{"title": "What Does the Hippocampus Compute?: A Precis of the 1993 NIPS Workshop", "book": "Advances in Neural Information Processing Systems", "page_first": 1173, "page_last": 1175, "abstract": null, "full_text": "What Does the Hippocampus Compute?: \n\nA Precis of the 1993 NIPS Workshop \n\nMark A. Gluck \n\nCenter for Molecular and Behavioral Neuroscience \n\nRutgers University \nNewark, NJ  07102 \n\ngluck@pavlov.rutgers.edu \n\nComputational models of the hippocampal-region provide an important method for \nunderstanding the functional role of this brain system in learning and memory. The \npresentations in this workshop focused on how modeling can lead to a unified \nunderstanding of the interplay among hippocampal physiology, anatomy, and \nbehavior. Several approaches were presented. One approach can be characterized as \n\"top-down\" analyses of the neuropsychology of memory, drawing upon brain-lesion \nstudies in animals and humans. Other models take a \"bottom-up\" approach, seeking \nto infer emergent computational and functional properties from detailed analyses of \ncircuit connectivity and physiology  (see Gluck &  Granger,  1993, for a review). \nAmong the issues discussed were:  (1) integration of physiological and behavioral \ntheories of hippocampal function,  (2) similarities and differences between animal \nand human studies, (3) representational vs.  temporal properties of hippocampal(cid:173)\ndependent behaviors, (4) rapid vs. incremental learning, (5) mUltiple vs. unitary \nmemory systems, (5) spatial navigation and memory, and (6) hippocampal \ninteraction with other brain systems. \n\nJay McClelland, of Carnegie-Mellon University, presented one example of a top(cid:173)\ndown approach to theory development in his talk, \"Complementary roles of \nneocortex and hippocampus in learning and memory\"  McClelland reviewed \nfindings indicating that the hippocampus appears necessary for the initial acquisition \nof some forms of memory, but that ultimately all forms of memory are stored \nindependently of the hippocampal system.  Consolidation in the neocortex appears \nto occur over an extended period -- in humans the process appears to extend over \nseveral years.  McClelland suggested that the cortex uses interleaved learning to \nextract the structure of events and experiences while the hippocampus provides a \nspecial system for the rapid initial storage of traces of specific events and \nexperiences in a form that minimizes mutual interference between memory traces. \nAccording to this view, the hippocampus is necessary to avoid the catastrophic \n\n1173 \n\n\f1174 \n\nGluck \n\ninterference that would result if memories were stored directly in the neocortex. \nConsolidation is slow to allow the gradual integration of new knowledge via \ncontinuing interleaved learning (McClelland,  1994/in press). \n\nIn another example of top-down modeling, Mark Gluck of Rutgers University \ndiscussed \"Stimulus representation and hippocampal function in animal and human \nlearning.\" He described a computational account of hippocampal-region function in \nclassical conditioning (Gluck & Myers,  1993; Myers & Gluck,  1994). In this model, \nthe hippocampal region constructs new stimulus representations biased by two \nopponent constraints:  first, to differentiate representations of stimuli which predict \ndifferent future events, and second, to compress together representations of co(cid:173)\noccurring or redundant stimuli.  This theory accurately describe the role of the \nhippocampal region in a wide range of conditioning paradigms.  Gluck also \npresented an extension of this theory which suggests that stimulus compression may \narise from the operation of circuitry in the superficial layers of entorhinal cortex, \nwhereas stimulus differentiation may arise from the operation of constituent circuits \nof the hippocampal formation. \n\nDiscussing more physiologically-motivated \"bottom-up\" research, Michael \nHasselmo, of Harvard University, talked about  \"The septohippocampal system: \nFeedback regulation of cholinergic modulation.\" Hasselmo presented a model in \nwhich feedback regulation sets appropriate dynamics for learning of novel input or \nrecall of familiar input.  This model extends previous work on cholinergic \nmodulation of the piriform cortex (Hasselmo,  1993; Hasselmo,  1994). This model \ndepends upon a comparison in region CAl between self-organized input from \nentorhinal cortex and recall of patterns of activity associated with CA3 input.  When \nnovel afferent input is presented, the inputs to CA 1 do not match, and cholinergic \nmodulation remains high, allowing storage of a new association.  For familiar input, \nthe match between input patterns suppresses modulation, allowing recall dynamics \ndominated by input from CA3. \n\nMichael Recce and Neil Burgess, from England, presented their work on \"Using \nphase coding and wave packets to represent places.\" They are attempting to model \nthe spatial behavior of rats in terms of the firing of single cells in the hippocampus. \nA reinforcement signal enables a set of \"goal cells\" to learn a population vector \nencoding the direction of the rat from the goal. This is achieved by exploiting the \napparent phase-coding of place cell firing,  and the presence of head-direction cells. \nThe model shows rapid latent-learning and robust navigation to previously \nencountered goal locations (Burgess, O'Keefe, & Recce,  1993; Burgess, Recce, & \n0' Keefe,  1994). Spatial trajectories and cell firing characteristics compare well with \nexperimental data. \n\nRichard Granger, of U .C.  Irvine, was originally scheduled to talk on \"Distinct \nbiology and computation of entorhinal, dentate, CA3 and CAl.\" Granger and \ncolleagues have noted that synaptic changes in each component of the hippocampus \n(i.e., DG, CA3 and CAl) exhibit different time courses, specificities, and \nreversibility. As such, they suggest that subtypes of memory operate serially, in an \n\n\fWhat Does the Hippocampus Compute?: A Precis of the  1993 NIPS Workshop \n\n1175 \n\n\"assembly line\" of specialized functions, each of which adds a unique aspect to the \nprocessing of memories (Granger et al,  1994). \n\nIn other talks, Bruce McNaughton of the University of Arizona discussed models of \nspatial navigation (McNaughton et aI,  1991) and William Levy from the University \nof Virginia presented a theory of how sparse recurrence of CA3 and several other, \nless direct feedback systems, leads to an ability to learn and compress sequences \n(Levy,  1989). Mathew Shapiro, of McGill University, had been scheduled to talk on \ncomputing locations and trajectories with simulated hippocampal place fields. \n\nReferences \n\nBurgess N, O'Keefe 1 &  Recce M (1993)  Using hippocampal \"place cells\" for \n\nnavigation, exploiting phase coding, in:  Hanson S 1, Giles C L &  Cowan 1 D \n(eds.) Advances in Neural Information Processing Systems 5. San Mateo, CA: \nMorgan Kaufmann. \n\nBurgess N,  Recce M and O'Keefe 1 (1994) A model of hippocampal function, \n\nNeural Networks, Special Issue on Neurodynamics and Behavior, to be \npublished. \n\nGluck, M. and Granger, R. (1993). Computational models of the neural bases of \n\nlearning and memory. Annual Review of Neuroscience.  16, 667-706. \n\nGluck, M., &  Myers, C. (1993). Hippocampal mediation of stimulus \n\nrepresentation:  A computational theory. Hippocampus, 3.,  491-516. \n\nGranger, R., Whitson, 1., Larson, 1.  and Lynch, G.  (1994). Non-Hebbian \n\nproperties of L TP enable high-capacity encoding of temporal sequences. Proc. \nNat'l. Acad. Sci., (in press). \n\nHasselmo, M.E.  (1993) Acetylcholine and learning in a cortical associative \n\nmemory. Neural Computation  5,32-44. \n\nHasselmo, M.E. (1994) Runaway synaptic modification in models of cortex: \n\nImplications for Alzheimer's disease. Neural Networks, in press. \n\nLevy, W.  B (1989)  A computational approach to hippocampal function.  In: \n\nComputational Models of Learning in Simple Neural Systems. (R.D. Hawkins \nand G.H. Bower, Eds.), New York:  Academic Press, pp. 243-305. \n\nMcClelland, 1. L.  (1994/in press).  The organization of memory:  A parallel \ndistributed processing perspective.  Revue Neurologique, Masson, Paris \n\nMcNaughton, B., Chen, L., &  Markus, E.  (1991).  \"Dead reckoning\", landmark \n\nlearning, and the sense of direction:  A neurophysiological and computational \nhypothesis. 10urnal of Cognitive Neuroscience, 3.(2),  190-202. \n\n\f", "award": [], "sourceid": 733, "authors": [{"given_name": "Mark", "family_name": "Gluck", "institution": null}]}