Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 252

ject recognition this problem would be so reduced to
enable Hebbian learning to be successful on its own,
we suspect that error-driven learning will still play an
important role.

tions in controlling attention), the principles apply quite
generally across many other domains and processing
pathways.
Before we turn to the models, we need to reempha-
size the point made in chapter 7 that attention is not
a separate mechanism in our framework, but rather an
emergent property of the activation dynamics and repre-
sentational structure of the network. This contrasts with
the traditional view, where attention is seen as a distinct
mechanism. In the context of visual processing, atten-
tion has specifically been associated with the types of
spatially mediated effects that we model here. Attention
has also been used to refer to the role of the prefrontal
cortex in focusing processing in a task-relevant manner
(see chapters 7 and 11).
In our models, attention emerges from inhibitory
competition (which is responsible for imposing a lim-
itation on the total amount of activity within a given
set of representations), and constraint satisfaction op-
erating throughout the network (which determines the
representations that will become active in a given con-
text). Similar views of attention as a ubiquitous prop-
erty of the cortex have been articulated by Desimone
and Duncan (1995) and Allport (1989). Even though
we consider attention to be an emergent property, we
often refer to it as though it were a single mechanism
for convenience.
Because the mechanisms that underlie attention are
ubiquitous throughout the cortex, one would expect to
find attentional effects associated with all different sorts
of processing. Furthermore, the same kinds of process-
ing capacity limitations and binding problems discussed
here in the context of object recognition apply generally
as well. Thus, the models here provide a good demon-
stration of a very general cortical function. Also, you
will see that although spatial representations provide a
useful organizing substrate for visual attention, other
kinds of representations (e.g., in the object processing
pathway) can also make contributions to visual atten-
tion.
In the models that follow, we will also see that the
structural principles developed in chapter 7 (hierarchies
of specialized processing pathways with lateral inter-
actions) are critical for determining the ways in which
these functional considerations are resolved in the spe-

8.5 Spatial Attention: A Simple Model

The models we develop next build on the previous
model by showing how attention to specific spatial lo-
cations, mediated by the dorsal “where” visual path-
way, can enable objects to be recognized in a crowded
scene that would otherwise overwhelm the ventral ob-
ject recognition pathway. Objects tend to be spatially
contiguous, so object recognition benefits from group-
ing together and focusing processing on visual features
located within a contiguous spatial region.
As described in chapter 7, attention plays an impor-
tant role in solving the binding problem by restricting
the focus of processing to one related set of features
(e.g., one object). This focus enables the resulting pat-
tern of distributed representations to be interpretable,
because they will apply to features that should actually
be related to each other, as opposed to random combina-
tions of features from different objects (or other aspects
of the environment). We avoided this problem in the
previous model by presenting the simulation with only
one object at a time.
This multiple-object binding problem is somewhat
different than the multiple-feature binding problem con-
fronted in the previous model. With multiple objects,
we do not necessarily want to form conjunctive repre-
sentations that encode multiple objects simultaneously,
especially if these objects are completely unrelated.
However, for objects commonly seen together (e.g.,
eyes within a face), the network could use the same kind
of hierarchical feature-conjunctions and spatial invari-
ance solution to the multiple object binding problem as
it did with the multiple-feature one.
In this section and the next we develop two closely re-
lated models of attention, with the second model (which
is an extension of the above object recognition model)
perceiving multiple objects in the environment by se-
quentially focusing attention on each in turn. Although
we focus on this case of attentional effects in object
recognition (specifically the role of spatial representa-