A Critical Review of Neural Net Theories of REM Sleep
Total Page:16
File Type:pdf, Size:1020Kb
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").