{"title": "A model of the hippocampus combining self-organization and associative memory function", "book": "Advances in Neural Information Processing Systems", "page_first": 77, "page_last": 84, "abstract": "", "full_text": "A model of the hippocampus combining self(cid:173)\norganization and associative memory function. \n\nMichael E. Hasselmo, Eric Schnell \n\nJoshua Berke and Edi Barkai \n\nDept. of Psychology, Harvard University \n33 Kirkland St., Cambridge, MA 02138 \n\nhasselmo@katla.harvard.edu \n\nAbstract \n\nA model of the hippocampus is presented which forms rapid self -orga(cid:173)\nnized representations of input arriving via the perforant path, performs \nrecall of previous associations in region CA3, and performs comparison \nof this recall with afferent input in region CA 1.  This comparison drives \nfeedback regulation of cholinergic modulation to set appropriate \ndynamics for learning of new representations in region CA3 and CA 1. \nThe network responds to novel patterns with increased cholinergic mod(cid:173)\nulation, allowing storage of new self-organized representations, but \nresponds to familiar patterns with a decrease in acetylcholine,  allowing \nrecall based on previous representations.  This requires selectivity of the \ncholinergic suppression of synaptic transmission in stratum radiatum of \nregions CA3 and CAl, which has been demonstrated experimentally. \n\n1 \n\nINTRODUCTION \n\nA number of models of hippocampal function have been developed (Burgess et aI., 1994; \nMyers and Gluck,  1994; Touretzky et al., 1994), but remarkably few simulations have \naddressed hippocampal function within the constraints provided by physiological and ana(cid:173)\ntomical data.  Theories of the function of specific subregions of the hippocampal forma(cid:173)\ntion often do not address physiological mechanisms for changing dynamics between \nlearning of novel stimuli and recall of familiar stimuli.  For example, the afferent input to \nthe hippocampus  has been proposed to form orthogonal representations of entorhinal \nactivity (Marr, 1971; McNaughton and Morris, 1987; Eichenbaum and Buckingham, \n1990), but simulations have not addressed the problem of when these representations \n\n\f78 \n\nMichael  E.  Hasselmo.  Eric  Schnell.  Joshua  Berke.  Edi Barkai \n\nshould remain stable, and when they should be altered.  In addition, models of autoasso(cid:173)\nciative memory function in region CA3 (Marr, 1971; McNaughton and Morris, 1987; \nLevy, 1989; Eichenbaum and Buckingham, 1990) and  heteroassociative memory function \nat the Schaffer collaterals projecting from region CA3 to CAl (Levy, 1989; McNaughton, \n1991) require very different activation dynamics during learning versus recall. \n\nAcetylcholine may set appropriate dynamics for storing new information in the cortex \n(Hasselmo et aI., 1992, 1993; Hasselmo, 1993, 1994; Hasselmo and Bower, 1993).  Ace(cid:173)\ntylcholine has been shown to  selectively suppress synaptic transmission at intrinsic but \nnot afferent fiber synapses (Hasselmo and Bower, 1992), to suppress the neuronal adapta(cid:173)\ntion of cortical pyramidal cells (Hasselmo et aI., 1994; Barkai and Hasselmo, 1994), and \nto enhance long-term potentiation of synaptic potentials (Hasselmo, 1994b).  Models \nshow that suppression of synaptic transmission during learning prevents recall of previ(cid:173)\nously stored information from  interfering with  the storage of new information (Hasselmo \net al., 1992, 1993; Hasselmo, 1993, 1994a), while cholinergic enhancement of synaptic \nmodification enhances the rate of learning (Hasselmo, 1994b). \n\nFeedback regulation of cholinergic modulation may set the appropriate level of cholin(cid:173)\nergic modulation dependent upon the novelty or familiarity of a particular input pattern. \nWe have explored possible mechanisms for the feedback regulation of cholinergic modu(cid:173)\nlation in simulations of region CAl (Hasselmo and Schnell, 1994) and region CA3.  Here \nwe show that self-regulated learning and recall of self-organized representations can be \nobtained in a network simulation of the hippocampal formation.  This model utilizes selec(cid:173)\ntive cholinergic suppression of synaptic transmission in stratum radiatum of region CA3, \nwhich has been demonstrated in brain slice preparations of the hippocampus. \n\n2  METHODS \n\n2.1.  SIMPLIFIED REPRESENTA nON OF HIPPOCAMPAL NEURONS. \n\nIn place of the sigmoid input-output functions used in many models, this model uses  a \nsimple representation in which the output of a neuron is not explicitly constrained, but the \ntotal network activity is regulated by feedback from  inhibitory interneurons and adapta(cid:173)\ntion due to intracellular calcium concentration.  Separate variables represent pyramidal \ncell membrane potential a, intracellular calcium concentration c, and the membrane poten(cid:173)\ntial of inhibitory interneurons h: \nl1ai  =  Ai -l1 ai - J..l.C + L Wijg(aj - e) - Hikg(hk - e h) \nI1c\u00b7  =  'Vg(a . - e ) - Qc \nI1hk  = IWkjg(aj-eo)-l1hk- IHk/g(h/-e) \nwhere A = afferent input, \" = passive decay of membrane potential, Il = strength of cal-\n\n/ \n\nj \n\nj \n\nI \n\ni\u00b7 \n\nI \n\nC \n\n\fA Model of Hippocampus \n\n79 \n\ncium-dependent potassium current (proportional to intracellular calcium), Wij = excitatory \nrecurrent synapses (longitudinal association path tenninating in stratum radiatum), gO is a \nthreshold linear function proportional to the amount by which membrane potential \nexceeds an output threshold 00  or threshold for calcium current Oc' 'Y = strength of voltage(cid:173)\ndependent calcium current, n = diffusion constant of calcium, Wki = excitatory synapses \ninhibitory interneurons, Hilc = inhibitory synapses from interneurons to pyramidal cells, \nHk}= inhibitory synapses between interneurons.  This representation gives neurons adapta(cid:173)\ntion characteristics similar to those observed with intracellular recording (Barkai and Has(cid:173)\nselmo, 1994), including a prominent afterhyperpolarization potential (see Figure 1). \nAn \n\nB \n\n.... ~ .. \n\n....J \n\n'N~JJL \n\n\\..--\n\nC lO \n\n-14 \n\nFigure 1. Comparison of pyramidal cell model with experimental data. \n\nIn Figure I, A shows the membrane potential of a modeled pyramidal cell in response to \nsimulated current injection.  Output of this model is a continuous variable proportional to \nhow much membrane potential exceeds threshold.  This is analogous to the reciprocal of \ninterspike interval in real neuronal recordings.  Note that the model displays adaptation \nduring current injection and afterhyperpolarization afterwards, due to the calcium-depen(cid:173)\ndent potassium current.  B shows the intracellularly recorded membrane potential in a pir(cid:173)\nifonn cortex pyramidal cell, demonstrating adaptation of firing frequency due to \nactivation of calcium-dependent potassium current. The firing rate falls off in a manner \nsimilar to the smooth decrease in firing rate in the simplified representation.  C shows an \nintracellular recording illustrating long-tenn afterhyperpolarization caused by calcium \ninflux induced by spiking of the neuron during current injection. \n\n2.2.  NETWORK CONNECTIVITY \n\nA schematic representation of the network simulation of the hippocampal fonnation is \nshown in Figure 2. The anatomy of the hippocampal fonnation is summarized on the left \nin A, and the function of these different subregions in the model is shown on the right in \nB. Each of the subregions in the model contained  a population of excitatory neurons with \na single inhibitory interneuron mediating feedback inhibition and keeping excitatory \nactivity bounded.  Thus, the local activation dynamics in each region follow the equations \npresented above.  The connectivity of the network is further summarized in Figure 3 in the \nResults section.  A learning rule of the Hebbian type was utilized at all synaptic connec(cid:173)\ntions, with the exception of the mossy fibers from the dentate gyrus to region CA3, and the \nconnections to and from  the medial septum. Self-organization of perforant path synapses \nwas obtained through decay of synapses with only pre or post-synaptic activity, and \ngrowth of synapses with combined activity.  Associative memory function at synapses \n\n\f80 \n\nMichael  E.  Hasse/mo,  Eric  Schnell, Joshua  Berke,  Edi Barkai \n\narising from region CA3 was obtained through synaptic modification during cholinergic \nsuppression of synaptic transmission. \n\nB \n\nEntorhinal cortex \n\n\"\"\"\"\"\"\"\"\"\" \n\n\u2022 \u2022 \u2022 \u2022 \u2022 \u2022 \n\n: Self-organized \n: representation \n:  r-----..L.\\ \n\u2022 \n~  Comparison \n.L-______ ~~.-----~~~--~~ \n\nFeedback regulation of \ncholinergic modulation \n\nRegulation of \nlearning dynamics \n\nFigure 2.  Schematic representation of hippocampal circuitry \nand the corresponding function of connections in the model. \n\n2.3. CHOLINERGIC MODULA nON \n\nThe total output from region CAl determined the level of cholinergic modulation within \nboth region CA3 and CAl, with increased output causing decreased modulation.  This is \nconsistent with experimental evidence suggesting that activity in region CAl and region \nCA3 can inhibit activity in the medial septum, and thereby downregulate cholinergic mod(cid:173)\nulation.  This effect was obtained in the model by excitatory connections from region CAl \nto an inhibitory interneuron in the medial septum, which suppressed the activity of a cho(cid:173)\nlinergic neuron providing modulation to the full network.  When levels of cholinergic \nmodulation were high, there was strong suppression of synaptic transmission at the excita(cid:173)\ntory recurrent synapses in CA3 and the Schaffer collaterals projecting from region CA3 to \nCAL  This prevented the spread of activity due to previous learning from  interfering with \nself-organization.  When levels of cholinergic modulation were decreased, the strength of \nsynaptic transmission was increased, allowing associative recall to dominate.  Cholinergic \nmodulation also increased the rate of synaptic modification and depolarized neurons. \n\n2.4. TESTS OF SELF-REGULATED LEARNING AND RECALL \n\nSimulations of the full hippocampal network evaluated the response  to the sequential pre(cid:173)\nsentation of a series of highly overlapping activity patterns in the entorhinal cortex.  Recall \nwas tested with interspersed presentation of degraded versions of previously presented \nactivity patterns.  For effective recall, the pattern of activity in entorhinal cortex layer IV \nevoked by degraded patterns matched the pattern evoked by the full learned version of \nthese patterns.  The function of the full network is illustrated in Figure 3.  In simulations \n\n\fA Model  of Hippocampus \n\n81 \n\nfocused on region CA3, activity patterns were induced sequentially in  region CA3, repre(cid:173)\nsenting afferent input from the entorhinal cortex.  Different levels of external activation of \nthe cholinergic neuron resulted in different levels of learning of new overlapping patterns. \nThese results are illustrated in Figure 4. \n\n2.5. BRAIN SLICE EXPERIMENTS \n\nThe effects in the simulations of region CA3 depended upon the cholinergic suppression \nof synaptic transmission in stratum radiatum of this region  The cholinergic suppression of \nglutamatergic synaptic transmission in region CA3  was tested in brain slice preparations \nby analysis of the influence of the cholinergic agonist carbachol on the size of field poten(cid:173)\ntials elicited by stimulation of stratum radiatum.  These experiments used techniques sim(cid:173)\nilar to previously published work in region CAl (Hasselmo and Schnell, 1994). \n\n3  RESULTS \n\nIn the full hippocampal simulation, input of an unfamiliar pattern to entorhinal cortex \nlayer II resulted in high levels of acetylcholine.  This allowed rapid self-organization of \nthe perforant path input to the dentate gyrus and region CAl.  Cholinergic suppression of \nsynaptic transmission in region CAl prevented recall from  interfering with self-organiza(cid:173)\ntion.  Instead, recurrent collaterals in region CA3 stored an autoassociative representation \nof the input from  the dentate gyrus to region CA3, and connections from CA3 to CA 1 \nstored associations between the pattern of activity in CA3 and the associated self-orga(cid:173)\nnized representation in region CAl. \n\nidentity \n\nself-org \"matrix  ~ \n\n,,~> \n\n--I.~  \u00a3 \nSelf-org  ~ \n~ \nI  \"T I' \n\n\u2022 \n\niden~ity \nmatrix \n, \n\njld II \n\n111111 \n2  I I I  II \nr \nr \nQ)2dl' \n~  311111 \n4  n , n \n3d I I \nI \n4d  II \nI \nld II \nI \n2d \n\nI  'I  I \n\" I  \"'\"' \nII'  I \nI \nI' I  I \n\" \nI \nII \n\nI  I I \n\n1 I \n\nIf n  I \nII \n, \nI'll \nn \n'f \nn \nI \nII \n\n, \n'I \n\n~ \" \n,auto-\" \nM  assoc \n-\n\u00ab \n\u00ab \nu \nu \nc \nc \n\u2022 \n\u2022 \n.9  hetero-\n.9 \nhetero-\n~  assoc \n~  assoc \n~ \n~ \n, r \nI \n, , \nI \n\n..  , \n\u2022 \nf'l't \ni \nItt  r \n\" \nI  I\"'\" \n~' \n\n\" \n\u2022 \n'I  .. \n\n'I \n\n1  I' ,I \n\n\" \n\n, I  I , , \n\n1 II  r \n\n1 \n\nII \nU \nI I \nU \nI \nH \nI \nU \nI I \n\nI \n\n, , \nI \n\nNeuron # \n\nu  ..... \n'at) \nc  >. \n:.a.\u00a3:i \n8 \nC! \nI.Ll \n\nl  11  I J 1 \n't \nIII I' \nI  I  I \n\" \n, \n'(  1'1  I I \n( U.  I  I  I  n \n, 'I \nI  II \nI  W.  ,  I  I  II \nII  \u2022 .1 \nI \nl  \" \nI  I \nI  III I I \nJl  1 \nlU \n\nII \n\nI \n\nFigure 3.  Activity in each subregion of the full network simulation of the hippocampal \nformation during presentation of a sequence of activity patterns in entorhinal cortex. \n\n\f82 \n\nMichael E.  Hasselmo,  Eric Schnell,  Joshua  Berke,  Edi Barkai \n\nIn Figure 3. width of the lines represents the activity of each neuron at a particular time \nstep.  As seen here. the network forms a self-organized representation of each new pattern \nconsisting of active neurons in the dentate gyrus and region CAL  At the same time. an \nassociation is formed between the self-organized representation in region CAl and the \nsame afferent input pattern presented to entorhinal cortex layer IV.  Four overlapping pat(cid:173)\nterns (1-4) are presented sequentially. each of which results in learning of a separate self(cid:173)\norganized representation in the dentate gyrus and region CAl. with an association formed \nbetween this representation and the full input pattern in entorhinal cortex. \n\nThe recall characteristics of the network are apparent when degraded versions of the affer(cid:173)\nent input patterns are presented in the sequence (ld-4d).  This degraded afferent input \nweakly activates the same representations previously formed in the dentate gyrus.  Recur(cid:173)\nrent excitation in region CA3 enhances this activity. giving robust recall of the full version \nof this pattern.  This activity then reaches CA 1. where it causes strong activation if it \nmatches the pattern of afferent input from  the entorhinal cortex.  Strong activation in \nregion CAl decreases cholinergic modulation. preventing formation of a new representa(cid:173)\ntion and allowing recall to dominate.  Strong activation of the representation stored in \nregion CAl then activates the full representation of the pattern in entorhinal cortex layer \nIV.  Thus. the network can accurately recall each of many highly overlapping patterns. \n\nThe effect of cholinergic modulation on the level of learning or recall can be seen more \nclearly in a simulation of auto-associative memory function in region CA3 as shown in \nFigure 4.  Each box shows the response of the network to sequential presentation of full \nand degraded versions of two highly overlapping input patterns.  The width of the black \ntraces represents the activity of each of 10 CA3 pyramidal cells during each simulation \nstep.  In the top row. level of cholinergic modulation (ACh) is plotted.  In A. external acti(cid:173)\nvation of the cholinergic neuron is absent. so there is no cholinergic suppression of synap(cid:173)\ntic transmission.  In this case. the first pattern is learned and recalled properly. but \nsubsequent presentation of a second overlapping pattern results only in recall of the previ(cid:173)\nously learned pattern.  In B. with greater cholinergic suppression. recall is suppressed suf(cid:173)\nficiently to allow learning of a combination of the two input patterns.  Finally. in C. strong \ncholinergic suppression prevents recall. allowing learning of the new overlapping pattern \nto dominate over the previously stored pattern. \n\nA \n\nB  ACh input = 0.15 \n\nACh input = 0.0 \n\n.....  ... - .. -\n,11\",,11.  . ... -..... -. \n\nN \n\nC  ACh input = 0.3 \n\n' ... _ .. , ....... , .. _ ...... \n\n11'_  I \u2022 \u2022  \" '  \u2022 \u2022  __  '111\" \n\n111 ... 111 ...... 11 .... \"' ... 11 \u2022 \u2022  111 \u2022\u2022 , \n\n\u00b7' .. _\"11 .. \u00b7\u00b7.'.\u00b7 \n\n. . . . . . . . . . 111 . . . .... . \n\n.,~ . \",. \n\n.u . . . . .... , \n\n. \n\nStored \npatterns \n\nQ\\ \n\nACh \n\u2022\u2022\u2022  Inhib \n\u2022\u2022 \n\u2022\u2022 \n\u2022 \n\u2022 \n-.gN.g - N \n\n.';:: \n~ \n~ \n\n0 \n\nFigure 4.  Increased cholinergic suppression of synaptic transmission in region CA3 \n\ncauses greater learning of new aspects of afferent input patterns. \n\n\fA  Model  of Hippocampus \n\n83 \n\nExtracellular recording in brain slice preparations of hippocampal region CA3 have dem(cid:173)\nonstrated that perfusion of the cholinergic agonist carbachol strongly suppresses synaptic \npotentials recorded in stratum radiatum, as shown in Figure 5.  In contrast, suppression of \nsynaptic transmission at the afferent fiber synapses arising from entorhinal cortex is much \nweaker.  At a concentration of 20J..tM, carbachol suppressed synaptic potentials in stratum \nradiatum on average by 54.4% (n=5).  Synaptic potentials elicited in stratum lacunosum \nwere more weakly suppressed, with an average suppression of28%. \n\nControl \n\nCarbachol \n(20JlM) \n\nWash \n\nFigure 5.  Cholinergic suppression of synaptic transmission in stratum radiatum of CA3. \n\n4  DISCUSSION \n\nIn this model of the hippocampus, self-organization at perforant path synapses forms com(cid:173)\npressed representations of specific patterns of cortical activity associated with events in \nthe environment.  Feedback regulation of cholinergic modulation sets appropriate dynam(cid:173)\nics for learning in response to novel stimuli, allowing predominance of self-organization, \nand appropriate dynamics for recall in response to familiar stimuli, allowing predomi(cid:173)\nnance of associative memory function.  This combination of self-organization and associa(cid:173)\ntive memory function may also occur in neocortical structures.  The selective cholinergic \nsuppression of feedback and intrinsic synapses has been proposed to allow self-organiza(cid:173)\ntion of feedforward synapses while feedback synapses mediate storage of associations \nbetween higher level representations and activity in primary cortical areas (Hasselmo, \n1994b).  This previous proposal could provide a physiological justification for a similar \nmechanism utilized in recent models (Dayan et al.,  1995).  Detailed modeling of cholin(cid:173)\nergic effects in the hippocampus  provides a theoretical framework for linking the consid(cid:173)\nerable behavioral evidence for a role of acetylcholine in  memory function (Hagan and \nMorris,  1989) to  the neurophysiological evidence for the effects of acetylcholine within \ncortical structures (Hasselmo and Bower, 1992; 1993; Hasselmo, 1994a, 1994b). \n\nAcknowledgements \n\nThis work supported by a pilot grant from the Massachusetts Alzheimer's Disease \nResearch Center and by an NIMH FIRST award MH52732-01. \n\nReferences \n\nBarkai E, Hasselmo ME (1994)  Modulation of the input/output function of rat piriform \ncortex pyramidal cells.  J. Neurophysiol. 72: 644-658. \n\n\f84 \n\nMichael  E.  Hasselmo,  Eric Schnell, Joshua  Berke,  Edi Barkai \n\nBarkai E, Bergman RE, Horwitz G, Hasselmo ME (1994) Modulation of associative mem(cid:173)\nory function in a biophysical simulation of rat pirifonn cortex.  J. Neurophysiol. 72:659-\n677. \nBurgess N, Reece M, O'Keefe J (1994) A model of hippocampal function.  Neural Net(cid:173)\nworks 7:  1065-1081. \nDayan P, Hinton GE, Neal RM and Zemel RS  (1995) The Helmholtz machine. Neural \ncomputation in press. \nEichenbaum, H. and Buckingham, J. (1990) Studies on hippocampal processing: experi(cid:173)\nment, theory and model.  In: Learning and computational neuroscience:  foundations of \nadaptive networks, M. Gabriel and J.  Moore, eds., Cambridge, MA:  MIT Press. \nHagan, JJ and Morris, RGM (1989) The cholinergic hypothesis of memory: A review of \nanimal experiments. In Psychopharmacology of the Aging Nervous System, L.L. Iversen, \nS.D. Iversen and S.H. Snyder, eds. New York: Plenum Press, p. 237-324. \nHasselmo, M.E. (1993) Acetylcholine and learning in a cortical associative memory. Neu(cid:173)\nral Compo 5: 22-34. \nHasselmo ME (1994a) Runaway synaptic modification in models of cortex: Implications \nfor Alzheimer's disease. Neural Networks 7:  13-40. \nHasselmo ME (1994b) Neuromodulation and cortical function.  Behav. Brain Res. in press \nHasselmo ME, Anderson, BP and Bower, JM (1992) Cholinergic modulation of cortical \nassociative memory function. J. Neurophysiol. 67(5):  1230-1246. \nHasselmo ME, Bower JM (1992) Cholinergic suppression specific to intrinsic not afferent \nfiber synapses in rat pirifonn (olfactory) cortex. J. Neurophysiol. 67(5):  1222-1229. \nHasselmo ME, Bower JM (1993) Acetylcholine and memory. Trends Neurosci 16:218-\n222. \nHasselmo ME, Barkai E, Horwitz G, Bergman RE (1993) Modulation of neuronal adapta(cid:173)\ntion and cortical associative memory function. In: Computation and Neural Systems II \n(Eeckman F, Bower JM, ed). Norwell, MA:  Kluwer Academic Publishers. \nHasselmo ME, Schnell E (1994) Laminar selectivity of the cholinergic suppression of syn(cid:173)\naptic transmission in rat hippocampal region CAl: Computational modeling and brain \nslice physiology.  J. Neurosci.  14: 3898-3914. \nLevy WB (1989) A computational approach to hippocampal function.  In:  Computational \nmodels of learning in simple neural systems (Hawkins RD, Bower GH, ed), pp. 243-305. \nOrlando, FL: Academic Press. \nMyers CE and Gluck M (1994) Context, conditioning and hippocampal rerepresentation \nin animal learning.  Behav. Neurosci.  108: 835-847. \nMarr 0  (1971) Simple memory: A theory for archicortex. Phil. Trans. Roy. Soc. B \nB262:23-81 \nMcNaughton BL (1991) Associative pattern completion in hippocampal circuits: New \nevidence and new questions. Brain Res. Rev.  16:193-220. \nMcNaughton BL, Morris RGM (1987) Hippocampal synaptic enhancement and infonna(cid:173)\ntion storage within a distributed memory system. Trends Neurosci.  10:408-415. \nTouretzky OS, Wan HS and Redish AD (1994) Neural representation of space in rats and \nrobots.  In Zurada JM and Marks RJ (eds) Computational Intelligence: Imitating life. \nIEEE Press. \n\n\f", "award": [], "sourceid": 972, "authors": [{"given_name": "Michael", "family_name": "Hasselmo", "institution": null}, {"given_name": "Eric", "family_name": "Schnell", "institution": null}, {"given_name": "Joshua", "family_name": "Berke", "institution": null}, {"given_name": "Edi", "family_name": "Barkai", "institution": null}]}