Memory - Computational Explorations in Cognitive Neuroscience - page 284

rated into the aggregation learned by the cortex itself,
through the standard slow learning process. This phe-
nomenon is known as consolidation , and it can explain
some interesting properties of memory loss (amnesia)
that occurs when the hippocampus is damaged.
Consider the phenomenon of retrograde amnesia ,
where the most recent memories up to the point of the
hippocampal damage are lost, while memories from the
more distant past are preserved. This goes against the
usual time course of forgetting, where memories get
worse as more time passes. Retrograde amnesia can be
explained if these distant memories were consolidated
into the intact cortical system, whereas more recent
memories remained dependent on the now-damaged
hippocampus. For a more detailed treatment of this
set of ideas and issues, see McClelland et al. (1995).
We will cover other examples of hippocampal-cortical
memory interactions in section 9.7.
In what follows, we describe the critical functional
properties of the hippocampal system in terms of two
competing mechanisms: pattern separation and pat-
tern completion . Pattern separation leads to the forma-
tion of different, relatively non-overlapping represen-
tations in the hippocampus. As discussed previously,
when different subsets of units are used to encode dif-
ferent memories, there is no interference because dif-
ferent sets of weights are involved. Pattern completion
enables partial cues to trigger the activation of a com-
plete, previously encoded memory. Thus, pattern sep-
aration operates during the encoding of new memories,
and pattern completion operates during the retrieval of
existing memories.

tion). Simultaneously, activation flows from the EC to
the CA1, forming a somewhat pattern separated but also
invertible representation there — that is, the CA1 repre-
sentation can be inverted to reinstate the corresponding
pattern of activity over the EC that originally gave rise
to the CA1 pattern in the first place (McClelland &
Goddard, 1996). An association between the CA3 and
CA1 representations is encoded by learning in the con-
nections between them.
Having encoded the information in this way, retrieval
from a partial input cue can occur as follows. Again,
the EC representation of the partial cue (based on in-
puts from the cortex) goes up to the DG and CA3.
Now, the prior learning in the feedforward pathway and
the recurrent CA3 connections leads to the ability to
complete this partial input cue and recover the origi-
nal CA3 representation. Pattern completion in the CA3
works much like the exploration from chapter 3. This
completed CA3 representation then activates the corre-
sponding CA1 representation, which, because it is in-
vertible, is capable of recreating the complete original
EC representation.
If, on the other hand, the EC input pattern is novel,
then the weights will not have been facilitated for this
particular activity pattern, and the CA1 will not be
strongly driven by the CA3. Even if the EC activity
pattern corresponds to two components that were pre-
viously studied, but not together, the conjunctive nature
of the CA3 representations will prevent recall (O'Reilly
et al., 1998).

Pattern Separation

The main mechanism that the hippocampus uses to
achieve pattern separation is to make the representations
sparser (having fewer units active). This is the princi-
ple we tried to exploit in our “cortical” model of AB-
AC list learning in section 9.2.2, with relatively small
effects (but in the right direction) — as we summarized
previously, the hippocampus employs sparseness on a
fairly dramatic scale in the DG, CA3 and CA1 areas,
which consequently has much larger effects.
To understand why sparse representations lead to pat-
tern separation, first imagine a situation where the hip-
pocampal representation is generated at random with

Encoding and Retrieval

We can summarize the basic operations of the model
by explaining how the encoding and retrieval of mem-
ories works in terms of the areas and projections of the
hippocampus. The general scheme for encoding is that
activation comes into the EC from the cortex, and then
flows to the DG and CA3, forming a pattern separated
representation across a sparse, distributed set of units
that are then bound together by rapid Hebbian learn-
ing within the recurrent collaterals (also, learning in the
feedforward pathway helps to encode the representa-