{"title": "Cholinergic suppression of transmission may allow combined associative memory function and self-organization in the neocortex", "book": "Advances in Neural Information Processing Systems", "page_first": 131, "page_last": 137, "abstract": null, "full_text": "Cholinergic suppression of transmission may \n\nallow combined associative memory function and \n\nself-organization in the neocortex. \n\nMichael E. Hasselmo and Milos Cekic \n\nDepartment of Psychology and Program in Neurosciences, \nHarvard University, 33 Kirkland St., Cambridge, MA 02138 \n\nhasselmo@katIa.harvard.edu \n\nAbstract \n\nSelective  suppression  of  transmission  at  feedback  synapses  during \nlearning  is  proposed  as  a  mechanism  for  combining  associative feed(cid:173)\nback  with  self-organization  of feed forward  synapses.  Experimental \ndata  demonstrates  cholinergic  suppression  of synaptic  transmission  in \nlayer I (feedback synapses), and a lack of suppression in layer IV (feed(cid:173)\nforward synapses).  A network with this feature uses local rules to learn \nmappings  which  are not  linearly  separable.  During  learning,  sensory \nstimuli  and  desired  response  are  simultaneously  presented  as  input. \nFeedforward connections  form  self-organized representations  of input, \nwhile suppressed feedback connections learn  the  transpose of feedfor(cid:173)\nward connectivity. During recall, suppression is removed, sensory input \nactivates  the  self-organized  representation,  and  activity  generates  the \nlearned response. \n\n1 \n\nINTRODUCTION \n\nThe synaptic connections in most models of the cortex can be defined as either associative \nor self-organizing on the basis of a single feature:  the relative infl uence of modifiable syn(cid:173)\napses on post-synaptic activity during learning (figure 1).  In associative memories, post(cid:173)\nsynaptic  activity during  learning  is  determined  by nonmodifiable afferent input connec(cid:173)\ntions,  with no change in  the storage due to  synaptic  transmission  at modifiable synapses \n(Anderson,  1983;  McNaughton  and  Morris,  1987).  In  self-organization,  post-synaptic \nactivity is predominantly influenced by the modifiable synapses, such that modification of \nsynapses  influences  subsequent learning  (Von  der  Malsburg,  1973;  Miller et al.,  1990). \nModels of cortical  function  must combine the capacity  to  form  new  representations and \nstore  associations  between  these representations.  Networks  combining  self-organization \nand associative memory function can learn complex mapping functions with more biolog(cid:173)\nically plausible learning rules (Hecht-Nielsen,  1987; Carpenter et al.,  1991;  Dayan et at., \n\n\f132 \n\nM. E. HASSELMO, M. CEKIC \n\n1995), but must control the influence of feedback associative connections on self-organi(cid:173)\nzation.  Some  networks  use  special  activation  dynamics  which  prevent  feedback  from \ninfluencing activity unless it coincides with  feedforward activity  (Carpenter et al.,  1991). \nA  new  network  alternately  shuts  off feedforward  and  feedback  synaptic  transmission \n(Dayan et al., 1995). \n\nA. \n\nSelf-organizing \n\nc. \n\nAfferent \n\nSelf-organizing \nfeedforward \n\nAssociative \nfeedback \n\nFigure 1 - Defining  characteristics of self-organization and associative memory.  A.  At \nself-organizing  synapses,  post-synaptic  activity  during  learning  depends  predominantly \nupon transmission at the modifiable synapses.  B. At synapses mediating associative mem(cid:173)\nory function, post-synaptic activity during learning does not depend primarily on the mod(cid:173)\nifiable  synapses,  but  is  predominantly  influenced  by  separate  afferent  input.  C.  Self(cid:173)\norganization and  associative memory  function  can  be combined  if associative  feedback \nsynapses are selectively suppressed during learning but not recall. \n\nHere  we present a  model  using  selective suppression of feedback  synaptic  transmission \nduring  learning  to  allow  simultaneous  self-organization  and  association  between  two \nregions.  Previous  experiments  show  that  the  neuromodulator  acetylcholine  selectively \nsuppresses synaptic transmission within the olfactory cortex (Hasselmo and Bower, 1992; \n1993) and hippocampus (Hasselmo and Schnell, 1994).  If the model is valid for neocorti(cid:173)\ncal  structures,  cholinergic  suppression should be  stronger for  feedback  but  not  feedfor(cid:173)\nward synapses. Here we review experimental data (Hasselmo and Cekic, 1996) comparing \ncholinergic  suppression  of synaptic  transmission  in  layers  with  predominantly  feedfor(cid:173)\nward or feedback synapses. \n\n2.  BRAIN SLICE PHYSIOLOGY \n\nAs shown in  Figure 2, we utilized brain  slice preparations of the rat somatosensory neo(cid:173)\ncortex to investigate whether cholinergic suppression of synaptic transmission is selective \nfor feedback but not feedforward  synaptic connections.  This was possible because feed(cid:173)\nforward and feedback connections show different patterns of termination in neocortex.  As \nshown  in  Figure  2,  Layer  I  contains  primarily  feedback  synapses  from  other  cortical \nregions  (Cauller and  Connors,  1994), whereas  layer  IV  contains primarily afferent syn(cid:173)\napses from  the thalamus and feedforward  synapses from  more primary neocortical struc(cid:173)\ntures (Van  Essen and Maunsell,  1983).  Using previously developed  techniques (Cauller \nand  Connors, 1994; Li and Cauller, 1995) for testing of the predominantly feedback con(cid:173)\nnections in layer I, we stimulated layer I and recorded in layer I (a cut prevented spread of \n\n\fCholinergic  Suppression  of Transmission  in  the  Neocortex \n\n133 \n\nactivity  from  layers II  and  III).  For testing  the predominantly  feedforward  connections \nterminating  in  layer  IV,  we  elicited  synaptic  potentials  by  stimulating  the  white  matter \ndeep to layer VI and recorded in layer IV.  We tested suppression by measuring the change \nin  height of synaptic potentials during  perfusion  of the  cholinergic agonist carbachol at \nlOOJ,1M.  Figure 3 shows that perfusion of carbachol caused much stronger suppression of \nsynaptic transmission in layer I as compared to layer IV (Hasselmo and Cekic, 1996), sug(cid:173)\ngesting that cholinergic suppression of transmission is selective for feedback synapses and \nnot for feedforward synapses. \n\nI  ~~ I \n\nll-ill \n\nIV \n\n.1  Foedback \n\nll-ill \n\nIV \n\nV-VI \n\nV-VI \n\nRegion 1 \n\nRegion 2 \n\nI \n\nLayer IV \nrecording \n\nWhite matter  / \n\nstimulation 1/ \n\nFigure 2.  A. Brain slice preparation of somatosensory cortex showing location of stimula(cid:173)\ntion  and recording  electrodes  for  testing  suppression  of synaptic  transmission  in  layer  I \nand in layer IV.  Experiment based on procedures developed by Cauller (Cauller and Con(cid:173)\nnors,  1994;  Li and Cauller,  1995).  B.  Anatomical pattern  of feedforward and  feedback \nconnectivity within cortical structures (based on Van Essen and Maunsell, 1983). \nFeedforward -layer IV \n\nControl \n\nCarbachol (1 OOJlM) \n\nWash \n\nFeedback - layer I \n\nI~ -0  \n\n5ms \n\n'!oi' \n\nControl \n\nCarbachol (1 OOJlM) \n\nWash \n\nFigure 3 - Suppression of transmission in somatosensory neocortex.  Top: Synaptic poten(cid:173)\ntials  recorded in  layer IV  (where  feedforward  and  afferent synapses predominate) show \nlittle effect of l00J.tM carbachol.  Bottom: Synaptic potentials recorded in layer I (where \nfeedback synapses predominate) show suppression in the presence of lOOJ,1M carbachol. \n\n\f134 \n\nM. E. HASSELMO, M. CEKIC \n\n3.  COMPUTATIONAL MODELING \n\nThese experimental results supported the use of selective suppression in a computational \nmodel (Hasselmo and Cekic, 1996) with self-organization in its feedforward synaptic con(cid:173)\nnections and associative memory function in its feedback synaptic connections (Figs 1 and \n4).  The proposed network uses local, Hebb-type learning rules supported by evidence on \nthe physiology of long-tenn potentiation in the hippocampus (Gustafsson and Wigstrom, \n1986).  The learning rule for each set of connections in the network takes the fonn: \n\ntlWS:'Y)  = 11  (a?) - 9(Y\u00bb  g (ar\u00bb \n\nWhere W(x. Y)  designates the connections from region x to  region  y,  9 is the threshold of \nsynaptic modification in region  y, 11  is  the rate of modification, and the output function is \ng(a;.(x~ = [tanh(~(x) - J.1(x~]+ where []+  represents the constraint to positive values only. \nFeedforward connections (Wi/x<y\u00bb  have self-organizing properties, while feedback con(cid:173)\nnections (Wir>=Y~ have associative memory properties.  This difference depends entirely \nupon  the  selective  suppression  of feedback  synapses  during  learning,  which  is  imple(cid:173)\nmented in the activation rule in the form  (I-c).  For the entire network, the activation rule \ntakes the fonn: \n\nM  II(X) \n\nN  II(X) \n\nII(Y) \n\na?)  = A?) + 2,  2, Wi~<Y) g (a~x\u00bb  +  2,  2,  (1- c) Wi~~Y) g (a~x\u00bb  - 2, Hi~) (g (af\u00bb) \n\nx=lk=l \n\nx=lk=l \n\nk=l \n\nwhere a;.(y) represents the activity of each of the n(y) neurons in region y, ~ (x)  is the activ(cid:173)\nity of each of the n(x)  neurons in  other regions x,  M is the total number of regions provid(cid:173)\ning feedforward input, N is the total number of regions providing feedback input, Aj(y) is \nthe input pattern to  region  y,  H(Y)  represents the inhibition between neurons in region  y, \nand (1  - c) represents the suppression of synaptic transmission.  During learning, c takes a \nvalue between 0 and 1.  During recall, suppression is removed, c = O.  In this network, syn(cid:173)\napses  (W)  between regions only  take positive values,  reflecting  the  fact that long-range \nconnections between cortical regions consist of excitatory synapses arising from  pyrami(cid:173)\ndal cells.  Thus, inhibition mediated by the local inhibitory interneurons within a region is \nrepresented by a separate inhibitory connectivity matrix H. \n\nAfter each step of learning, the total weight of synaptic connections is nonnalized pre-syn(cid:173)\naptically for each neuron j in each region: \n\nWij (t+l)  = [Wij(t)  + l1W;j(t)]I(  .i  [Wij(t)  +l1Wij (t)] 2) \n\n~--------------\n\n1= 1 \n\nSynaptic  weights are then  normalized post-synaptically for  each neuron i in each region \n(replacing  i  with j  in the sum in  the denominator in  equation 3).  This  nonnalization  of \nsynaptic strength  represents slower cellular mechanisms which redistribute pre and post(cid:173)\nsynaptic resources for maintaining synapses depending upon local influences. \n\nIn  these simulations, both the sensory input stimuli and the desired output response to be \nlearned are presented as afferent input to  the neurons in region  1.  Most networks  using \nerror-based  learning  rules  consist  of feedforward  architectures  with  separate  layers  of \ninput and output units.  One can imagine this network as an  auto-encoder network folded \nback on itself, with both input and output units in region 1, and hidden units in region 2. \n\n\fCholinergic  Suppression  of Transmission  in  the  Neocortex \n\n135 \n\nAs an  example of its functional properties, the network presented here was trained on the \nXOR problem.  The XOR problem has previously been used as an example of the capabil(cid:173)\nity of error based training schemes for solving problems which are not linearly separable. \nThe specific characteristics of the network and patterns used for this simulation are shown \nin figure 4.  The two logical states of each component of the XOR problem are represented \nby two separate units (designated on or off in figures 4 and 5), ensuring that activation of \nthe  network is equal  for  each input condition.  The problem  has  the appearance of two \nXOR problems with inverse logical states being solved simultaneously. \n\nAs shown in  figure 4, the input and desired output of the network are presented simulta(cid:173)\nneously during  learning to  region  1.  The six neurons  in  region  1 project along  feedfor(cid:173)\nward connections to four neurons in region 2, the hidden units of the network.  These four \nneurons project along feedback connections  to the six neurons in  region  1.  All  connec(cid:173)\ntions take random initial weights.  During learning, the feedforward connections undergo \nself-organization  which  ultimately  causes  the  hidden  units  to  become  feature  detectors \nresponding to each of the four patterns of input to region 1.  Thus, the rows of the feedfor(cid:173)\nward synaptic connectivity matrix gradually take the form of the individual input patterns. \n\nSTIMULUS \n\nRESPONSE \nyes  no \n\non  off  on  off \n\n1.  oeeo  eo \n2.  oeoe  oe  Afferent \neooe  eo  input \n4.  9  9  9 \n\n3. \n\n\"'-Region 1 \n\nFigure 4 - Network for  learning the XOR problem, with 6  units in  region  1 and 4 units in \nregion 2.  Four different patterns of afferent input are presented successively to  region 1. \nThe input stimuli of the XOR problem are represented by the four units on the left, and the \ndesired  output designation  of XOR  or  not-XOR  is  represented  by  the  two  units  on  the \nright.  The XOR problem has four basic states:  on-off and off-on on  the input is catego(cid:173)\nrized by yes on the output, while on-on and off-off on the input is categorized by no on the \noutput. \n\nModulation  is  applied  during  learning  in  the  form  of selective  suppression  of synaptic \ntransmission along feedback connections (this suppression need not be complete), giving \nthese  connections  associative  memory  function.  Hebbian  synaptic  modification  causes \nthese connections  to  link each of the  feature detecting hidden  units in  region 2  with  the \ncells in region 1 activated by the pattern to which the hidden unit responds.  Gradually, the \nfeedback synaptic connectivity matrix becomes the transpose of the feedforward connec(cid:173)\ntivity  matrix.  (parameters used in simulation:  Aj(l) = 0 or  I, h = 2.0, q(l) = 0.5, q(2) = \n0.6, (1) = 0.2, (2) = 0.5, c = 1.0 and Hik(2) = 0.6).  Function was similar and convergence \nwas obtained more rapidly with c = 0.5.  Feedback synaptic transmission prevented con-\n\n\f136 \n\nM. E.  HASSELMO. M. CEKIC \n\nvergence during learning when c = 0.367). \n\nDuring recall, modulation of synaptic transmission is removed, and the various input stim(cid:173)\nuli  of the XOR problem  are presented to region  1 without the corresponding output pat(cid:173)\ntern.  Activity  spreads along  the self-organized  feedforward  connections  to activate  the \nspecific  hidden  layer  unit  responding  to  that  pattern.  Activity  then  spreads  back along \nfeedback  connections  from  that particular unit  to activate the desired  output  units.  The \nactivity in the two regions settles into a final pattern of recall.  Figure 5 shows the settled \nrecall of the network at different stages of learning.  It can  be seen that the network ini(cid:173)\ntially may show little recall activity, or erroneous recall activity, but after several cycles of \nlearning, the network settles into the proper response to each of the XOR problem states. \nConvergence during learning and recall have been obtained with other problems, includ(cid:173)\ning  recognition of whether on  units were on  the left or right, symmetry of on  units, and \nnumber of on units.  In addition, larger scale problems involving multiple feedforward and \nfeedback layers have been shown to converge. \n\n1. \n\n3. \n\n~ \n\u00a7L \n3 \n\n4. \n\n2. \n\noff  off  no \n\n1 \n\n0 \n\n.... \n\non  on  no \n\noff  on  yes \n\non  off  yes \n\n--- - --\n-- -\n-- -\n-\n--\n- -\n-\n_ .. --- - \u00b7 - .-- . - - \u00b7 - -\u00b7 \u00b7  -\nR  ----- - ------- - - -- -\u00b7 \u00b7 \n--\n.  -----. \u00b7  \u00b7 \n:-=.:: - \u00b7 - ---- - - \u00b7 - -\u00b7 \u00b7 \n- \u00b7 - -- - - \u00b7 - --\u00b7  - ---- - \u00b7  \u00b7 \n.  11----. \u00b7  \u00b7 \n-- ---- \u00b7 - --- - - \u00b7 - -. \u00b7  - ----- -\n::=::  - -- --- - - \u00b7 -- -. \u00b7  - 11---- -\n--- --\u00b7  \u00b7 \n- - =-- - - -- -. \u00b7  - ---- -\n:11-==  - - .-- - -\n.- - - - -- \u00b7  - -----\n----- - - .- - - .:  \u2022\u2022\u2022  - 11:11::  -\n-- -- - -- \u00b7  - ----- -\n-\n----- - - -.- -- - -- \u00b7  - ---- - \u00b7 \n-.  -\n----- -\n----- - - -.--- - -- - - --.- \u2022 \n>  ----- - - ---- \u2022 .: - - .~.: - -\n-\n-\n-\n-:==~:  - - ----\n- --- - - --\n- --\n:11 := -- - --\n-\n- --\n<  -. --\n-- \u2022  .:  - \u2022 \u2022 \n~.  =-=: -- -\n- -\n- - - =: \n- - -\n\u2022 \u2022 - -\n--- - -\n- -\u2022 \n-\n- - ---- - --\n-<  I. - -\n-- -\n-\n- -\n-- - - --- - - =:  --- -\n- -- -\n-- - - -- - - -\n\u2022 \u2022 - - \u2022 .:  - \u2022 \u2022  -\n--- - - =: \n-- - -\n- -- -\n\u2022 \u2022 - - \u2022 .:  - \u2022 \u2022  -\n--- -\n:=.  - - --- - -\n- -\n\u2022\u2022 - -\n-\n-\n--\n--- -\n= -- - - --\n--\n\u2022\u2022 - -\n--- -\n-\n-\n- - - - --\n- -\n- -= -\n- - --- - - --\n- -\n-- - - --- - - --\n- - -\n\u2022 :. - - \u2022 .: \n==  - -\n- \u2022 \u2022  -\n-- - -\n= -- - - =:  --= -\n.- - - -\n\u2022\u2022  - -\n-\n-\n- -\n- - -\n- - -.- -- - -- --- -\n- -\n:==  - -\n26  -- - - --- - - --\n- -- -\n\n\u2022 \n\n-\n\n:;0  I \nen \n0 \n~. \n3  ~ \n\nt \n\nRegion 1  Region 2 \n\nFigure 5 - Output neuronal activity in  the network shown at different learning steps.  The \nfour input patterns are shown at top.  Below these are degraded patterns presented during \nrecall, missing the response components of the input pattern.  The output of the 6 region 1 \nunits and the 4 region 2 units are shown at each stage of learning.  As learning progresses, \ngradually one region 2 unit starts to respond selectively to each input pattern, and the cor(cid:173)\nrect output unit becomes active in response to the degraded input.  Note that as learning \nprogresses the response to pattern 4 changes gradually from incorrect (yes) to correct (no). \n\n\fCholinergic  Suppression  of Transmission  in  the  Neocortex \n\n137 \n\nReferences \n\nAnderson, 1.A.  (1983)  Cognitive and  psychological computation  with  neural  models. \nIEEE Trans. Systems, Man, Cybem.  SMC-13,799-815. \nCarpenter, G.A.,  Grossberg,  S.  and Reynolds, 1.H.  (1991)  ARTMAP:  Supervised real(cid:173)\ntime learning and classification of nonstationary data by a self-organizing neural network. \nNeural Networks  4:  565-588. \nCauller,  LJ.  and  Connors,  B.W.  (1994)  Synaptic  physiology  of horizontal  afferents  to \nlayer I in slices of rat SI neocortex.  1. Neurosci.  14: 751-762. \nDayan,  P.,  Hinton,  G.E.,  Neal,  RM.  and  Zemel,  RS.  (1995)  The  Helmholtz  machine. \nNeural Computation. \nGustafsson, B. and Wigstrom, H.  (1988) Physiological mechanisms underlying long-term \npotentiation.  Trends Neurosci.  11:  156-162. \nHasselmo,  M.E.  (1993)  Acetylcholine  and  learning  in  a  cortical  associative  memory. \nNeural Computation. 5(1}: 32-44. \nHasselmo M.E. and Bower 1.M.  (1992)  Cholinergic suppression specific to intrinsic not \nafferent fiber synapses in rat piriform (olfactory) cortex.  1. Neurophysiol.  67:  1222-1229. \nHasselmo,  M.E.  and  Bower, 1.M.  (1993)  Acetylcholine and  Memory.  Trends Neurosci. \n26: 218-222. \nHasselmo,  M.E.  and  Cekic,  M.  (1996)  Suppression of synaptic  transmission  may allow \ncombination of associative feedback  and  self-organizing  feed forward  connections in  the \nneocortex.  Behav. Brain Res. in press. \nHasselmo M.E., Anderson B.P. and Bower 1.M. (1992) Cholinergic modulation of cortical \nassociative memory function.  1. Neurophysiol.  67:  1230-1246. \nHasselmo M.E. and Schnell, E.  (1994) Laminar selectivity of the cholinergic suppression \nof synaptic  transmission  in  rat  hippocampal  region  CAl:  Computational  modeling  and \nbrain slice physiology.  1. Neurosci.  15: 3898-3914. \nHecht-Nielsen, R  (1987) Counterpropagation networks.  Applied Optics  26: 4979-4984. \nLi,  H.  and Cauller,  L.l.  (1995)  Acetylcholine  modulation  of excitatory  synaptic  inputs \nfrom layer I to the superficial layers of rat somatosensory neocortex in vitro.  Soc. Neuro(cid:173)\nsci. Abstr. 21: 68. \nLinsker, R  (1988)  Self-organization in a perceptual network.  Computer 21:  105-117. \nMcNaughton  B.L.  and  Morris RG.M.  (1987)  Hippocampal synaptic enhancement and \ninformation storage within a distributed memory system. Trends in Neurosci.  10:408-415. \nMiller,  K.D.,  Keller, 1.B.  and  Stryker,  M.P.  (1989)  Ocular dominance column develop(cid:173)\nment Analysis and simulation.  Science 245:  605-615. \nvan  Essen,  D.C.  and Maunsell, 1.H.R.  (1983)  Heirarchical  organization and  functional \nstreams in the visual cortex.  Trends Neurosci.  6:  370-375. \nvon der Malsburg, C. (1973)  Self-organization of orientation sensitive cells in the striate \ncortex.  Kybemetik  14:  85-100. \n\n\f", "award": [], "sourceid": 1105, "authors": [{"given_name": "Michael", "family_name": "Hasselmo", "institution": null}, {"given_name": "Milos", "family_name": "Cekic", "institution": null}]}