Memory - Computational Explorations in Cognitive Neuroscience - page 271

tem to learn rapidly without suffering undue interfer-
ence, specifically the use of sparse conjunctive repre-
sentations. Such representations result in pattern sep-
aration , so that information is encoded very distinctly
(separately), which reduces interference. The conjunc-
tive nature of the representations can also bind together
many different features, enabling entire episodes to be
recalled from partial cues, for example.
In the domain of activation-based memory, the
generic posterior cortex model is challenged by work-
ing memory tasks, such as mental arithmetic (e.g., mul-
tiplying 42 times 7). Such tasks typically involve both
rapid updating of information (e.g., rapidly storing the
partial products of 7x2 and 7x4 )and robust mainte-
nance of this information (e.g., not forgetting 14 while
computing 7x4 ). We will see that our generic corti-
cal model cannot satisfy both of these demands because
they are mutually contradictory. Thus, we suggest that
the posterior cortex is supplemented by the specialized
frontal cortex (and more specifically the prefrontal cor-
tex ), which has a dynamic gating mechanism that can
dynamically switch between rapid updating and robust
maintenance.
When the dynamic frontal gating mechanism is
“closed,” the frontal cortex maintains activation patterns
with little interference from outside activity, and when
it “opens,” the frontal cortex can be rapidly updated.
We show how such a gating mechanism can be imple-
mented by the neuromodulatory substance dopamine
under the control of the temporal-differences learning
algorithm described in chapter 6. Thus, our frontal cor-
tex model can learn when it is appropriate to maintain
and when it is appropriate to update based on task de-
mands.
We divide our exploration of memory according to
the broad mechanisms of weight- and activation-based
memory, and begin by exploring weight-based priming
in our generic cortical model (i.e., the posterior cor-
tex). We then challenge this model to rapidly learn lists
of arbitrary but overlapping items, where we encounter
catastrophic interference. This motivates the need for
the hippocampal memory system — we will see that
our model of the hippocampus is much more capable
of rapid learning of arbitrary information. We next
move on to activation-based memories, first by explor-

ing short-term priming in the generic cortical model.
We then challenge this model to perform more demand-
ing working memory tasks, where we see that it can-
not both robustly maintain and rapidly update infor-
mation. This motivates the need for the frontal mem-
ory system - we will see that our frontal model is
much more capable of handling working memory tasks.
Then, we explore the development and interaction of
both activation-based and weight-base d memory mech-
anisms in a model of the A-not-B ( AB ) task. Finally,
we summarize some of the ways that the basic mech-
anisms developed in this chapter can be applied to ex-
plain a wider range of memory phenomena.

9.2 Weight-Based Memory in a Generic Model of

Cortex

In previous chapters we explored the formation of rep-
resentations that capture the structure of the environ-
ment ( model learning ) and the structure of tasks ( task
learning ). Chapters 4-6 discussed the basic princi-
ples behind these types of learning, and how these
principles can be implemented by changes in weights
(synaptic efficacies) on the connections between neu-
rons. Chapter 7 discussed in general terms how multi-
ple layers of neurons organized into multiple processing
streams in the cortex can produce complex transforma-
tions of input signals, which emphasize some distinc-
tions and deemphasize or collapse across others. Chap-
ter 8 showed how this works in the case of spatially in-
variant object recognition, where the network collapsed
across distinctions in the spatial locations of objects,
while retaining distinctions between different objects.
We can now summarize these earlier chapters as ex-
plorations into the ways that semantic/procedural mem-
ories are formed over long-term exposure to, and inter-
action with, the environment. The formation of such
representations (and their use in constraint-satisfaction
style processing) is perhaps the most important con-
tribution of the posterior cortex to cognition. In the
following section, we explore another manifestation of
memory in the posterior cortex, long-term or weight-
based priming. This form of memory is manifest after
single exposures to stimuli, and results in our model
from the same kinds of weight changes that, over a