activity is then modified so that it is less likely to occur in the future. The theory holds that this will make the memory more efficient. With regard to its emphasis on the functional significance of PGO stimulation Crick and Mitchison's theory has similarities with the Activation-Synthesis (A-S) theory of Hobson and McCarley /52/. In the A-S theory the activated forebrain "synthesises the dream by comparing information generated in specific brain-stem circuits with information stored in memory", whereas in the unlearning theory the brain-stem stimulation contains no information at all. Instead, information (usually parasitic) is said to be evoked by it. The Activation-Synthesis theory states that the best fits to the incomplete data provided by the primary stimuli are called up from memory. The primary sensorimotor stimuli thus provide a frame into which ideational, volitional and emotional content is projected to produce the final REM dream images. Such features as scene shifts, time compression and condensations may thus be dictated by the brain-stem, they claim. Similarly, the unlearning theory contends that condensation (i.e. mixed memories) will occur following a PGO burst. Crick and Mitchison comment on this point that either "the brain is put into a condition such that mixed responses (fantasies) are more common, possibly induced by a reduction in the activity of some or all of the inhibitory neurons" or that the shock nature of the PGO waves, perhaps because they are so unlike the normal somewhat structured inputs to the brain, may tend to promote such mixed responses" /23, p. 242 [112]/. Such a connection between dreams and insanity has a long history, witness the statement by Hartley /49, p. 389/ cited in /69/ that:

"The wilderness of our dreams seems to be of singular use to us, for interrupting and breaking the course of our associations. For, if we

234 Mark Blagrove Journal of Intelligent Systems

were always awake, some accidental associations would be so much cemented by us that nothing could afterwards disjoin them; which would be madness."

Robert /85/ had a similar idea of dreams as the excretion of pathological thoughts. (However, although early work by Dement /26/ reported that REM deprivation may precipitate psychosis in human subjects this was not replicated and was later denied by him 111 I.) Whereas the "unlearning" algorithm is applied after learning, the "brainwashing" mechanism of Clark, Winston and Rafelski /18/ and Clark, Rafelski and Winston /19/ is used to prepare a newly formed net for learning. It is not intended to reduce the accessibility of spurious states. Brainwashing "punishes" all active units, in the sense that when unit i fires at time Τ and unit j fires at time T+t then their connection weight's absolute value decreases, whether the connection is excitatory or inhibitory. Clark et al /18/ applied this to nets with initial quasi-random connectivity in order to make the net have more complex cycle modes, which are more useful for learning. It also prevents catastrophic behaviour such as "epilepsy" (many neurones staying on) or dying (many neurones staying off). The net is made less dependent on its initial conditions (such as initial activity, refractory time, threshold values) and hence better able to learn; the initial weights distribution also has less influence on the ultimate distribution in a brainwashed net. Irregular wide swings of activity occur, unlike the initial regular, high level activity /18/. Like increasing the number of interconnections this change in the weights has the effect of making the units more dissimilar, leading to longer periods of cyclic modes, and more "eligible" neurones, that is, those that are not constantly either on or off. Also not too many fire at once, for otherwise next they would all be within their refractory period at the same time. Their initial speculation concerning the biological analogue of this procedure was:

"that a type of synaptic modification akin to the brainwashing procedure may be found to play a role in the late foetal and early postnatal development of vertebrate nervous systems... There is not only widespread and systematic cell death, generally at the time that the first synaptic contacts are being made, but also, at later stages, widespread and selective weeding out of redundant or erroneous

235 Vol. 1, No. 3,1991 A Critical Review of Neural Net Theories of REM Sleep

synaptic connections ... to correct critical gross errors made during prior formation of connections..." /18, p. 210-211/.

Such a suggestion, of course, does not address the question of why the mechanism persists into adult life rather than disappearing, for instance as Babinsky's reflex does in the newborn. They proceeded to note the work of Crick and Mitchison, which (they stated) proposed "that a process similar to brainwashing ... may play a key role in dream sleep" /18, p. 211/ but then they change in the next paper to the suggestion that brainwashing occurs during NREM sleep in order to "wipe out all but the deepest traces of the newly learned patterns ... the "nonessentials" of the previous day's recorded experiences are eradicated..." /19, p. 246/. There is no mention here of parasitic memories, just ones that are weak and unimportant. They suggest that once these nonessential memories are removed there will be a temporary period of "positive learning" with "heightened noise", during which excitatory connection weights between positively firing neurones are increased, and inhibitory connection weights between positively firing neurones are made less inhibitory. The heightened noise will result in a greater stochastic element in neural firing. Noise is increased by having the units fire probabilistically, rather than just summing up their inputs and firing accordingly. This will "promote spontaneous associations among and abstractions upon the recently formed memory traces as well as the more firmly established ones - this is the model analog of dreaming." /19, p. 246/ The usefulness of probabilistic firing is that it can cause a net to jump from one cyclic mode to another, so that the system doesn't get stuck in unwanted local minima. It is also useful in pattern recognition, for finding the nearest fit, as proposed by Kirkpatrick, Gelatt and Vecchi /62/ for the process of "simulated annealing" in the Boltzmann machine, in which the system starts in a highly stochastic state which is gradually made less so. Under this algorithm the system ends up in the trough of lowest energy, rather than falling into and remaining in smaller troughs on the way. This may have some relation to de Bono's concept of lateral thinking, which involves looking at a problem in a new way, sometimes by having a new starting point in assessing a problem. "A random input from outside can serve to disrupt the old pattern and allow it to reform in a new way." /24, p. 243/ De Bono likens this new way of thinking to poetry, where words

236 Mark Blagrove Journal of Intelligent Systems are used in provocative and extraordinary ways. (The similarity of dreaming to "involuntary" poetry and to metaphor has been elaborated by the psychoanalyst Charles Rycroft /89/ and also by Baylor and Deslauriers 16,11.) This extraordinary way of using a word has no validity in itself, it may be completely nonsensical, but it results in a new arrangement of information, as with insight and humour, which both raise a smile and release tension, and "involve the restructuring of patterns" /25, p. 10/. He gives the example of Leonardo da Vinci's diaries being lost for centuries because of their having been misfiled in a library, and states that we similarly need to shake up our stored information in order to avoid dogmatic thinking, despite its advantages of quickness of response. Rational thinking is then applied to the results of the lateral thinking. De Bono states that "Vertical thinking is used to dig the same hole deeper, lateral thinking is used to dig a hole in a different place" /25, p. 12/ and that "The distinction between the two sorts of thinking is sharp. For instance in lateral thinking one uses information not for its own sake but for its effect. In lateral thinking one may have to be wrong at some stage [a bizarre intrusion?] in order to achieve a correct solution; in vertical thinking (logic and mathematics) this would be impossible. In lateral thinking one may deliberately seek out irrelevant information." /25, p. 11/ Roberts /86/ makes the similar point that dreams are concerned with updating our personal myths, with "reconstructing" our cognitive worlds. In addition, Krippner /63/ hypothesises that dreams are the synthesising of one's existing myths with one's life experiences, but also that they can sometimes be conservative rather than instigators of change. Clark et al /19/ relate their neural theory to sleep by suggesting that during NREM sleep non-essential experiences are eradicated and that during REM sleep important experiences are associated and abstracted. This is certainly in line with the difficulty in remembering mentation in the former case and the creativity of mentation in the latter. It is the former that they associate with brainwashing, the latter with "positive learning" plus "heightened noise". This would tie in with the claim of Schatzman /90,92/ that some dreams actively solve waking problems. Conversely, Crick and Mitchison /23/ state that the bizarreness of REM sleep dreams argues in favour of its being the time for brainwashing of connections already present, rather than that it is "abstracted' learning, the finding of new connections, a kind of loose thinking, that is causing the bizarreness. They state that for REM dreams to be involved in the consolidation of memories,

237 Vol. 1, No. 3.1991 Λ Critical Review of Neural Net Theories of REM Sleep

"such memories seem to us a very odd choice to reinforce... it is possible that the other type of dreams, more typically found in NREM sleep, may be part of a memory consolidation process, especially, for more recent events. It will be recalled that in more normal circumstances a period of REM is preceded by a period of non-REM, and it would not be unreasonable if recent memories were "consolidated" before they were "edited" by the postulated reverse learning process." /23, p. 246[116]/

To conclude this section, psychological analogues of two neural network algorithms have been proposed, although the evidence for the models is meagre.

3. Relevant Evidence from Waking Psychology

Although Crick and Mitchison /23/ state that "it is practically impossible to test such a theory decisively at the macroscopic level" (p. 24), some evidence from psychology may be useful in deciding the value of their theory. The relevance of Hopfield's parasitic memories to mammalian brains depends upon the presence of physical interference between traces in biological memories. Such a notion was once popular in psychology, but according to Baddeley /4, p. 93/:

"Interference theory has difficulty demonstrating its relevance outside the rigid confines of the verbal-learning laboratory and faces increasing competition from more cognitive views of memory."

Postman /80/ reviewed the history of research into interference theory and decided that response competition is not accompanied by the physical deterioration of old memories by newly learnt ones. Although the overlapping of traces is obviously a disadvantage for a simple storage system, it is essential for an intelligent system. Gardner- Medwin /46/ states, with regard to his own design of an associative net, "The associations between elements in human experience are by no means random and equiprobable in the way assumed for the analysis in this paper." Similarly, G. Hinton /51/states:

238 Mark Blagrove Journal of Intelligent Systems

"If there is some neat regularity in the mapping from input vectors to output vectors, we would like the [system] to 'capture' this regularity ... to give the 'correct' output vectors for input vectors it has never seen before."

Surely unlearning, the orthogonalizing of memories which in reality are correlated, will decrease the possibility of this "capturing". The reliance of the Hopfield net on orthogonal memories is illustrated by its microchip version. This is designed to work with almost equal numbers of +1 and -1 units at any one time (completely different from the operating of cortex), and, related to this, in that a preponderance of positive or negative units means that memories are not orthogonal, "... any correlation between memories causes some memories to be weighted more heavily than others. Statistical orthogonality between memories minimises this effect." /98/ The importance of this problem is understated by Crick and Mitchison, who merely comment: "Of course, this confusion of memories is not always undesirable. It is at the root of imagination, fantasy, poetry, etc. ... [As a result of unlearning] the system becomes less imaginative and more prosaic in its behaviour." /23/ The system must show less intelligence as well. A further problem with the analogy of the Hopfield net with human brains is in the method of recall. Gardner-Medwin /46/ notes that during waking life "deductive processes may assist the associative processes in recall", by generating correct details and inferences or by eliminating spurious details. This may be a functional alternative to the unlearning mechanism if parasitic traces do exist, for the intelligent recall will simply ignore them. He goes on to note that dreams have many of the characteristics of recall in which the constraints of logic have been removed, such as when we overlook bizarreness. Foulkes /41, p. 178/ makes the similar point that "in dreaming there often seems to be a dissociation of inference markers [knowledge of permissible inferences filed with a proposition] from the propositions to which they are attached." Gardner- Medwin says that the reasoning processes and "the stringent demands of waking life" may actually act as a censor for some memories, and that dreaming is needed in order to keep them strong - this is thus the opposite of the unlearning theory! (This also resembles the psychodynamic theory that dreams may express thoughts which are morally censored in the day.) Note that he is suggesting that REM sleep reruns actual memories, whereas Clark et al suggest REM sleep forms new associations, while Crick and Mitchison propose that spurious memories are accessed and unlearned.

239 Vol. 1, No. 3,1991 A Critical Review of Neural Net Theories of REM Sleep

4. Relevant Evidence from the Physiology of REM Sleep

Although the Crick and Mitchison theory seems to predict that more primitive mammals will need less REM sleep (in fact, humans have the same proportion as moles) the additional variable of brain size means that no such simple prediction can really be made. This is because a greater brain size could also supposedly separate memories, and hence obviate the need for an increase in REM sleep time in a less primitive mammal. As empirical evidence from natural history Crick and Mitchison 1221 noted that the two extant examples of monotremes (the most primitive mammals, which hatch their young from eggs), the duck-billed platypus and the spiny anteater, not only do not have REM sleep but also have frontal cortex which is disproportionately large in comparison with other mammals. They suggest that the absence of dreaming means that the memories must be made separate by the simple alternative solution of having a larger brain, which would accord with Rumelhart et al's /88/ statement that networks show undesirable behaviour if they don't have enough neurones or neural connections present. Winson /106/ explains this finding by the claim that the larger cortex is needed to perform the waking processing of connecting memories together, which he says other mammals achieve in REM sleep. This explanation would accord with the ideas of Clark et al, and also with the work of Palombo /77,78/, who holds that in dreams condensations work for the comparison of past memories with recent memories as a method of finding the correct place for their storage. A problem with the claim that it is the brain's memory areas that are changed during dreaming is that PGO stimulation, a basic component of all mammalian REM sleep, is found in perceptual rather than memory areas of the brain. It travels from the brain-stem and ends up in the occipital cortex, and the saw-tooth waves produced by this then travel from the occipital cortex over the top of the head to the frontal lobes /107/. The link made between Hopfield's work and the physiology of dreaming is that the random stimulation, which leads to the brain state which is unlearned, is analogous to the periodic stimulation ('PGO bursts') which the cortex receives from the brain stem during REM sleep, and which correlates with the production of individual REM bursts. It was initially speculated /3,76/ that dreaming subjects should have more vivid images at the moment that a 'phasic' burst of either REM or PGO activity occurs, and

240 Mark Blagrove Journal of Intelligent Systems that they may also be more likely to experience a 'discontinuity' in the dream experience, or a bizarre scene, than the subject would during the 'tonic', quiescent pauses during REM sleep. Molinari and Foulkes /75/ initially found evidence for this, but their work did not adequately obtain the objective and unbiased measurement of the subjects' experiences. The work was only 'modestly' replicated /43/ and a study using saw-tooth EEG waves as an indicator of phasic events found no correlation with specific vivid images /73/. Instead, there is persuasive evidence that PGO stimulation relates to the amount of general activity there is in the dream /9,32/ and to the overall vividness of the dream /79/. In addition there is the work of Fiss, Klein and Shollar /38/ who subjected two subjects to a regimen of REM deprivation. When the deprivation ended the subjects "caught up" on the REM sleep missed by having longer REM times for a few nights, coupled with an increase in eye movement density. It was found that the dreams of the subjects were more intense, vivid, and more clearly narrated, but were not more bizarre, than they were before the deprivation procedure. Contrary to the unlearning hypothesis, bizarreness in dreams correlates with the subject's waking personality /93/ rather than with PGO burst density; furthermore, Rechtschaffen and Buchigiani /84/ found that the vividness of dreams bears no relation to their bizarreness, and Antrobus /2/ gives persuasive evidence against the relation of phasic events with either dream breaks or bizarreness. Obviously, the existence of NREM dreams /39/, the increased bizarreness of NREMS dreams through the night in the absence of PGO bursts /40/, and the cases of Sleep Onset mentation and day-dreaming /42/ show that PGO stimulation is not a necessary condition for the phenomena emphasised by Crick and Mitchison.

Thus it must be concluded that the analogy between the random stimulation of the Hopfield net and the PGO stimulation of the brain does not hold. It has been proposed that the PGO stimulation is instead part of a cortical activation system akin to waking Ascending Reticular Activating System general stimulation /44, pp. 153-159/. Bowker & Morrison /ll/ provide an alternative account of the nature of PGO bursts with their discovery that the bursts are a startle reflex which can be produced in NREM sleep and in awake animals simply by taping the animals' cage. The bursts may therefore just be an indication of the animals' alertness during REM sleep, or may even be produced by startling or vivid dream images, as opposed to causing the vividness. Such an account is devastating for the Crick and Mitchison theory, which is based upon the notion of the uniqueness of PGO bursts to REM sleep.

241 Vol. 1, No. 3,1991 A Critical Review of Neural Net Theories of REM Sleep

The idea that REM sleep is an aid to the eradication of putative mixed memories is opposed by the finding that the proportion of REM sleep is greatest in the foetus, and declines rapidly in early infancy /87/ just at the time that memory is enlarging. It may be predicted that under the unlearning theory the brain will be credited with a larger number of false connections and more parasitic information as it grows older, which would mean a greater need for REM sleep later in life. It could be replied that REM sleep is less important in later life because we are learning more of the same things, rather than anything conceptually new (see Cartwright /14/ on changes in REM sleep characteristics following divorce). Alternatively, the massive proportion of REM sleep in the foetus may not be functional but may rather be due to the immaturity of the brain /72, p. 94/, as evidenced by the presence of intermediate sleep and the 'trace alternant' EEG burst pattern /20, chp. 13/. To avoid this criticism Crick and Mitchison postulated, like Clark et al, that REM sleep also removes catastrophic problems inherent in the net before the problems of mixed traces and learning arise: "... during development, the semi-random process of making synaptic connections is likely to produce parasitic modes. It is these which must be 'unlearned' in order to obtain a well-behaved system" 122, p. 113/. They later stated that unlearning may "help to remove inappropriate connections made by the somewhat random nature of neural development." /23, p. 235[105]/ REM sleep is thus claimed to have a dual function because of the random links and resonating structures present in the foetus and neonate (for myelinization of neurones occurs until age 2 years and interneural connections are still being made). If such a preparatory function were true for REM sleep throughout life this would accord with the decrease in REM time with age only if it were also true that there were fewer and fewer neurones available, or fewer and fewer needed, for new learning, as we get older. Conversely, if the brain has ample neurones for the experiences of a lifetime, or if previously used ones are recycled, and if just as many are needed for new learning at all ages, then we would expect the amount of such preparation to stay constant with age, as is found with non-REM sleep. The latter case is predicted by Clark et al, but both meet in the suggestion by Hughlings Jackson that sleep fulfils the function of sweeping away unnecessary memories from the previous day and of consolidating or maintaining more necessary ones. Clark et al /19/ propose that brainwashing occurs during non-REM

242 Mark Blagrove Journal of Intelligent Systems sleep in order to provide a competitive disadvantage to active links while connections are still forming. However, there is a precocious development of synaptic inhibition in the newborn /82/, maybe as a brake against overexcitation before the appropriate sensorimotor systems are mature; the threshold of seizure induction is actually greater at earlier ages and spontaneous activity is virtually absent. Notwithstanding this finding, the unlearning theory could be reformulated to emphasise this preparatory function, with the conclusion that dreams are bizarre because the physiological basis for them is either an atavism from earlier life, which is possibly even maladaptive in later life, or is directed throughout life at as yet unused neurones, while coincidentally stimulating memories already in place. The idea that REM sleep protects the brain from overexcitation was hypothesised by Cohen and Dement /21/ after noting increasingly severe convulsive spasms of face and body in the REM deprived animal, and the increased availability of cells able to respond to a second click in REM deprived cats. A problem with such evidence is the extent to which the physiological and psychological changes after REM sleep deprivation are in fact due to the disturbance or loss of sleep in general (Cohen and Dement's paper only mentions 2 NREM sleep deprived controls). Furthermore, Kaelbling et al /59/ found that induced electrical seizures can substitute for REM sleep as shown by a decrease in the REM rebound effect after REM deprivation. Cartwright /13, p. 92/ states that the ability of the MAOIs to eradicate REM sleep is an indicator of their anti-depressant activity, and REM deprivation itself has a beneficial effect on depressives /103/. One proposed explanation for this is that drive-disinhibition occurs due to the REM sleep deprivation /56/.

5. Relevant Evidence from the Psychology of REM Sleep

A major problem for both neural net theories is that the information- processing account of REM sleep is contested by many authors. A common paradigm is to REM deprive animals and measure either the impairment in incorporation of amino-acids into brain proteins /68/ (which assumes that the proteins concerned are those involved with memory storage), or measure behavioral deficiencies due to memory loss /33,34/. However, whether pharmacological or physical means are used to deprive an animal

243 Vol, 1. No. 3,1991 A Critical Review of Neural Net Theories of REM Sleep of REM sleep, the concomitant stress induced, or stereotypical movements used by the animal to counteract stress, may be the actual causes of deficiencies in learning. Alternatively, if drive disinhibition occurs due to sleep or REM sleep deprivation /56, pp. 278-279/ this may itself affect learning because of attention changes irrespective of any specific information-processing changes. Even though memories are retained better after periods of REM sleep than NREM sleep or wakefulness /94,95/, this may be because of the differences in level of arousal during REM and NREM sleep, differences in arousal upon waking from REM and NREM sleep, as well as the confabulating effect of time of day and position in circadian cycle at which the measurements are taken. The best evidence comes from the work on REM time augmentation following learning Π 1,108/ although it is difficult to ensure that the only difference between experimental subjects and controls is that the former have only learnt more than the latter, rather than having had more exercise, for example. Further opposition to the REM sleep information-processing hypothesis is based on the findings of the relative unimportance of REM sleep as compared to stages 3 and 4 of deep sleep. Not only can stress diminish the amount of time devoted to REM sleep /72, p. 78/, showing its fragility in comparison to deep sleep, but after total sleep deprivation stage 4 of NREM sleep is made up in preference to REM sleep /105/. The Crick and Mitchison neural net theory treats dream imagery as both epiphenomenal and meaningless, although allowing for confabulation after waking. Against this is the finding of Fiss /37/ that the need to complete an interrupted dream by mentation in waking life was quite independent of any need to have REM sleep itself. He showed that dreams must be completed, and not just substituted for by more REM time. Being prevented from finishing a dream was found to be more disruptive, as measured by projected anger and frustration, than if the subject was not allowed to dream at all. (However, note that to be halted during a meal may be more psychically disruptive than not to be allowed to start eating!) Crick and Mitchison's theory accords with the observation that most dreams are unremembered (humans have 4-5 REM periods each night) and, if ever they are remembered, easily forgotten - any theory of dreams must take account of the fact that the vast majority of dreams never reach waking consciousness. At first sight the Crick and Mitchison theory seems to imply that chronic dream-rememberers will be more prone to instabilities, on average, than non-rememberers, because an association or

244 Mark Blagrove Journal of Intelligent Systems event in the dream should have been unlearned rather than entered into consciousness. Such a prediction accords with Crick's anti-psychoanalytic bias, in that the remembering of dreams would no longer be a virtue. In contrast to this, Singer and Schonbar /97/ found that "high and low daydreamers differ along a dimension which might be termed self- awareness, or acceptance of inner experience", and there is no study which shows a greater psychopathology among rememberers of night dreams. Schatzman /91/ stated that "If their theory were true, and their advice [of not attempting to remember dreams] valid, much psychotherapeutic practice and many people's habits of recalling dreams would have to change". However, this assertion does not follow from the theory. To remember a dream may reinforce the parasitic memories best forgotten, but the subject's musings about the source of the dream images (even if such musings are complete confabulations /23, pp. 239[109]-240[110]/) would add to the subject's knowledge of him/herself, which is beneficial, rather than a cause of 'instability'. An additional problem for the model is that bizarre links from a dream may be found to be immensely enlightening, to such an extent that even if, as the theory's proponents say, the free associations produced on waking are later confabulations and not the real causes of the dream, the cause of the enlightenment is not deserving of erasure. Such a belief that dreams display or produce knowledge is the basis of the 'positive learning' hypothesis of Clark et al /19/, and has been widely studied. The amount of enlightenment and creativity associated with dreams varies between people. Adelson /l/ reported that dreams of more creative college women are more innovative and interesting than those of their less creative peers. Similarly dream diary reports of art students were found to be more imaginative than those of science or engineering students /93/. Hartmann, Baekeland and Zwilling /50/ reported that people who sleep for long periods (averaging 8 hours) had much the same slow wave deep sleep as subjects who sleep for short periods (average of 5.5 hours): the difference between the two was in the amount of REM sleep and light stage 2 obtained. Compared to the latter group, the former group were found to be more artistic, less sure of themselves, less ambitious and less politically conformist. Hartmann concluded that they were more in need of reprogramming than 'pre-programmed' short sleepers. These results can be accepted by either of the neural net theories, which illustrates their problem with falsifiability. Crick and Mitchison would claim that the great

245 VoL 1. No. 3,1991 Λ Critical Review of Neural Net Theories of REM Sleep

amount of fantasy in some subjects during the day requires a diligent effort by the brain during REM sleep to be damped down, whereas Clark et al can claim that bizarreness during sleep partly causes the subject to be more creative during waking hours. Incidentally, Hargreaves and Bolton /48/ provide evidence that creative thinking correlates with the ability to form vivid images when awake, although they do not relate this to dreaming. However, opponents of information-processing theories would claim that differences in REM dream mentation styles passively reflect and are caused by differences in waking mentation. Although thematic apperception test stories obtained immediately after REM periods are more dreamlike than those toid after NREM awakenings /70/, and divergent thinking tasks set before sleep have better responses after NREM deprivation than after REMS deprivation /47/, this does not prove that information-processing occurs during REM sleep. Temporary arousal differences between the two states provide a simpler explanation than mentation differences with storage function and long-term effects. Similarly, the stochastic neural activations found by Warren /104/ and Pritchard /81/ do not have any information-processing function nor long- term effects. Crick and Mitchison do differ from Clark et al with regard to what would be predicted about the incorporation of external stimuli into dreams when the stimuli are presented to the sleeping subject below the threshold needed for arousal. Berger /8/ found that stimuli were incorporated by virtue of, most commonly, (a) their assonant properties, (b) directly, (c) by the use of symbolisation, or, least often, (d) by the use of association. According to the Crick and Mitchison theory, learning and thinking do not occur during dreams, hence the theory cannot explain why a stimulus is frequently incorporated as a distorted element in the dream, looking much like the "parasitic" or mixed-up elements present in many dreams. The dream in this instance is more a creator of bizarre images than a destroyer of them. The incorporated element was not randomly stimulated out of the memory system by PGO stimulation, and yet it became part of the dream. Berger /8/ also reported that meaningful and neutral names were incorporated to an equal extent, but that the former often appeared in a manner which was dependent on their personal meaning to the dreamer, showing the intelligent nature of dream mentation. Incorporation was also "at a point in the dream appropriate to the time of presentation of the stimulus", and hence definitely not the result of PGO stimulation. The

246 Mark Blagrove Journal of Intelligent Systems activity of incorporation shows that dreams are more than replayed memory traces: dreams seem to be more akin to thinking than to reminiscing, as evidenced by the greater use of real memories in sleep onset imagery and day-dreams than in REM dreams /5,17,42/. As Crick and Mitchison 1221 note, experiments that claim to show the positive effect that dreaming has on memory cannot be used to decide between their theory and more orthodox memory consolidation theories, or even Clark et al's 'positive learning' and NREM sleep brainwashing theory. From the literature on dreams, however, there are the following few reports which argue against the unlearning theory.

1. Cartwright and Monroe /16/ awakened two groups of subjects during REM sleep. Those subjects who were allowed to report and think about their dreams had a smaller REM rebound effect (the 'catching up' with lost REM sleep) than did subjects who had to repeat lists of numbers. According to the unlearning theory fantasy without PGO stimulation and its concomitant putative unlearning function could never substitute for REM sleep, and would even require greater amounts of REM sleep in the future to counteract it.

2. Fishbein, McGaugh and Swarz /26/ showed that REMS deprivation did not destroy memory traces in mice, but rather left them liable to subsequent disruption by electroconvulsive shock, which usually only causes amnesia if given immediately after training. Similarly, lesions of the locus coeruleus (an area involved with the production of REM sleep) prolong the period during which an established memory is susceptible to disruption /109/. The only maladaptive feature of the memory traces was their fragility, not any putative interference between traces.

3. Verdone /102/ showed that REM periods from early in the night have fewer memories from the distant past than do late REM periods. Although this is not incompatible with the unlearning theory, Cartwright also noted that 'dreams begin with the feelings and concerns the person was experiencing just before sleep. Those [dreams] that follow are related to the first, but are older examples of situations in which the same feelings are experienced' /12, p. 129/. She claimed that the later scenes contemplated solutions to the concerns. The importance

247 VoL 1. No. 3,199 J A Critical Review of Neural Net Theories of REM Sleep

of this narrative nature of dreams is evidenced by patients who are found to have narrative dreaming and yet no visual imagery /60,61/. Furthermore, Kuper /64,65,66/ and Kuper and Stone /67/ and Blagrove /10/ offer persuasive evidence that not only does each dream exist as a reasoned argument and narrative, but that successive dreams of each night follow on from each other in this way also.

On this theme Crick and Mitchison /23/ write that: 'As will be seen, our theory provides a good explanation of the nature of the bizarre intrusions. It has nothing to say about the narrative.' Although great importance must be attached to this issue, Crick and Mitchison are too hard on themselves at this point. A period of ratiocination could always follow each phasic burst, for they state that 'the way the brain deals with the results of PGO stimulation - especially as far as the narrative may be concerned - need not be at all random' /23/. Indeed, Seligman and Yellen /96/ speculate upon how such overlaying of the putative basic dream with narrative may occur. It may also be necessary to invoke some narrative in order just to define what the components of the hypothesised mixed memories are anyway, although the empirical question remains of how much narrative is there in a night of dreams, and why they take this form rather than just isolated pictures.

6. Conclusion

Temporary altered states of functioning have been found to be useful in associative neural networks, with the altered state either preceding, or following, learning. The authors mentioned above have hence claimed that altered states of functioning in mammals occur for a similar reason, the low-level 'cleaning' and enhancing of memory. However, much evidence linking sleep (and specifically REM sleep) with memory consolidation is contested, and there is no evidence, apart from the rather stretched analogy of the random stimulation of neural nets with PGO bursts, for the operation of a brainwashing or unlearning function in mammals. The phenomenological evidence for the mundaneness of most of the images of REM sleep indicates the simpler solution that dreams are just a more vivid and 'single-minded' /83/ form of daydreaming, with such states as hypnogogic imagery and lucid dreams on the continuum between. Whether

248 Mark Blagrove Journal of Intelligent Systems the imaginativeness of this REM dream state requires a change in the stochastic firing of neurones, as both the positive learning and unlearning theses indicate, is also speculative. The properties of dreams may well derive from changes in the higher levels of the brain, with the hardware working much as in the waking state. Certainly it is difficult to envisage a mechanism that would cause a sudden change, not just in the stochastic element, but in the rules of operation of individual synapses across the brain, such that unlearning, brainwashing, and positive learning require. The claims for unlearning in REM sleep and brainwashing in NREM sleep are bold; the former claim predicts most, such as about the function of PGO stimulation and the deleterious side-effects of REM sleep deprivation, and accordingly more psychological and physiological evidence is found against it. The postulation of positive learning must similarly await the necessary neurological evidence, when the mechanism for the physical summation of neural input to brain cells is known, before we can know if it is any more than a metaphor for the features of thought we have during REM sleep. To conclude about the analogy of the REM sleep change of state to that of temporary changes in network algorithms, associative network theorists have exceeded the experimental evidence in trying to explain the high-level symbolic activity of intelligent thinking in terms of the brain's hardware, as if all cognitive rules will simply emerge from a simple network. Large networks may have some properties not present in simpler networks, for example, Minsky /74/ suggested that small barriers may exist inside minima in order to differentiate concepts within a class, an effect which would not be seen in small networks. Similarly, Johnson-Laird /58/ drew attention to the differences between mistakes in language learning by children and in the learning of rules for the past tense by associative nets. He warns that it may be that 'high-level principles are paramount, and that parallel processes are merely the brain's low-level language into which all mental life must ultimately be translated... At the other extreme, another option is that the mind contains no high-level principles: there is only parallel distributed processing from which all behaviour emerges.' (As an example, the 'cocktail-party effect' in psychology works on a higher level than the 'attention' algorithm of Clark et al /19/.) Similarly, Hopfield and Tank's /55/ comment that: 'Hierarchy is necessary to keep the number of synaptic connections to a reasonable level. To extend the present ideas from neural circuit to neural system, such notions will be essential'. In that case the

249 VoL 1, No. 3,1991 A Critical Review of Neural Net Theories of REM Sleep hierarchical organisation of nets, as in the brain, may obviate the need for the reverse learning algorithm anyway. The issue may be best summed up by the conclusion of Crick and Mitchison's original /22/ article:

'even if it turns out that our ideas are wrong and that nature does not employ the reverse learning mechanism we have postulated, the process may well be useful for artificial intelligence machines of the future, especially those having extensive parallel processing, a learning mechanism and a certain amount of randomness in their construction.'

References

1. Adelson J. 1959 Creativity and the dream. Merrill-Palmer Quarterly, 6: 91-97. 2. Antrobus J. 1986 Dreaming: cortical activation & perceptual thresholds. Journal of Mind & Behavior, 7(2&3): 193[63]-212[82]. 3. Aserinsky E. 1967 Physiological activity associated with segments of the rapid eye movement period. In: S. Kety et al. (eds.) Sleep and altered states of consciousness, Baltimore, Williams & Wilkins, pp. 333-350. 4. Baddeley A. 1976 The psychology of memory, New York: Basic Books. 5. Battaglia D., Cavallero C. & Cicogna P. 1987 Temporal reference of the mnemonic sources of dreams. Perceptual & Motor Skills, 64: 979-983. 6. Baylor G. & Deslauriers D. 1986 Understanding dreams: method, maps and metaphor. Dreamworks, 5: 46-57. 7. Baylor G. & Deslauriers D. 1986-1987 Dreams as problem solving: part 1: background and theory. Imagination, Cognition and Personality, 6:105-118. 8. Berger R. 1963 Experimental modification of dream content by meaningful verbal stimuli. British Journal of Psychiatry, 109: 722-740. 9. Berger R & Oswald I. 1962 Eye movements during active and passive dreams. Science, 137: 601. 10. Blagrove M. 1989 The structuralist analysis of dream series. Paper

250 Mark Blagrove Journal of Intelligent Systems

presented at the Sixth Annual Conference of the Association for the Study of Dreams, London University. 11. Bowker R. & Morrison A. 1977 The PGO spike: an indicator of hyperalertness. In: Koella W. & Levin P. (eds.), Sleep 1976, Basel: Karger, pp. 23-27. 12. Cartwright R. 1977 Night Life, Explorations in Dreaming, Englewood Cliffs, N.J., Prentice-Hall. 13. Cartwright R. 1978 A Primer on Sleep and Dreaming, Reading, Mass.: Addison-Wesley. 14. Cartwright R. 1983 REM sleep characteristics during and following mood-disturbing events. Archives of General Psychiatry, 40: 197-201. 15. Cartwright R. 1986 Affect and dream work from an information processing point of view. Journal of Mind and Behavior, 7(2&3): 411[281]-428[298]. 16. Cartwright R. & Monroe L. 1968 Relation of dreaming and REM sleep: The effects of REM deprivation under two conditions. Journal of Personality and Social Psychology, 1: 69-74. 17. Cicogna P., Cavallero C. & Bosinelli M. 1986 Differential access to memory traces in the production of mental experience. International Journal of Psychophysiology, 4: 209-216. 18. Clark J., Winston J. & Rafelski J. 1984 Self-organization of neural networks. Physics Letters, 102A(4): 207-211. 19. Clark J., Rafelski J. & Winston J. 1985 Brain without mind: computer simulation of neural networks with modifiable neuronal interactions. Physics Reports, 123: 216-273. 20. demente C., Purpura D. & Mayer F. (eds.) 1972 Sleep and the maturing nervous system. Academic Press 21. Cohen D. & Dement W. 1965 Sleep: changes in threshold to ECS in rats after deprivation of PS. Science, 150:1318-1319. 22. Crick F. & Mitchison G. The function of dream sleep. 1983 Nature, 304:111-114. 23. Crick F. & Mitchison G. 1986 REM sleep and neural nets. Journal of Mind & Behavior, 7(2&3): 229[99]-250[120], 24. de Bono E. 1969 The mechanism of mind. Pelican Books. 25. de Bono E. 1970 Lateral thinking. Pelican Books. 26. Dement W. 1960 The effect of dream deprivation. Science, 131: 1705-1707. 27. Dement W. 1969 The biological role of REM sleep. In: A. Kales

251 Vol 1, No. 3, 1991 A Critical Review of Neural Net Theories of REM Sleep

(ed.), Sleep: physiology and pathology, New York: Lippencott. 28. Dement W. & Kleitman N. 1957 Cyclic variations in EEG during sleep and their relation to eye movements, body motility and dreaming. Electroencephalography & Clinical Neurophysiology, 9: 673-690. 29. Dement W. & Kleitman N. 1957 The relation of eye movements during sleep to dream activity: An objective method for the study of dreaming. Journal of Experimental Psychology, S3:339-346. 30. Drucker-Colin R., Spanis C., Cotman C. & McGaugh J. 1975 Proteins and REM sleep. In: P. Levin & W. Koella (eds.), Sleep 1974: Instinct, neurophysiology, endocrinology, episodes, dreams, epilepsy and intracranial pathology. Basel: S. Karger. 31. Evans C. & Newman E. 1964. Dreaming: an analogy from computers. New Scientist, 24(419): 577-579. 32. Firth H. & Oswald I. 1975 Eye movements and visually active dreams. Psychophysiology, 12: 602-606. 33. Fishbein W. 1970 Interference with conversion of memory from short-term to long-term storage by partial sleep deprivation. Commun. Behav. Biol., 5:171-175. 34. Fishbein W. 1972 Alterations of the sleep states: sleep deprivation: total, partial, and selective. In: M. Chase (ed.), The sleeping brain, UCLA Brain Information Service, pp. 347-355. 35. Fishbein W. & Gutwein B. 1977 Paradoxical sleep and memory storage processes. Behavioral Biology, 19:425-464. 36. Fishbein W., McGaugh J. & Swarz J. 1971 Retrograde amnesia: ECS effects after termination of REM sleep deprivation. Science, 172: 80-82. 37. Fiss H. 1969 The need to complete one's dreams. In: J. Fisher & L. Breger (eds.), The meaning of dreams: some insights from the laboratory. California Mental Health Research Symposium Monograph No. 3, pp. 38-63. 38. Fiss H., Klein G. & Shollar S. 1974 "Dream intensification" as a function of prolonged REM period interruption. In: Psychoanalysis & Contemporary Science, New York: IUP, pp. 399-424. 39. Foulkes D. 1962 Dream reports from different sleep stages. J. of Abnormal and Social Psychology, 65: 14-25. 40. Foulkes D. 1966 The psychology of sleep, New York: Scribners.

252 Mark Blagrove Journal of Intelligent Systems

41. Foulkes D. 1982 A cognitive-psychological model of REM dream production. Sleep, 5(2): 169-187. 42. Foulkes D. & Fleisher S. 1975 Mental activity in relaxed wakefulness. /. of Abnormal Psychology, 84: 66-75. 43. Foulkes D. & Pope R. 1973 Primary visual experience and secondary cognitive elaboration in stage REM; a modest confirmation and an extension. Perceptual and Motor Skills, 37: 107- 118. 44. Freemon F. 1972 Sleep research: a critical review. Springfield, IL: C. Thomas. 45. Freud S. 1976 The interpretation of dreams. Pelican. 46. Gardner-Medwin A. 1976 The recall of events through the learning of associations between their parts. Proc. Roy. Soc. Lond. B., 194: 375-402. 47. Glaubman H. 1978 REM deprivation and divergent thinking. Psychophysiology, 15:75-79. 48. Hargreaves D. & Bolton N. 1972 Selecting creativity tests for use in research. British Journal of Psychology, 63:451-462. 49. Hartley D. 1801 Observations on man, his frame, his duty, and his expectations. London: Johnson. [Cited in Lavie & Hobson, 1986.] 50. Hartmann Ε., Baekeland F. & Zwilling G. 1972 Psychological differences between long and short sleepers. Archives of General Psychiatry, 26:463-468. 51. Hinton G. April 1985 Learning in parallel networks. BYTE. 52. Hobson J.A. & McCarley R. 1977 The brain as a dream state generator: An activation-synthesis hypothesis of the dream process. American Journal of Psychiatry, 134(12): 1335-1348. 53. Hopfield J. 1982 Neural networks and physical systems with emergent collective computational abilities. Proc. Natl. Acad. Sei. USA, 79:2554-2558. 54. Hopfield J., Feinstein D. & Palmer R. 1983 'Unlearning' has a stabilizing effect in collective memories. Nature, 304:158-159. 55. Hopfield J. & Tank W. 1986 Computing with neural circuits: a model. Science, 233: 625-633. 56. Hörne J. 1988 Why we sleep: The function of sleep in humans and other mammals. Oxford University Press. 57. Hunt H. 1982 Forms of dreaming. Perceptual & Motor Skills, 54: 559-633.

253 Vol. 1, No. 3,1991 A Critical Review of Neural Net Theories of REM Sleep

58. Johnson-Laird P. 1987 Connections and controversy. Nature, 330: 12-13. 59. Kaelbling R., Koski E. & Hartwig C. 1968 Reduction of REM sleep in cats after electroconvulsions. Psychophysiology, 4: 381. 60. Kerr N., Foulkes D. & Jurkovic G. 1978 Reported absence of visual dream imagery in a normally sighted person with Turner's syndrome. Journal of Mental Imagery, 2: 247-264. 61. Kerr N. & Foulkes D. 1981 Right hemisphere mediation of dream visualisation: a case study. Cortex, 17: 603-609. 62. Kirkpatrick S., Gelatt C. & Vecchi M. 1983 Optimization by simulated annealing. Science, 220: 671-680. 63. Krippner S. 1986 Dreams and the development of a personal mythology. Journal of Mind and Behavior, 7(2&3): 449[319]-462[332]. 64. Kuper A. 1979 A structural approach to dreams. Man, 14(4): 645- 662. 65. Kuper A. 1983 The structure of dream sequences. Culture, Medicine and Psychiatry, 7: 153-175. 66. Kuper A. 1986 Structural anthropology and the psychology of dreams. Journal of Mind & Behavior, 7(2&3): 333[203]-344[214], 67. Kuper A. and Stone A. 1982 The dream of Irma's injection: a structural analysis. American Journal of Psychiatry, 139(10): 1225- 1234. 68. Lambrey-Sakai F. 1972 Cited from the French by Fishbein W. & Gutwein B. Paradoxical sleep and memory storage processes. Behavioral Biology 1977; 19:425-464. 69. Lavie P. & Hobson J. 1986 Origin of dreams: anticipation of modern theories in the philosophy and physiology of the eighteenth and nineteenth centuries. Psychological Bulletin, 100:229-240. 70. Lewin I. & Glaubman H. 1975 The effect of REM deprivation: is it detrimental, beneficial or neutral? Psychophysiology, 12:349-353. 71. Lucero M. 1970 Lengthening of REM sleep duration consecutive to learning in the rat. Brain Research, 20:319-322. 72. Meddis R. 1977 The sleep instinct. London: Routledge & Kegan Paul. 73. Medoff L. 1972 Psychological correlates of the EEG sawtooth wave. Unpublished Master's Thesis, University of Wyoming. 74. Minsky M. 1975 The psychology of computer vision. McGraw-Hill. 75. Molinari S. & Foulkes D. 1969 Tonic and phasic events during

254 Mark Blagrove Journal of Intelligent Systems

sleep: psychological correlates and implications. Perceptual & Motor Skills, 29:343-368. 76. Moruzzi G. 1963 Active processes in the brain stem during sleep. The Harvey lectures, 58:297. 77. Palombo S. 1976 The dream and the memory cycle. International Review of Psychoanalysis, 3: 65-83. 78. Palombo S. 1978 Dreaming and memory: a new information- processing model. New York: Basic Books. 79. Pivik T. & Foulkes D. 1966 'Dream deprivation': effects on dream content. Science, 153:1282-1284. 80. Postman L. 1969 Mechanisms of interference in forgetting. In: Talland G. & Waugh N. (eds.), The pathology of memory. New York and London: Academic Press. 81. Pritchard R. 1961 Stabilised images on the retina. Scientific American, 204: 72-78. 82. Purpura D., Shofer R. & Scarff T. 1965 Properties of synaptic activities and spike potentials of neurons in immature neocortex. Journal of Neurophysiology, 28:925-942. 83. Rechtschaffen A. 1978 The smgle-mindedness and isolation of dreams. Sleep, 1:97-109. 84. Rechtschaffen A. & Buchigiani C. 1983 Visual dimensions and correlates of dream images. Sleep Research, 12:189. 85. Robert W. 1886 [The dream explained as waste-disposal.] Hamburg. 86. Roberts K.R. 1985 World and world reconstruction. Advances in Descriptive Psychology, 4:17-53. 87. Roffwarg H., Muzio J. & Dement W. 1966 The ontogenetic development of the human sleep dream cycle. Science, 152:604-618. 88. Rumelhart D., Hinton G. & Williams R. 1986 Learning representations by back-propagating errors. Nature, 323: 533-536. 89. Rycroft C. 1981 The innocence of dreams. Oxford University Press. 90. Schatzman M. 1983 Solve your problems in your sleep. New Scientist, 98(1361): 692-693. 91. Schatzman M. 1983 Such stuff as REM sleep is made of. New Scientist, 99(1375): 796-797. 92. Schatzman M. 1986 The meaning of dreaming. New Scientist, 101(1540/1541): 36-39. 93. Schechter N., Schmeidler G. & Staal M. 1965 Dream reports and

255 VoL 1, No. 3,1991 A Critical Review of Neural Net Theories of REM Sleep

creative tendencies in students of the arts, sciences, and engineering. Journal of Consulting Psychology, 29: 415-421. 94. Scrima L. 1982 Isolated REM sleep facilitates recall of complex associative information. Psychophysiology, 19: 252-259. 95. Scrima L. 1984 Dream sleep and memory: New findings with diverse implications. Integrative Psychiatry, 2: 211-216. 97. Singer J. & Schonbar R. 1961 Correlates of daydreaming: a dimension of self-awareness. Journal of Consulting Psychology, 25: Ι- ό. 98. Sivilotti M., Emmerling M. & Mead C. 1985 A novel associative memory implemented using collective computation. Chapel Hill Conference on VLSI. 99. Solodkin M., Cardona A. & Corsi-Cabrera M. 1985 Paradoxical sleep augmentation after imprinting in the domestic chick. Physiology and Behavior, 35(3): 343-348. 100. Tank D. & Hopfield J. 1987 Collective computation in neuronlike circuits. Scientific American, 257(6): 62-70. 101. Vellutti R. 1985 An electrochemical approach to sleep metabolism:

a p0o paradoxical sleep system. Physiology & Behavior, 34: 355-358. 102. Verdone P. 1965 Temporal reference of manifest dream content. Perceptual & Motor Skills, 20:1253-68. 103. Vogel G., Thurmond Α., Gibbons P., Sloan K., Boyd M. & Walker M. 1975 REM sleep reduction effects on depression syndromes. Archives of General Psychiatry, 32:765-777. 104. Warren R. 1961 Illusory changes of distinct speech upon repetition - ihe verbal transformation effect. British Journal of Psychology, 52: 249-258. 105. Webb W. & Agnew H. 1971 Stage 4 sleep: influence of time-course variables. Science, 174:1354-1356. 106. Winson J. 1985 Brain and psyche - The biology of the unconscious. New York: Anchor Press/Doubleday. 107. Yasoshima Α., Hayashi H., Iijima S., Sugita Y., Teshima Y., Shimizu T. and Hishikawa Y. 1984 Potential distribution of vertex sharp wave and saw-toothed wave on the scalp. Electroencephalography & Clinical Neurophysiology, 58:73-76. 108. Zimmerman J., Stoyva J. & Metcalf D. 1969 Distorted visual experience and REM sleep. Report to the 9th Annual Meeting of the

256 Mark Blagrove Journal of Intelligent Systems

Association for the Psychophysiological Study of Sleep, Boston, Mass. 109. Zornetzer S. & Gold M. 1976 The locus coeruleus: its possible role in memory consolidation. Physiology & Behavior, 16:331-336.

257