The Operation of Memory Systems in the Brain - Computational Neuroscience: A Comprehensive Approach - page 492

pus. Instead, what is applicable to both the primate and rat hippocampal recordings
is that hippocampal neurons contain a representation of space (for the rat, primar-
ily where the rat is, and for the primate primarily of positions 'out there' in space)
which is a suitable representation for an episodic memory system. In primates, this
would enable one to remember, for example, where an object was seen. In rats, it
might enable memories to be formed of where particular objects (for example those
defined by olfactory, tactile, and taste inputs) were found. Thus at least in primates,
and possibly also in rats, the neuronal representation of space in the hippocampus
may be appropriate for forming memories of events (which usually in these animals
have a spatial component). Such memories would be useful for spatial navigation,
for which according to the present hypothesis the hippocampus would implement
the memory component but not the spatial computation component. Evidence that
what neuronal recordings have shown is represented in the non-human primate hip-
pocampal system may also be present in humans is that regions of the hippocampal
formation can be activated when humans look at spatial views [21, 55].

16.2.3 Hippocampal models

These neuropsychological and neurophysiological analyses are complemented by
neuronal network models of how the hippocampus could operate to store and re-
trieve large numbers of memories [73, 75, 76, 92, 114, 115]). One key hypothesis
(adopted also by [46]) is that the hippocampal CA3 recurrent collateral connections
which spread throughout the CA3 region provide a single autoassociation network
that enables the firing of any set of CA3 neurons representing one part of a memory
to be associated together with the firing of any other set of CA3 neurons representing
another part of the same memory (cf. [44]). The generic architecture of an attractor
network is shown in Figure 16.5. Associatively modifiable synapses in the recurrent
collateral synapses allow memories to be stored, and then later retrieved from only a
part, as described by [4, 32, 33, 82, 92]. The number of patterns p each representing
a different memory that could be stored in the CA3 system operating as an autoas-
sociation network would be as shown in equation 16.1 (see [82, 92], which describe
extensions to the analysis developed by [33]).

C RC
a ln

p ≈ k (16.1)

1
a )

(

where C RC is the number of synapses on the dendrites of each neuron devoted to the
recurrent collaterals from other CA3 neurons in the network, a is the sparseness of
the representation, and k is a factor that depends weakly on the detailed structure
of the rate distribution, on the connectivity pattern, etc., but is roughly in the order
of 0.2-0.3. Given that C RC is approximately 12,000 in the rat, the resulting storage
capacity would be greater than 12,000 memories, and perhaps up to 36,000 memories
if the sparseness a of the representation was as low as 0.02 [114, 115].
Another part of the hypothesis is that the very sparse (see Figure 16.3) but power-
ful connectivity of the mossy fibre inputs to the CA3 cells from the dentate granule