Memory - Computational Explorations in Cognitive Neuroscience - page 270

mechanisms (e.g., how effects of prior experience are
physically maintained in the neural network, and how
differences in neural network properties can result in
different kinds of representations).
Persistent effects of experience can take one of two
general forms in a neural network — changes in the
weights ( weight-based memory ) or persistent activ-
ity ( activation-based memory ). Because weight-based
memories result from weight changes, we typically
view them as the product of the kinds of learning mech-
anisms discussed in chapters 4-6. Similarly, the archi-
tectural properties that determine how activation flows
through the network (chapter 3) play an important role
in our understanding of the nature of activation-based
memories. Thus, you have already learned a great deal
about memory without necessarily realizing it! This
chapter builds on these earlier foundations by apply-
ing principles of learning and activation propagation to-
ward understanding a range of memory phenomena in
humans and other animals.
The learning and processing principles developed to
this point are intended to apply generally to the cortex.
Thus, our first goal is to explore the general memory ca-
pacities of a generic cortical model. We will see that this
cortical model of memory seems to apply more specif-
ically to the posterior areas, and not necessarily to the
frontal lobe. Thus, we will use our generic model to
characterize the posterior-cortical memory system (as
discussed in chapter 7).
The representations that result from the integrated ef-
fects of learning over many experiences comprise per-
haps the most important form of memory exhibited by
the posterior cortical system. These memories have tra-
ditionally been characterized as semantic and procedu-
ral memory , emphasizing the extent to which they cap-
ture stable aspects of the environment (“semantics”), or
often-repeated task components (“procedures”). For ex-
ample, the object recognition model in chapter 8, em-
ploying the basic learning principles developed in chap-
ters 4-6, can be said to have developed semantic mem-
ories for the different objects it can recognize.
Short-term and long-term priming are two other im-
portant and related memory phenomena that the pos-
terior memory system can subserve. Priming refers to
a facilitory effect on processing items from prior pro-

cessing of similar items (just as priming a pump facil-
itates subsequent pumping). We can understand short-
term priming in terms of the residual activation of im-
mediately prior processing, while long-term priming re-
flects the weight changes resulting from prior process-
ing. This latter form of priming is relatively long-
term due to the persistence of the underlying weight
changes, whereas the residual activation underlying
short-term priming dissipates relatively rapidly. Inter-
estingly, long-term priming can be understood using the
same underlying mechanisms that ultimately lead to se-
mantic/procedural memories — the difference is that
priming results from single “tweaks” of the weights,
while semantic/procedural memories require the accu-
mulation of many incremental weight changes over a
relatively long period of time.
Although the posterior-cortical system has powerful
learning and activation mechanisms and provides the
foundation for many memories, it has some important
limitations that necessitate the use of additional, spe-
cialized memory systems. In the domain of weight-
based memories, we will see that our generic cortical
model can suffer from a catastrophic level of interfer-
ence from subsequent learning, such that almost every-
thing the network learned about one list of items is lost
due to interference from a second list of similar items.
People show some level of interference on such tasks,
but not nearly as much as our generic cortical model.
As mentioned in chapter 7, we suggest that this discrep-
ancy is due to an intrinsic tradeoff between the long-
term learning of semantic representations and the abil-
ity to learn rapidly arbitrary information such as a list of
words. The hippocampus and related structures (which
we refer to simply as the hippocampus) seem special-
ized for this latter form of rapid arbitrary learning, while
the cortex slowly accumulates semantic/procedural rep-
resentations based on many experiences with the envi-
ronment.
Hippocampally mediated memories have been char-
acterized as episodic (representing the contents of
episodes or events; Tulving, 1972), declarative (rep-
resenting explicitly accessible information; Squire,
1992), or spatial (representing locations in space;
O'Keefe & Nadel, 1978). We instead emphasize the un-
derlying mechanisms that enable the hippocampal sys-