The Hippocampus and Memory
Edmund T. Rolls
DOI: https://doi.org/10.1093/acprof:oso/9780198784852.003.0024
2016-01-01
Abstract:A description of the operation of the hippocampus in episodic memory is provided, with a theory of how the hippocampus operates, and how the hippocampal memory system illustrates many of the principles of operation of the cerebral cortex. A feature of the hippocampal network is that the CA3 system has a highly developed recurrent collateral system which enables it to associate together as an autoassociation or attractor network places and objects, with the inputs coming from very different neocortical systems. The single attractor network in CA3 enables any object to be associated with any place in an episodic or event memory, and for the place to be recalled from the object, and the object from the place. This is completion of a memory in an autoassociation network. Although the representation of place in rodents is of the place where the rodent is located, in primates there is a representation of spatial view, enabling humans to remember where they have seen a person or object. The number of memories that can be stored in the hippocampus is determined largely by the number of recurrent collateral connections onto any one neuron (in the order of 12,000 in the rat) and the sparseness of the representation. The dentate gyrus input to CA3 performs pattern separation of the inputs before CA3 to help achieve the relatively uncorrelated representations that help to achieve high memory capacity in CA3. The dentate granule cells perform this operation by competitive learning; by the low probability of contact of the dentate granule cell to CA3 connections (0.00015 in the rat: a granule cell makes 46 mossy fibre connections onto 300,000 CA3 cells); and by neurogenesis with the new dentate granule cells helping to ensure that new memories are distinct from old memories by making new random connections to CA3 neurons. The dentate gyrus is the only part of cortex where new neurons can develop in adulthood, for here the premium is on selecting new random sets of neurons in CA3 for new memories distinct from previous memories, whereas elsewhere in cortex the premium is on building and reusing useful new representations. The CA1 which follows the CA3 prepares for the recall of memories, completed in CA3 by a partial retrieval cue, back to the neocortex. The CA1 has the architecture of a competitive network, useful for recoding the parts of an episodic memory necessarily represented as separate in CA3 for autoassociation in CA3. Backprojections from the hippocampal system to neocortical areas have a large number of synapses onto each neocortical neuron to achieve high recall capacity. This accounts for why there are as many backprojections as forward connections between adjacent cortical areas in a hierarchy. For the recall to work efficiently, the backprojection synapses at one or more stages of the backprojection pathway must be associatively modifiable. Once set up in this way, the backprojection connections can also be used for top-down attentional control (see Chapter 6). In addition to place cells in rodents and spatial view cells in primates, there are also time cells in the hippocampus which increase their firing at different times in a delay period. These neurons may be thought of as similar to place cells, and when associated with objects or places, enable the order in which the objects or places occurred to be recalled. CA1 is especially implicated in this operation. The medial entorhinal cortex contains grid cells in rodents which fire with several peaks of firing as a rodent traverses places. In primates, there are spatial view grid cells, which are likely to provide inputs to the spatial view cells found in the primate hippocampus. Different entorhinal grid cells operate at different spatial scales. The medial entorhinal cortex appears to operate as a continuous attractor network which can perform path integration, updating the spatial representation based on motion signals. The same type of attractor network may implement the time cells also present in the medial entorhinal cortex. The lateral entorhinal cortex provides a route for object-related information from the ventral cortical systems to reach the hippocampus.