{"title": "Models Wanted: Must Fit Dimensions of Sleep and Dreaming", "book": "Advances in Neural Information Processing Systems", "page_first": 3, "page_last": 10, "abstract": null, "full_text": "MODELS  WANTED:  MUST FIT DIMENSIONS \n\nOF  SLEEP  AND  DREAMING* \n\nJ. Allan Hohson,  Adam N.  Mamelakt  and  Jeffrey  P.  Suttont \n\nLaboratory of Neurophysiology and Department of Psychiatry \n\nHarvard Medical  School \n\n74  Fenwood  Road,  Boston,  MA  02115 \n\nAbstract \n\nDuring waking and sleep, the brain and mind undergo a  tightly linked and \nprecisely  specified  set  of changes  in state.  At  the  level  of neurons,  this \nprocess  has  been  modeled  by  variations  of Volterra-Lotka  equations  for \ncyclic fluctuations of brainstem cell populations.  However, neural network \nmodels based upon rapidly developing knowledge ofthe specific population \nconnectivities  and  their  differential responses  to drugs  have  not  yet  been \ndeveloped.  Furthermore, only  the  most  preliminary  attempts  have  been \nmade  to model across states.  Some  of our own attempts  to link rapid eye \nmovement (REM) sleep neurophysiology and dream cognition using neural \nnetwork approaches  are summarized in this  paper. \n\n1 \n\nINTRODUCTION \n\nNew  models are  needed  to test  the closely linked  neurophysiological and cognitive \ntheories that are emerging from recent scientific studies of sleep and dreaming.  This \nsection describes four separate but related levels of analysis at which modeling may \n\n\u00b7Based, in part, upon an invited address by J.A.H. at NIPS, Denver, Dec.  2 1991  and, \nin part,  upon a  review paper by J.P.S.,  A.N.M.  and  J.A.H.  published in  the  P.ychiatric \nAnnal \u2022. \n\nt Currently in the Department of Neurosurgery, University of California,  San Francisco, \n\nCA  94143 \n\n: Also in the  Center for  Biological Information Processing,  Whitaker College, E25-201, \n\nMassachusetts Institute of Technology,  Cambridge, MA 02139 \n\n3 \n\n\f4 \n\nHobson,  Mamelak, and Sutton \n\nbe applied and outlines some of the desirable features of such models in terms of the \nburgeoning data of sleep and dream science.  In the subsequent sections, we  review \nour own preliminary efforts  to develop models at some of the levels discussed. \n\n1.1  THE INDIVIDUAL NEURON \n\nExisting models  or  \"neuromines\"  faithfully  represent  membrane properties but ig(cid:173)\nnore the dynamic biochemical changes that change neural excitability over the long \nterm.  This  is  particularly  important  in  the  modeling  of state  control  where  the \ncrucial neurons appear to act  more like hormone pumps than like simple electrical \ntransducers.  Put succinctly, we need models that consider the biochemical or \"wet\" \naspects of nerve cells,  as well as  the  \"dry\"  or electrical aspects (cf.  McKenna et  al., \nin press). \n\n1.2  NEURAL POPULATION INTERACTIONS \n\nTo mimic the changes in excitability of the modulatory neurons which control sleep \nand dreaming, new models are needed which incorporate both the engineering prin(cid:173)\nciples  of oscillators  and  the  biological  principles  of time-keeping.  The latter prin(cid:173)\nciple  is  especially  relevant  in  determining  the  dramatica.lly  variable  long  period \ntime-constants that are  observed  within and  across  species.  For example, we  need \nto  equip  population  models  borrowed  from  field  biology  (McCarley  and  Hobson, \n1975)  with specialized  properties of \"wet\" neurons  as mentioned in section 1.1. \n\n1.3  COGNITIVE  CONSEQUENCES  OF MODULATION  OF \n\nNEURAL  NETWORKS \n\nTo understand the state-dependent changes in cognition, such as those that distin(cid:173)\nguish waking and dreaming,  a  potentially fruitful  approach is  to mimic  the known \neffects  of neuromod ulation  and  examine  the  information  processing  properties  of \nneural  networks.  For  example,  if the  input-output  fidelity  of networks  can  be  al(cid:173)\ntered by changing their mode  (see  Sutton  et  al.,  this  volume),  we  might  be  better \nable  to  understand  the  changes  in  both  instantaneous  associative  properties  and \nlong  term plasticity  alterations  that occur in sleep  and dreaming.  We  might  thus \ntrap the  brain-mind  into  revealing its  rules  for  making  moment-to-moment  cross(cid:173)\ncorrelations  of its  data  and  for  changing  the  content  and  status  of its  storage  in \nmemory. \n\n1.4  STATE-DEPENDENT CHANGES IN  COGNITION \n\nAt  the  highest  level  of analysis,  psychological  data,  even  that  obtained  from  the \nintrospection of waking and dreaming subjects, need to be more creatively reduced \nwith a  view  to modeling the dramatic alterations that occur with changes in brain \nstate.  As  an  example,  consider  the  instability  of orientation  of dreaming,  where \ntimes,  places,  persons  and  actions  change  without  notice.  Short  of mastering  the \nthorny problem  of generating narrative text from  a  data base,  and  thus synthesiz(cid:173)\ning  an  artificial  dream,  we  need  to  formulate  rules  and  measures  of categorizing \nconstancy and transformations  (Sutton and Hobson,  1991).  Such an approach is  a \n\n\fModels Wanted: Must Fit Dimensions of Sleep  and Dreaming \n\n5 \n\nmeans  of further  refining  the algorithms  of cognition itself,  an effort  which is  now \nlimited to simple activation models that cannot  change mode. \n\nAn important characteristic of the set of new models that are proposed is that each \nlevel informs, and is informed by,  the other levels.  This nested, interlocking feature \nis  represented in figure  1.  It should be noted that any erroneous assumptions made \nat  level  1  will  have  effects  at  levels  2  and  3  and  these  will,  in  turn,  impede  our \ncapacity  to integrate levels  3  and 4.  Level 4  models  can and  should  thus  proceed \nwith a degree ofindependence from levels 1, 2 and 3.  Proceeding from level 1 upward \nis the  \"bottom-up\" approach, while  proceeding from level 4 downward is  the  \"top(cid:173)\ndown\"  approach.  We like  to think it might be possible  to take both approaches in \nour work while  according equal respect  to each. \n\nLEVEL \n\nSCHEMA \n\nFEATURES \n\nIV  COGNITIVE \n\nSTATES \n\n(eg.  dream plot \nsequences) \n\nA-.B \n\n<C--+D \nE--+F \n\nvariable associative \nand learning states \n\nIII  MODULATION \nOF NETWORKS \n(eg.  hippocampus, \n\ncortex) \n\n~~ ,~ \n~  (~2~ \n\nmodulation of \n1-0 processing \n\nt?-(7  ~ \n\nvariable time-\n\nconstant oscillator \n\nwet  hormonal \n\naspects \n\nIl  NEURAL \n\nPOPULATIONS \n\n(eg.  pontine \nbrainstem) \n\nI  SINGLE \nNEURONS \n(eg.  NE,  5HT, \nACh neurons) \n\nFigure  1:  Four  levels  at  which  modeling  innovations  are  needed  to  provide  more \nrealistic  simulations  of brain-mind states  such as  waking  and  dreaming.  See  text \nfor  discussion. \n\n\f6 \n\nHobson,  Marnelak, and Sutton \n\n2  STATES  OF  WAKING AND  SLEEPING \n\nThe states  of waking and  sleeping,  including  REM  and  non-REM  (NREM)  sleep, \nhave  characteristic  behavioral,  neuronal,  polygraphic  and  psychological  features \nthat  span  all  four  levels.  These  properties  are  summarized  in  figures  2  and  3. \nChanges  occurring  within  and  between  different  levels  are  affected  by  the  sleep(cid:173)\nwake or circadian cycle  and by the relative shifts in brain chemistry. \n\nA \n\nWAKE \n\nNREM  SLEEP  REM  SLEEP \n\nE~I~-------- -------------1----------\nEEGI:_==:::==::::;::: 1~~':':.fJd,~ 1===::::::::: \nEOG ~ ~1--..\"l'-'o...J \n\nSensation  and \n\nVivid, \n\nPerception  Externally Generated \n\nDull  or  Absent \n\nVivid, \n\nIntemoUy Generated \n\nTlJougllf \n\nMovement \n\nLogical \n\nProgressive \n\nContiooous \nVoluntary \n\nLogical \n\nPerseverotive \n\nEpisodic \nInvoluntary \n\nIllogical \nBizarre \n\nCommanded \ntM  Inhibited \n\nB \n\nc \n\nD \n\no \n\no \n\n, \n\n, \n\nt \n\nTime  (hours) \n\n_\n\n__ ...,(1  - - l l -\n\n_ \n\n0_  \n\n_ \n\n..,[) \n\n_ \n\nMl \n\n_ - - - .,-\n-- - - -\n\nTime  (lJours) \n\nFigure  2:  (a)  States of waking and  NREM and  REM sleeping in  humans.  Charac(cid:173)\nteristic behavioral, polygraphic and psychological features are shown for  each state. \n(b)  Ultradian  sleep  cycle  of NREM  and  REM  sleep  shown in  detailed  sleep-stage \ngraphs  of 3  subjects.  (c)  REM  sleep  periodograms  of 15  subjects.  From  Hobson \nand Steriade  (1986),  with permission. \n\n\fModels Wanted:  Must Fit  Dimensions of Sleep  and Dreaming \n\n7 \n\n2.1  CIRCADIAN RHYTHMS \n\nThe  circadian  cycle  has  been  studied  mathematically  using  oscillator  and  other \nnon-linear dynamical models to capture features of sleep-wake rhythms (Moore-Ede \nand  Czeisler,  1984;  figure  2).  Shorter  (infradian)  and longer  (ultradian)  rhythms, \nrelative  to  the  circadian  rhythm,  have  also  been examined.  In general,  oscillators \nare  used  to couple  neural,  endocrine  and other pathways important in controlling \na  variety of functions,  such as  periods of rest and activity,  energy conservation and \nthermoregulation.  The  oscillators  can  be  sensitive  to  external  cues  or  zeitgebers, \nsuch as light and daily routines,  and there is  a  stong linkage  between the circadian \nclock and the  NREM-REM sleep oscillator. \n\n2.2  RECIPROCAL INTERACTION MODEL \n\nIn the 1970s, a brainstem oscillator became identified that was central to regulating \nsleeping  and  waking.  Discrete  cell  populations  in  the  pons  that  were  most  active \nduring waking, less active in NREM sleep and silent  during REM sleep were found \nto contain the monoamines norepinephrine  (NE)  and serotonin  (5HT).  Among the \nmany  cell  populations  that  became  active  during  REM  sleep,  but  were  generally \nquiescent otherwise,  were cells  associated with acetylcholine (ACh)  release. \n\nA \n\nC \n15 \n\nI \n\nBlillJ \n\nt-\n\no \n\n, \n\n\\ \n\n20 \n\n40 \n\n10 \n\n10 \n\n100 \n\no \n\n'vi \n\n4 \n\n3 \n\n2 \n\no \n\nFigure  3:  (a)  Reciprocal  interaction model  of REM  sleep  generation  showing  the \nstructural  interaction  between  cholinergic  and  monoaminergic  cell  populations. \nPlus  sign  implies  excitatory  influences;  minus  sign  implies  inhibitory  influences. \n(b)  Model  output  of the  cholinergic  unit  derived  from  Lotka-Volterra  equations. \n(c)  Histogram of the discharge rate from  a cholinergic related pontine cell  recorded \nover 12  normalized sleep-wake cycles.  Model cholinergic (solid line)  and monoamin(cid:173)\nergic  (dotted  line)  outputs.  (d)  N oradrenergic  discharge  rates  before  (S),  during \n(D)  and  following  (W)  a  REM  sleep  episode.  From  Hobson  and  Steriade  (1986), \nwith permission. \n\n\f8 \n\nHobson, Mamelak, and Sutton \n\nBy  making  a  variety  of  simplifying  assumptions,  McCarley  and  Hobson  (1975) \nwere able  to structurally and mathematically model  the oscillations between these \nmonoaminergic  and  cholinergic  cell  populations  (figure  3).  This  level  2  model \nconsists  of two  compartments,  one  being monoaminergic-inhibitory  and  the  other \ncholinergic-excitatory.  It is  based pupon the assumptions offield biology (VoIterra(cid:173)\nLotka) and of dry neuromines (level 3).  The excitation (inhibition) originating from \neach compartment influences the other and also feeds back on itself.  Numerous pre(cid:173)\ndictions  generated  by  the  model  have  been  verified  experimentally  (Hobson  and \nSteriade, 1986). \nBecause  the  neural  population  model  shown  in  figure  3  uses  the  limited  passive \nmembrane type of neuromine discussed in the introduction,  the resulting oscillator \nhas a time-constant in the millisecond range, not even close to the real range of min(cid:173)\nutes to hours that characterize the sleep-dream cycle  (figure  2).  As such, the model \nis  clearly  incapable  of realistically  representing the long-term  dynamic  properties \nthat characterize interacting neuromodulatory populations.  To surmount this limi(cid:173)\ntation, two modifications  are possible:  one is  to remodel the individual neuromines \nequipping them  with  mathematics describing up and  down  regulation  of receptors \nand intracellular biochemistry that results in long-term changes in synaptic efficacy \n(c/.  McKenna  et  al.,  in  press);  another  is  to  model  the  longer  time  constants  of \nthe sleep  cycle  in terms  of protein transport  times between the  two populations in \nbrainstems of realistically varying width (c/.  Hobson and Steriade,  1986). \n\n3  NEUROCOGNITIVE ASPECTS  OF WAKING, \n\nSLEEPING AND DREAMING \n\nSince  the  discovery  that  REM  sleep  is  correlated  with  dreaming,  significant  ad(cid:173)\nvances have been  made in  understanding both the  neural and  cognitive  processes \noccurring  in  different  states  of the  sleep-wake  cycle.  During  waking,  wherein  the \nbrain is  in  a  state of relative aminergic  dominance,  thought  content and  cognition \ndisplay consistency  and  continuity.  NREM  sleep  mentation is  typically  character(cid:173)\nized by ruminative thoughts void of perceptual vividness or emotional tone.  Within \nthis state, the aminergic and cholinergic systems are more evenly balanced than in \neither  the  wake  or  REM  sleep  states.  As  previously  noted,  REM  sleep  is  a  state \nassociated with relative cholinergic activation.  Its mental status manifestations in(cid:173)\nclude graphic, emotionally charged and formally bizarre images encompassing visual \nhallucinations  and delusions. \n\n3.1  ACTIVATION.SYNTHESIS  MODEL \n\nThe  activation-synthesis  hypothesis  (Hobson  and  McCarley,  1977)  was  the  first \naccount  of dream mentation based  on  the  neurophysiological  state  of REM  sleep. \nIt considered factors present at levels 3 and 4,  according to the scheme in section 1, \nand attempted to bridge  these two levels.  In the model,  cholinergic activation and \nreciprocal monoaminergic disinhibition of neural networks in REM sleep generated \nthe source of dream formation.  However, the details  of how  neural networks  might \nactually synthesize information in the REM  sleep state was  not  specified. \n\n\fModels Wanted: Must Fit  Dimensions of Sleep and Dreaming \n\n9 \n\n3.2  NEURAL  NETWORK MODELS \n\nSeveral neural network models  have subsequently been proposed  that also attempt \nto bridge levels 3 and 4 (for example, Crick and Mitchison, 1983).  Recently,  Mame(cid:173)\nlak and Hobson (1989)  have suggested a neurocognitive model of dream bizarreness \nthat  extends  the  activation-synthesis  hypothesis.  In  the  model,  the  monoaminer(cid:173)\ngic  withdrawal in sleep  relative to waking leads to a  decrease in the signal-to-noise \nratio  in  neural  networks  (figure  4).  When  this  is  coupled  with  phasic  choliner(cid:173)\ngic  excitation of the cortex,  via brainstem ponto-geniculo-occipital  (PGO)  cell fir(cid:173)\ning (figure 5),  cognitive information becomes altered  and  discontinuous.  A  central \npremise  of the  model is  that  the monoamines  and acetylcholine function  as  neuro(cid:173)\nmodulators,  which modify ongoing activity in networks, without actually supplying \nafferent input information. \nImplementation of the  Mamelak and  Hobson model  as  a  temporal sequencing  net(cid:173)\nwork  is  described  by Sutton  et  al.  in  this  volume.  Computer simulations  demon(cid:173)\nstrate how  changes  in modulation  similar  to some  monoaminergic  and  cholinergic \neffects can completely alter the way information is collectively sequenced within the \nsame  network.  This  occurs  even in the  absence  of plastic  changes  in  the  weights \nconnecting the artificial neurons.  Incorporating plasticity,  which generally involves \nneuromodulators such as  the monoamines, is  a logical  next  step.  This would build \nimportant  level  1  features  into  a  level  3-4  model  and  potentially  provide  useful \ninsight into some state-dependent learning operations. \n\n\"  \u2022\u2022 10 ... \n\n\u2022 \u00b7r  11111 1111111111 11111 111111' II 111111111111 \n.-10  I  \"'\" 1111  fill  11111111111111' \nII \n\n1'.-.'0 ... \n\n\u2022 \u00b7r - -____  --+-__  \n.\u00b710  I  II  I \n\nIII  1 I \n\n1 \n\nI \n\nB \n\n.r \n... r-----\"77\"'. __ ::::--~ \n~  II J I I  \n\nD . \n\n........ \n\n. __ '_'.\",.' ........ te\"\"_ \n\nFigure  4:  (a)  Monoaminergic  innervation  of the  brain is  widespread.  (b)  Plot  of \nthe  neuron  firing  probability as  a  function  of the relative membrane  potential for \nvarious values  of monoaminergic  modulation (parameterized by Q).  Higher  (lower) \nmodulation is correlated with smaller (larger) Q  values.  (c)  Neuron firing  when sub(cid:173)\njected to supra- and sub-threshold inputs of +10 mvand -10 mv, respectively, for \nQ  = 2 and Q  = 10.  (d) For a given input, the repertoire of network outputs generally \nincreases as  Q  increases.  From Mamelak and  Hobson  (1989),  with permission. \n\n\f10 \n\nHobson, Mamelak, and Sutton \n\nA \n\nB \n\nUnil' i  r ~ g) ii' iI' \n\nLGlk~ \nLGBi--' ~ \n\nFigure 5:  (a)  Cholinergic  input from  the bramstem  to the  thalamus  and  cortex is \nwidespread.  (b)  Unit  recordings  from  PGO  burst  cells  in  the  pons  are  correlated \nwith PGO waves recorded in the lateral geniculate bodies (LGB)  of the thalamus. \n\n4  CONCLUSION \n\nAfter discussing four levels at which new models are needed,  we have outlined some \npreliminary efforts at modeling states of waking and sleeping.  We suggest that this \narea of research is ripe for  the development of integrative models of brain and mind. \n\nAcknowledgements \n\nSupported by NIH grant MH 13,923, the HMS/MMHC Research & Education Fund, \nthe  Livingston,  Dupont-Warren and  McDonnell-Pew  Foundations,  DARPA  under \nONR contract NOOOl4-85-K-0124,  the Sloan Foundation and Whitaker College. \n\nReferences \n\nCrick F,  Mitchison G  (1983)  The function  of dream sleep.  Nature  304 111-114. \nHobson  JA,  McCarley  RW  (1977)  The  brain  as  a  dream-state  generator:  An \nactivation-synthesis hypothesis  of the dream process.  Am J  P.ych 134 1335-1368. \nHobson  JA,  Steriade  M  (1986)  Neuronal  basis  of behavioral  state  control. \nIn: \nMountcastle  VB  (ed)  Handbook  of Phy.iology  - The  NeMJou.  Syltem,  Vol  IV. \nBethesda:  Am Physiol Soc,  701-823. \nMamelak  AN,  Hobson  JA  (1989)  Dream  bizarrenes  as  the  cognitive  correlate  of \naltered neuronal behavior in REM  sleep.  J  Cog  Neuro.ci 1(3) 201-22. \nMcCarley  RW,  Hobson  JA  (1975)  Neuronal  excitability  over  the  sleep  cycle:  A \nstructural and mathematical model.  Science 189 58-60. \n\nMcKenna T,  Davis J,  Zornetzer  (eds)  In  press.  Single  Neuron  Computation.  San \nDiego,  Academic. \n\nMoore-Ede  Me,  Czeisler  CA  (eds)  (1984)  Mathematical  Model.  of the  Circadian \nSleep- Wake  Cycle.  New York:  Raven. \nSutton JP,  Hobson  (1991)  Graph theoretical  representation of dream  content  and \ndiscontinuity.  Sleep  Re.earch 20 164. \n\n\f", "award": [], "sourceid": 460, "authors": [{"given_name": "J.", "family_name": "Hobson", "institution": null}, {"given_name": "Adam", "family_name": "Mamelak", "institution": null}, {"given_name": "Jeffrey", "family_name": "Sutton", "institution": null}]}