Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 258

120−

Typical reaction times for young adults (i.e., college
students) on this task are roughly: neutral, 370 ms; valid
350 ms; invalid 390 ms, showing about 20 ms on either
side of the neutral condition for the effects of attentional
focus. These data should agree in general with the pat-
tern of results you obtained, but to fit the data precisely
you would have to add a constant offset of roughly 310
ms to the number of cycles of settling for each trial
type. This constant offset can be thought of as the time
needed to perform all the other aspects of the task that
are not included in the simulation (e.g., generating a re-
sponse). Note also that one cycle of settling in the net-
work corresponds with one millisecond of processing
in humans. Of course, this relationship is not automatic
— we adjusted the time constant for activation updat-
ing ( vm_dt ) so that the two were in agreement in this
particular experiment.
Now let's explore the effects of the wt_scale pa-
rameters on the network's performance, which helps
to illuminate how the network captures normal perfor-
mance in this task. First, let's try reducing spat_obj
from 2 to 1, which reduces the influence of the spatial
system on the object system.

100−

Lesioned

80−

60−

Intact

40−

0 1 2

Neutral Valid Invalid

Figure 8.27: Graph log results for both the intact and le-
sioned network on the Posner spatial cuing task. The lesioned
case shows a specific slowing on the invalid trials.

and an Invalid cue. There is just one event for the
neutral case, which is the presentation of the target ob-
ject in the left location. For the valid case, the first event
is the cue presented on the left, followed by a target
event with the target also on the left. The invalid case
has the same cue event, but the target shows up on the
right (opposite) side of space. The network's activations
are not reset between the cue and target events within a
group, but they are reset at the end of a group (after the
target). Thus, residual activation from the cuing event
can persist and affect processing for the target event, but
the activation is cleared between different trial types.

Do a Batch run with spat_obj set to 1.
You should find that the network settling time for the
invalid condition has decreased from roughly 83 cycles
for the standard network to roughly 75 cycles for this
case. This faster settling for the invalid trials is just what
one would expect — because there is less of a spatial at-
tention effect on the object system, the invalid cue does
not slow it down as much. The effects on the other trial
types are less interesting, and involve some technical ef-
fects of the wt_scale parameters on processing speed
in the network.

Turn the network display back on (it was turned off
automatically by the previous batch run). Clear the
graph log, and then Step . Note how the network re-
sponds to each of the three conditions of the Posner task
as you continue to Step through these cases. Press
Batch on the attn_ctrl control panel to run a batch
of 10 runs.
Your graph log should resemble figure 8.27. Note
that it is impossible to add the “neutral,” “valid,” and
“invalid” labels to the graph in the simulator, so you will
have to remember that these are associated with the 0, 1
and 2 (respectively) points on the X axis, as is shown in
the figure.

, !

Set spat_obj back to 2.
Now let's try reducing the influence from the V1
layer to the spatial representations by reducing the value
of v1_spat from 2 to 1.5 and then to 1.

Do a Batch for v1_spat of 1.5 and then 1.
Again the interesting effect is in the invalid trials,
where the network takes an average of roughly 91 and
115 cycles for v1_spat values of 1.5 and 1, respec-
tively. As V1 has less effect on the spatial pathway, it
becomes more and more difficult for input in a novel
location (e.g., the target presented on the opposite side

Question 8.9 (a) By what mechanism does the spatial
cue influence the subsequent processing of the target
in the valid and invalid cases? (b) How is this influ-
ence reflected in the settling time (report results from
the batch run)?