Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 266

when the spatial representation is not clamped in by taking advantage of the accommodation (“fatigue”)

properties of neurons used in the previous simulation to
capture the inhibition of return effect (see section 2.9
for details). After a given object has been processed by
the network, the currently active neurons will fatigue
due to accommodation, and a new set of neurons can
then become activated. Presumably these new neurons
will be those that represent another object present in the
environment.

advance.
To test how much the spatial processing pathway con-
tributes to performance, we can reduce its level of influ-
ence over the object processing pathway.

Set the fm_spat1_scale parameter to .01 instead
of 1.
This will reduce the impact of the spatial pathway
significantly, and should render the network incapable
of focusing processing on individual objects. This may
provide a useful model of the object recognition deficits
(simultanagnosia) observed in patients with bilateral
parietal lobe lesions (Balint's syndrome patients), who
presumably have lost the equivalent of the spatial path-
way in our model. These patients cannot recognize ob-
jects when multiple are present in the field at once, but
can recognize individually-presented objects (just like
our earlier object recognition model without a spatial
pathway).

, !

Set accommodate field to ACCOMMODATE in the
control panel ( Apply ).
This will turn on the accommodation channels that
let potassium ions out after a unit has been active for
some period of time. It will also extend the maximum
settling time so that the activation dynamics have time
to play out.

Now, do InitTest and StepTest on the process
control panel.
You will see the same first event. But notice that after
it has correctly recognized the object in the lower right,
the focus of activation moves gradually over to the ob-
ject in the upper left, which is also correctly recognized.
Note that the spatial attention actually seems to move
continuously across space, and takes a little while to set-
tle over the second object. Thus, the output is momen-
tarily confused in the middle of this transition before it
correctly recognizes this second object. This continu-
ous movement of the spatial focus is probably due in
part to the self connections within the spatial system,
which tend to favor the activation of nearby locations.
Although this demonstration is interesting and il-
lustrates the potential of the network's emergent acti-
vation dynamics to achieve sensible-looking attention
switching, it is not a very reliable mechanism. Thus, if
you continue running on subsequent events, it does not
achieve correct recognition of both objects with particu-
larly high frequency. It is likely that better performance
on this kind of task will require the explicit training of a
control system that provides top-down activation of the
spatial representations to direct attention toward differ-
ent objects. Nevertheless, this kind of accommodation
mechanism is probably helpful in allowing attention to
be switched.

Do InitTest and then StepTest again.

Question 8.13 Describe the effects of this manipula-
tion on the patterns of activation in the network, and
the consequences thereof for the network's ability to tell
one object from another.

Set fm_spat1_scale back to 1 before continuing.
Note that although we want the spatial input to have a
strong impact on the object processing pathway, we do
not want this spatial input to be strong enough to pro-
duce activation in the object pathway all on its own. In-
stead, spatial attention should have an essentially mod-
ulatory role, which favors some activations over others,
but does not directly activate features all by itself. Com-
petitive inhibition produces this modulatory effect in the
model, because those units that receive activation from
both the spatial representations and the bottom-up feat-
ural inputs will win the competition over other units that
just receive activation from only one of these sources.
Although spatial attention is useful for enabling an
object processing pathway to focus on a single object
in an multi-object environment, it would be even more
useful if attention could be sequentially directed to the
different objects present in the scene. We can imple-
ment a simple version of this kind of attention switching

Before continuing, set accommodate to
DONT_ACCOMMODATE (or hit Defaults ).