Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 265

only the features for one of the two objects remaining
active. Thus, it is only these features that drive further
processing in the subsequent object-processing layers
(V2 and V4/IT). From the object processing pathway's
perspective, it is as if only one object had been pre-
sented (except for the transient activation of both ob-
ject's features, which can confuse the network some-
what).
In this particular case, the network should recognize
the object in the lower right hand corner (object num-
ber 12 in figure 8.12). Notice that the unit active in the
output layer is one of the two possibilities activated in
the Target layer, indicating it has correctly identified
one of the two possible objects. We can record the net-
work's performance in a testing grid log.

Do BuildNet in the objrec_ctrl control panel
to build and connect the network. Then, load trained
weights by doing LoadNet . Select r.wt in the network
window, and examine the connectivity patterns for the
spatial layers.
These spatial layers have a similar spatial extent
as their corresponding object pathway representations.
However, there is just one unit per unique location in
the spatial system, because unlike the object processing
system, there is no need to represent multiple features at
each location. Note also that there is an excitatory self-
connection in the Spat1 layer, which helps to ensure
that a contiguous region of spatial representations is ac-
tivated, as opposed to having multiple blobs of activity.
This encourages the system to focus completely on one
object, instead of on parts of two different objects.
The main test of this model involves presenting mul-
tiple objects in different locations, and letting the spa-
tial representations focus processing on one of these
objects, which should then be recognized by the ob-
ject recognition pathway. Thus, this is essentially the
same task as the MULTI_OBJS environment in the pre-
vious model. However, unlike that environment, we
present both objects with equal activation and let the
network select which to focus on. Small differences in
the strengths of the weights for the different features of
the objects will result in one being slightly more active
than another.
We present the network with patterns composed of
two objects in random locations. To fit two objects into
the relatively small input region with enough space in
between to clearly distinguish them, we present the ob-
jects in the smallest of the sizes used in the previous
object recognition model, even though this size is just
at the lower limit of the resolution of the network (and
thus causes more errors).

Do View , TEST_LOG to open the testing grid log.
Then do StepTest again.
When the settling has completed, you should notice
that the grid log is updated. The sum_Outp_wrng
column in this log contains the output of a “wrong on”
statistic — it gives a 1 if the unit in the output is not one
of the possibilities in the target layer (i.e., the output
unit was wrongly active), and a 0 otherwise. Note that
a standard squared-error statistic would not be useful in
this case, because the network never activates both out-
puts, and it chooses the object to activate essentially at
random, so a single-unit target could not be anticipated.
Also shown in the log are the actual output activations,
and the targets.

You can continue to StepTest through the other
patterns in the environment.
Notice that the network makes errors (i.e.,
sum_Outp_wrng values of 1) with some fre-
quency. There are several reasons for this. First, as we
mentioned, the objects are small and relatively difficult
for the network to recognize even when presented
singly. Second, the initial activation of all features in
V1 appears to confuse the network somewhat — when
one of the two possible spatial locations is activated at
the start of settling, the rate of errors is reduced by two
thirds. This latter version may be more representative
of human visual experience, where spatial attention is
directed in a more top-down manner as one actively
explores an environment. However, it is a much more
interesting demonstration of the network's dynamics

To run this test, press StepTest in the control
panel.
You will see two objects presented to the network,
with the activations updated every 5 cycles of settling.
Notice that in the first few updates, the V1 representa-
tions for both of the objects are activated, but shortly
thereafter, the spatial pathway settles on one of the two
locations where the objects are present. This then pro-
vides a top-down bias on the V1 system, that results in