Memory - Computational Explorations in Cognitive Neuroscience - page 292

Some of the computational models of the hippocam-
pal anatomy and physiology have provided quantitative
estimates of potential storage capacity of the hippocam-
pal system (e.g., Treves & Rolls, 1994; Moll & Miikku-
lainen, 1997). Other models have emphasized the role
of neuromodulators in memory storage and retrieval
(e.g., Hasselmo & Wyble, 1997). The issue of storage
of sequential information (e.g., for sequential locations
visited within an environment) has been addressed in
several models (e.g., Levy, 1989; Burgess et al., 1994;
Samsonovich & McNaughton, 1997). This rich compu-
tational literature has played a critical role in shaping
ideas about hippocampal function ever since the early
ideas of Marr (1971), and constitutes one of the great
successes of computational modeling in cognitive neu-
roscience.

similarity. The distinction between weight-based and
activation-based priming is one of duration, in that the
activation state typically decays within a couple of sec-
onds, whereas weight changes can last for long periods
of time. As we saw in chapters 4-6, learning is based on
activation values, so activation-based and weight-based
priming are similar on the other dimensions (content
and similarity). This has recently been supported in the
case of semantic priming (Becker et al., 1997), which
was previously thought to occur only in an activation-
based form.
In the following exploration, we will use the same
basic paradigm as the previous weight-based priming
exploration. Thus, the network is trained to perform
a one-to-many mapping as in a stem-completion or
homophone-priming task. However, in this case, we
will turn off any weight changes during the priming
test and just observe the effects of residual activation
on subsequent network performance. Note that in real-
ity one cannot just turn off learning in this way, so the
mechanistic basis of a given priming effect (i.e., weights
or activations) is not as clear as it is in our models. Thus,
behavioral experimentalists must resort to other tech-
niques of telling the two apart. One way to do this is to
pit the two types of priming against each other, so that
prior, but not immediately preceding, experience builds
up the weights to favor one response, but immediate ex-
perience favors another. We will e xp lore this idea in
greater detail in the context of the AB phenomenon in
section 9.6.

9.4 Activation-Based Memory in a Generic Model

of Cortex

Having explored weight-based memories in the cor-
tex and hippocampus, we now turn to activation-based
memories. A good example of a simple form of
activation-based memory can be found in the attentional
model from chapter 8. In this model, residual activa-
tion from processing a cue stimulus affects subsequent
processing of a target, leading to either faster process-
ing (when the cue and target are in the same location)
or slower processing (when they are in different loca-
tions). In this section we will explore how this kind of
residual activation can result in an activation based or
short-term priming effect, similar to the weight-based
one but lasting only as long as these residual activations.
This activation-based priming is a basic property of the
simple cortical model, just like weight-based priming.
After exploring the priming effects of residual acti-
vation, we will challenge our basic cortical model with
more demanding forms of active memory, where infor-
mation must be maintained for relatively longer periods
of time and also updated rapidly.

Exploring the Model

Open the project act_priming.proj.gz in
chapter_9 to begin.
This is essentially just the same simulation as the
weight priming one, so we will assume you are already
familiar with the network and environment. We begin
by loading a pre-trained network, and opening up the
test logs.

, !

Do LoadNet in the control panel, followed by View ,
TEST_LOGS .
Let's test this network (using a somewhat different
setup than last time) to obtain a baseline measure of per-
formance.

9.4.1 Short-Term Priming

As discussed in section 9.2.1, priming can be divided
into the three dimensions of duration, content, and