Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 392

produces exactly the opposite of the correct responses
(due to the reversal of the target stimuli). This rapid de-
activation of the PFC units facilitates switching to an-
other categorization rule. Next, we will observe as the
deactivated PFC units are reactivated.

StepEpoch , monitoring the NEpoch_0_TextLog .
Stop just after the network has completed all three
blocks (two training and one rule change) with both le-
sions — there are no interesting lesion effects for the
IDS task so we can just skip over these two lesions.
The cur_task column in both logs shows
fetles_ids and dimles_ids to indicate the
two different lesion conditions — the first with
lesioned PFC_Feat layer and the second with
lesioned PFC_Dim layer — tested on IDS. In
RepiBatch_0_TextLog ,the rule_change col-
umn reports a 2 and a 4 for both the lesioned networks.
Thus, the network, like Dias et al.'s (1997) monkeys,
does not seem to rely on prefrontal cortex for the IDS
task.

, !

Continue to StepSettle until the PFC units are
reactivated (even though they are in the wrong dimen-
sion). Monitor the PFC_td values associated with each
PFC layer.

, !

Question 11.5 (a) Report the PFC_td values that led
to the PFC units being reactivated. (b) Explain these
values in terms of what happened on the trial — did the
network produce a correct response, or was the activa-
tion due to random noise?

Question 11.4 Explain why PFC lesions do not affect
learning in the IDS task in the network (focus on the
advantages of the PFC, and why the demands of the
task do not require these advantages).

Continue to StepSettle for several more trials,
until PFC units in the A dimension units get activated
again.
Over this next set of trials, a combination of ran-
dom gating noise and initial correct responses caused
the PFC units to get deactivated and then reactivated
again in a new configuration, which facilitates produc-
ing a different response.
Because of the redundancy in the way the problem
can be solved, either by responding to the same side as
the target stimulus or to the side opposite of the nontar-
get, it is not always clear that the network has reversed
its focus from the previous target to the current target.
Thus, the critical property to observe is that different
PFC_Feat units get activated, and thereby support the
activation of different hidden units, which facilitates the
learning of a different response mapping. You should
see that this is indeed what has happened.

, !

Intradimensional Reversal

Now, we will change the categorization rule in a dif-
ferent way, with an intradimensional reversal. The net-
work is presented with the same stimuli during train-
ing and after the rule change, but the target reverses to
be the other feature within the previously rewarded di-
mension. As with the previous exercise, we will test
an intact network first and then two lesioned networks,
providing each network with 2 blocks of training before
the rule change.

StepEpoch , monitoring the NEpoch_0_TextLog ,
and stopping just after the first 2 training blocks have
been completed. Note the two PFC_Feat units that are
active.
Now, we are in the position to watch what hap-
pens during an intradimensional reversal trial. We will
switch back to stepping at the level of a single phase of
settling, and watch the network's response to the errors
that are produced.

, !

StepEpoch through the remaining trials in the IDR
block with the intact network.
You should see that the logs have recorded the
intact_idrrule_change value, which should be
9 epochs.
Next we will test the importance of the PFC areas
for performance on this intradimensional reversal task.
The automatic schedule calls for the network with a le-
sioned PFC_Feat layer (orbital prefrontal cortex) to
be run next, and then with a lesioned PFC_Dim layer
(dorsolateral prefrontal cortex).

, !

StepSettle through a few trials observing what
happens to the PFC units as errors are made.
You should observe that the PFC units are deactivated
as a result of the network's errors, because the network

, !