Information Technology Reference
In-Depth Information
+
Spatial
Object
Alert
Interrupt
V1
(features x
location)
Localize
Disengage
Move
Engage
Figure 8.23: Spatial representations can have an attentional
effect via direct connections to the object processing system
(which is spatially organized at its lower levels). The spa-
tial pathway can have a strong impact (thicker arrow) on ob-
ject processing, while remaining sensitive to bottom up inputs
(thicker bottom-up arrows than top-down ones).
Inhibit
Figure 8.24: Attentional selection theory proposed by Pos-
ner et al. (1984). The parietal cortex is thought to be respon-
sible for the disengage process.
the spatial pathway would provide top-down support for
this region in V1, meaning that the features in this re-
gion will be preferentially processed by the object path-
way. With the presence of lateral inhibition, the regions
not supported by the top-down spatial input will be at
a competitive disadvantage, and therefore less likely to
be active.
Although this model is simple and elegant, it has an
important limitation — it exacerbates the tradeoff be-
tween the ability to focus attention at a given location,
and the ability to respond to new inputs in novel loca-
tions. The attentional effects are mediated by what is
effectively the input layer of the network (V1). Thus,
the top-down spatial activation focused on a specific lo-
cation inhibits the ability of other locations in the input
layer to become active. However, these other locations
need to become active before attention can be switched
to them, creating a catch-22 situation — you have to al-
low other locations to be active for attention to switch,
but these other locations have to be suppressed to im-
plement attention in the first place. Thus, for this kind
of system to be able to switch attention to a new loca-
tion, the top-down projections from the spatial system
to V1 must be made relatively weak — so weak that the
resulting level of spatial attention is often incapable of
sufficiently focusing the object processing pathway on
one object over another.
The alternative model that we explore here is shown
in figure 8.23. The key difference is that this model
includes lateral interconnectivity between the spatial
pathway and the object processing pathway, that en-
ables spatial attention to affect object processing di-
rectly without having to go through the V1 input layer
first. This configuration avoids the catch-22 situation by
enabling strong attentional effects on object recognition
without requiring the model to ignore novel inputs via
strong top-down attentional modulation. Thus, the ef-
ficacy of the spatial modulation of object processing is
not limited by the need to keep the spatial system sen-
sitive to bottom-up input. The strong influence of spa-
tial processing on the object pathway is emphasized in
the figure by the thicker arrow. Similarly, the ability of
the model to be sensitive to bottom-up input is empha-
sized by thicker bottom-up arrows than top-down ones.
For this model to work, the spatial pathway must in-
teract with the lower, spatially organized levels of the
object processing pathway (e.g., areas V2 and V4 in the
object recognition model described in the previous sec-
tion). Our model also includes multiple spatial scales of
processing both in the object pathway (as in the previ-
ous model), and correspondingly in the spatial pathway.
Search WWH ::




Custom Search