Memory - Computational Explorations in Cognitive Neuroscience - page 286

1-5%) units are activated in the hippocampus at any
given time (table 9.1), but the same principle applies.
For pattern separation to work optimally, it is impor-
tant that different receiving units are maximally acti-
vated by different input patterns. This can be achieved
by having a high level of variance in the weights and/or
patterns of partial connectivity with the inputs. We im-
plemented this idea in the simple AB-AC list learning
model by increasing the variance of the random initial
weights, which reduced interference somewhat. In the
hippocampus, the perforant pathway has diffuse, ran-
dom connectivity, which ensures that individual DG and
CA3 neurons are maximally excited by different input
patterns, facilitating pattern separation in these areas.

a judgment call. Thus, there is a basic tradeoff operat-
ing within the hippocampus itself between pattern sep-
aration and completion. Optimizing this tradeoff can
actually be used to understand several features of the
hippocampal biology (O'Reilly & McClelland, 1994).

Details of the Model

We now describe further details about the model that
are generally based on the biology, but are also shaped
by the necessity of having a reasonably simple work-
ing model. These details include the sizes and activity
levels of the model layers, the structure of the EC input
representations, and the implementation of the invert-
ible CA1-EC mapping.
Figure 9.12 shows the structure of the model, and an
example activation pattern. Table 9.1 shows that the
model layers are roughly proportionately scaled based
on the anatomy of the rat, but the activation levels are
generally higher (less sparse) to obtain sufficient abso-
lute numbers of active units for reasonable distributed
representations given the small total number of units.
The activity levels are implemented using the basic
kWTA inhibition function. We use just CPCA Heb-
bian learning here because it is sufficient for simple in-
formation storage, but it is likely that the hippocampus
can also take advantage of error-driven learning in more
complex tasks (O'Reilly & Rudy, in press).
The EC input representations incorporate the topo-
graphic and columnar characteristics of the EC by hav-
ing different cortical areas and/or sub-areas represented
by different slots , which can be loosely thought of as
representing different feature dimensions of the input
(e.g., color, font, semantic features, etc.). Our EC has
36 slots with four units per slot; one unit per slot is ac-
tive, with each unit representing a particular “feature
value.” There are two functionally distinct layers of
the EC, one that receives input from cortical areas and
projects into the hippocampus (superficial or EC_in ),
and another that receives projections from the CA1 and
projects back out to the cortex (deep or EC_out ). Al-
though the representations in these layers are probably
different in their details, we assume that they are func-
tionally equivalent, and use the same representations
across both for convenience.

Pattern Completion

Pattern completion is as important as pattern separation
for understanding hippocampal function. If only pat-
tern separation were at work, then any time you wanted
to retrieve previously stored information using anything
other than exactly the same original input activation pat-
tern, the hippocampus would instead store a new pattern
separated version of the input instead of recognizing
it as a retrieval cue for an existing memory. Thus, to
actually use the memories stored in the hippocampus,
the countervailing mechanism of pattern completion is
needed.
There is a fundamental tension between pattern sep-
aration and pattern completion. Consider the follow-
ing event: a good friend starts telling you a story about
something that happened to her in college. You may
or may not have heard this story before, but you have
heard several stories about this friend's college days.
How does your hippocampus know whether to store this
information as a new memory and keep it separate (us-
ing pattern separation) from the other memories, or to
instead complete this information to an existing mem-
ory and reply “you told me this story before?” In one
case, your hippocampus has to produce a completely
new activity pattern, and in the other it has to produce a
completely old one. If you have perfect memory and the
stories are always presented exactly the same way each
time, this problem has an obvious solution. However,
imperfect memories and noisy inputs (friends) require