Large-Scale Brain Area Functional Organization - Computational Explorations in Cognitive Neuroscience - page 206

ficity causes a number of other problems, which must be
resolved in various different ways as discussed below.

chapter 3 (e.g., pattern completion, bootstrapping, mu-
tual and top-down support, etc.) play an important role
in processing. The role of mutual support in providing
aformof active memory is elaborated further in the fol-
lowing section, because it plays a particularly important
cognitive role, despite its relatively simple mechanistic
basis. Similarly, the role of inhibition in attention is
also mechanistically simple and cognitively important,
and is covered in the subsequent section.

7.2.2 Dynamic Principles

At a dynamic level, we view cognition as the result of
activation propagation through the brain's densely inter-
connected hierarchical structure of bidirectionally con-
nected processing layers as described above. Thus, via
the multiple constraint satisfaction property of bidi-
rectionally connected networks described in chapter 3,
the network will tend to produce an activation state
(e.g., a response or interpretation) that satisfies many of
the constraints imposed upon it from the environmental
inputs and the learned weights. Although many stimuli
(e.g., familiar ones) will result in straightforward, rapid
settling of the network into a relatively optimal state in
response to that stimulus, others may require more ex-
tended iterative “searching” for an appropriate activity
state.
In either case, the resulting activity state will not
typically be the same each time the same stimulus is
presented, due to a number of factors (e.g., learning,
habituation, and sensitization). Perhaps the most im-
portant variable in the response to a stimulus comes
from internally maintained activation states carried over
from prior processing, which provide additional con-
straints on the settling process. These internal states
(also known as internal context ) can dynamically al-
ter the interpretation and response to stimuli to provide
a more coherent, consistent and/or strategic set of re-
sponses over time (more on this below). For a simple
example of the role of internal context, note that the fol-
lowing two sentences produce very different interpreta-
tions of the word “bank”:

Mutual Support and Active Memory

Bidirectional excitatory connectivity allows different
representations to mutually support (excite) each other.
This mutual support is important because it enables rep-
resentations to remain active even in the absence of ex-
ternally derived excitation (e.g., from viewing a stim-
ulus), because they can instead rely on internal exci-
tation from other mutually supporting representations
to remain active over time. This can be viewed as a
form of memory , because it enables information to per-
sist over time. We will call this active memory ,which
can be contrasted with weight-based memory that re-
sults from changing weights to store new information
(see the following and chapter 9 for more on this dis-
tinction).
Active memory is not typically as long-lasting as
weight-based memory, because the active neurons ei-
ther fatigue or are interrupted by ongoing processing.
However, active memory has the distinct advantage that
it can directly influence ongoing processing in other
areas (e.g., by providing internal context as described
above), whereas weight-based memories only directly
affect those units whose weights are changed. We will
see that there are limits to the capabilities of mutual sup-
port for providing active memories, so that other active
maintenance mechanisms are needed for a more robust
and flexible active memory system. However, mutual
support provides the basic underlying mechanism, and
is the main form of active memory for those brain areas
that lack these more sophisticated active maintenance
mechanisms.

She swam from the overturned canoe to the
bank.
She walked from the post office to the bank.

Thus, the words preceding “bank” establish different
internal context representations that then alter its sub-
sequent interpretation (Cohen, Dunbar, & McClelland,
1990).
In addition to multiple constraint satisfaction, all of
the amplification -like attractor dynamics described in