DynaPsych Table of Contents


The Dreaming Glasses


R. Ernesto Blanco

Instituto de Física, Facultad de Ingeniería, Universidad de la República

Julio Herrera y Reissig 565, Montevideo, Uruguay,




In 1983, Crick & Mitchison proposed a function of  REM sleep as an unlearning mechanism (Ref. 1). They gives some phenomenological implications based on neural networks theory: the developing of hallucinations and obsessions in subjects deprived of REM sleep. This implications were not observed in further studies. Theoretical knowledge of the spin glasses with temperature effects is used here to explain why the implications suggested by Crick & Mitchison could be not observed even in the case that their main ideas could be right. Additionally I proposed here a explanation of some characteristics of dreams imaginery and other psychological implications derived from spin glasses models of neural networks.


In the early eighties Crick & Mitchison (Ref. 1) postulated that rapid-eye movement (REM) sleep could be a reverse learning mechanism. This mechanism was shown as very useful in enhancing the memory recovery performance of small networks of two states neurons with symmetric connections (Ref. 2). The main proposal, in Crick & Mitchison words, was: “we dream in order to forget”. Neural networks are associative memory storage device: starting from any arbitrary state it will evolve to a stable state representing a evokable pattern. This behaviour resembles some properties  of biological neural networks. In some simulations was seen  that, while some patterns are teached to the network, spurious ones are also created (Ref. 2). Another observation was that different learned patterns are not equally easy to recall. Crick & Mitchison’s hypothesis is that the spurious memories produce hallucinatory associations. Also, the unevenness of patterns recall (which Hopfield, Feinstein & Palmer described as the “50 % of all stimuli remind me of sex” effect)  produce obsessions. It was observed in small networks that reverse learning equalized the patterns recall and supressed the spurious ones (Ref. 2). A testable prediction of Crick & Mitchison proposal is that a lack of REM sleep should produce pathological responses such as hallucinations, delusions and obsessions. As a biological implication they suggested that the echidna’s large neocortex is related with lacking of REM sleep (Ref. 3): an overloading of its neural net could be tackled by making the net bigger.


The Crick & Mitchison program  (I mean: to apply simple neural networks models in understanding dreams) lost appeal because the proposed implications were not confirmed by experimental and clinical studies (Ref 4, 5 and 6). However I think that  developments in the physical theory of spin glasses (Ref 7), mathematically equivalent to the kind of neural networks that inspired the Crick & Mitchison hypothesis, can be used to suggest a mechanism that avoid the waking hallucinations and, as by product, qualitatively explain some aspects of the dream-work called condensation by Sigmund Freud (Ref. 8). I assume in the following that conscious states are in correspondence with stable states in the phase space of a spin-glass model of a neural network. Short living states reached by the system are considered here as an unconscious state that will evolve in a brief time to a stable one. It was shown theoretically that for a network with infinite number of neurons and with random noise effects (as a temperature T) the spurious states observed in the numerical studies (Ref 2), are not random but correspond to well-defined mixtures of a number n of the stored patterns (Ref 9). The n=1 states, corresponding to the learned patterns, are stable in a broad range of temperatures. The even n mixtures are always unstable. For odd n mixtures will there be a critical temperature below which the n-mixture states are stable. This critical temperature is smaller when the odd n is bigger. Then can be conclude that with a sufficiently high temperature (high levels of random noise in the system) will not be stable all the mixture states with n>1 (spurious patterns) being stable only the learned patterns. The previous result implies that, if Crick & Mitchison’s interpretations of spurious patterns are right, in waking time when there are high levels of random noise in the neural network produced by sensorial inputs, the spurious memories are not reached as stable (consciouss) states preventing hallucinations and delusions. The pathological effect of the spurious patterns was predicted by Crick & Mitchison in the less realistic approach of zero temperature used in the original Hopfield model (Ref. 2). However when the sensorial input is strongly reduced, as in REM sleep, those spurious states can be accesed as dream imaginery. This spurious mixture patterns recalled the phenomena that Freud described from its clinical and personal experience as condensation processes. More specifically he defined them as cumulative images (Sammelbild) and mixture persons (Mischperson): “From its own dreams each of us knows those products… I can compose a person features from one or another, or giving the figure of one and thinking about her in the dream with the name of other, or I can perform visually a person, but putting her in a situation that happens with other.” (Ref 8). A particularly simple example are the nonsense words appearing in dreams composed by proper ones. From Freud’s studies (Ref 8) due to its rather subjective methodology of interpretation is very difficult to quantify the number of words involved in the construction of those nonsense words. However is possible that carefully designed experimental studies helps to show if this words are created from a odd or even number of proper words.


It is expected, from the point of view stated here, that during extreme isolation and sensory deprivation, due to the lowering of random noise as sensory inputs, the spurious patterns can be more accesible by consciouss (stable states) process giving place to hallucinations. This implications are consistent with the general results of the seminal work of Bexton, Heron & Scott on sensory deprivation (Ref 10) and many further studies.  The model discussed here is too simple in order to provide a complete explanation of the dream psychological processes. However it provides a qualitative explanation of the relation of random noise (temperature) and the dream imaginery that could be further tested in properly designed experiments and confronted with more biological based theoretical neural network models. The Crick & Mitchinson’s predictions were not observed probably due to the temperature effect not considered by them. I hope that their way of thinking about dreams, including more realistic theoretical considerations, still have something to offer in the research about psychology of dreams.          




  1. F. Crick & G. Mitchison (1983). The function of dream sleep. Nature 304, 111-114.


  1. J.J. Hopfield, D.I. Feinstein & R.G. Palmer (1983) ‘Unlearning’ has a stabilizing effect in collective memories. Nature 304, 158-159.


  1. Allison, T., H. Van Twyver and W. R. Goff (1972). Electrophysiological studies of the echidna, Tachyglossus aculeatus. I. Waking and sleep. Arch. ital. Biol, 110,145-184.


  1. P.Lavie et al (1984). Localized pontine lesion: Nearly total absence of REM sleep. Neurology. 34, 118-120.


  1. R.P. Vertes, K.E. Eastman (2000). "The case against memory consolidation in REM sleep," Behavioral and Brain Sciences, 23:867-876.


  1. Siegel, J.M. (1999). The evolution of REM sleep. In Handbook of Behavioral State Control., Lydic, R and Baghdoyan (Eds.), pp 87-100, CRC Press, Boca Raton.


  1. ACC Coolen (2000). Statistical mechanics of Recurrent Neural Networks. I. Statics cond-mat/0006010.


  1. Freud, S. (1900 / 1950) The Interpretation of Dreams. Random House.

  2. Amit DJ, Gutfreund H & Sompolinsky H. (1985) Phys Rev A. 32, 1007-1018


  1. Bexton, W.H., Heron, W., & Scott, T.H. (1954). Effects of Decreased Variation in the Sensory Environment. Canadian Journal of Psychology (8) 70-76.