Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 264

Spat2 Output Target

Motter, 1993). These recordings show that stimulus-
specificneuralfiring in V4 is reduced when visual pro-
cessing (“attention”) is directed toward other stimuli
within the receptive field of the neuron, but not when
this attention is directed outside the receptive field.
Such attentional effects could result directly from lat-
eral inhibitory connectivity (i.e., competition) among
V4 neurons, but it is also possible that interactions be-
tween the spatial and object pathways could be impor-
tant for explaining these and other similar effects.
Our model has some important limitations. For ex-
ample, because we focused specifically on the contribu-
tion of the dorsal cortical spatial processing pathway to
attention, other potentially important contributors were
not represented. In particular, the thalamus has often
been discussed as playing an important role in attention
(e.g., LaBerge, 1990; Crick, 1984). Although a com-
plete discussion of the role of the thalamus is beyond
the present scope, it is worth noting that a model based
on many of the same principles we just explored is con-
sistent with much of the biological and behavioral data
on thalamic attention (Wager & O'Reilly, submitted).
One concern that readers may have about the present
model is the simplicity of the object representations.
Clearly, it is highly unrealistic to assume that there is
a unique one-to-one mapping between low-level visual
features and object representations. How much is this
simplification contributing to the observed results? The
next section helps to address this concern by using the
detailed object recognition system described previously
in conjunction with a spatial processing pathway.

V4/IT

Spat1

V2

V1

LGN_On LGN_Off

Figure 8.29: Object recognition model with interactive spa-
tial representations that enable sequential processing of mul-
tiple objects.

way to one object at a time, enabling it to successfully
perform in an environment containing multiple objects.
Because all of the essential ideas have been covered in
the discussion of the previous models, we proceed di-
rectly into the exploration of this model.

8.6.1 Exploring the Complex Attentional Model

[ Note: this simulation requires a minimum of 128Mb of
RAM to run. ]

Open the project objrec_multiobj.proj.gz in
chapter_8 to begin.
You will notice that the network is essentially the
same as in the object recognition model, except for
the addition of two spatial processing layers, Spat1
and Spat2 (figure 8.29). These two layers are lat-
erally interconnected with V1 and V2, respectively,
and amongst themselves, providing the interactions that
give rise to spatially mediated attentional effects. There
is also a Target layer next to the output layer, which
does not interact with the network, but is useful because
it displays the identities of the objects presented to the
network. Thus, the output activation should correspond
to one of the active target units.

, !

8.6 Spatial Attention: A More Complex Model

The simple model we just explored provides a useful
means of understanding the basic principles of spatial
attention as it interacts with object processing. How-
ever, the model does not really address one of the main
motivations for having spatial attention in the first place
— to avoid the binding problem when perceiving dis-
plays containing multiple objects. The next simulation
basically combines both of the previous two models,
the full-fledged object recognition model and the sim-
ple spatial attention model. The resulting model can use
spatial attention to restrict the object processing path-