Memory - Computational Explorations in Cognitive Neuroscience - page 287

a) b)

CA3 CA1 CA3 CA1

DG DG

EC_in EC_out EC_in EC_out

Input Input

Figure 9.12: The hippocampus model. a) The areas and connectivity, and the corresponding columns within the Input, EC, and
CA1. b) An example activity pattern. Note the sparse activity in the DG and CA3, and intermediate sparseness of the CA1.

We argued above that the CA1 translates the pattern-
separated CA3 representations back into activation pat-
terns on the EC during pattern completion using in-
vertible representations. At the same time it must also
achieve some amount of pattern separation to minimize
interference in the learning of CA3-CA1 mappings. In-
deed, this pattern separation in CA1 may explain why
the hippocampus actually has a CA1, instead of just as-
sociating CA3 directly back with the EC input. These
two functions of the CA1 are in conflict, because invert-
ibility requires a systematic mapping between CA1 and
EC, and pattern separation requires a highly nonlinear
(nonsystematic) mapping. The model achieves a com-
promise by training the CA1-EC mapping to be invert-
ible in pieces (referred to as columns ), using pattern-
separated CA1 representations. Thus, over the entire
CA1, the representation can be composed more sys-
tematically and invertibly (without doing any additional
learning) by using different combinations of representa-
tions within the different columns, but within each col-
umn, it is conjunctive and pattern separated

(i.e., 12 EC units), which is consistent with the rela-
tively point-to-point connectivity between these areas.
The weights for each CA1 column were trained by tak-
ing one such column with 9.4 percent activity level
and training it to reproduce any combination of pat-
terns over 3 EC_in slots (64 different combinations)
in a corresponding set of 3 EC_out slots. Thus, each
CA1 has a conjunctive, pattern separated representation
of the patterns within the 3 EC slots. The cost of this
scheme is that more CA1 units are required (32 vs12
per column in the EC), which is nonetheless consistent
with the relatively greater expansion in humans of the
CA1 relative to other hippocampal areas as a function
of cortical size (Seress, 1988). A further benefit is that
only certain combinations of active CA1 units (within a
column) correspond to valid EC patterns, allowing in-
valid combinations (e.g., due to interference) to be fil-
tered out. In the real system, slow learning can develop
these CA1 invertible mappings in all the columns sepa-
rately over time.

(McClel-

land & Goddard, 1996).
In the model, the CA1 columns have 32 units each,
so that the entire CA1 is composed of 12 such columns.
Each column receives input from 3 adjacent EC slots

9.3.3 Explorations of the Hippocampus

[Note: this model requires at least 64Mb of RAM to

run.]