Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 389

value is zero. This keeps the hidden to PFC weights
at their low tonic value, so the PFC tends to maintain
its current values. However, there is some noise added
to the gating function, so it is possible for transitions
to occur even with no AC change, especially if there is
nothing presently active in the PFC. Finally, if an er-
ror is made after correct performance, then there is a
negative Æ value, which causes the PFC weight gain to
decrease significantly in strength, including the gain on
the recurrent self-maintenance weights. This effectively
resets the PFC activations.
Note that although we generally describe this pro-
cess of AC-mediated activation and deactivation of
the frontal representations in terms of trial-and-error
search, the search process is not actually completely
random. The PFC units that are likely to get activated
are those that receive the strongest input from the hid-
den layer. Thus, the search process is guided by the
activations of the hidden units, which has the advantage
of restricting the search to only relevant aspects of the
input (i.e., it doesn't randomly search through the entire
space of possible PFC representations).
We test the model on the three types of rule changes
tested by Dias et al. (1997): IDR, IDS, and EDS. For
each of these rule changes, we test an intact network
first and then networks with lesions to the feature-level
(orbital) prefrontal cortex or to the dimension-level (lat-
eral) prefrontal cortex.
Before each rule change, each network receives 2
blocks of training (as did Dias et al.'s (1997) monkeys).
In the first block, features from only one dimension are
presented, where one feature is rewarded and the other
is not. In the second block (described earlier in our
initial presentation of the task), the two features from
the other dimension are added, but the categorization
rule remains the same as before. The network receives
multiple trials within each of the 2 training blocks and
proceeds from each block after completing two epochs
without error. After the 2 training blocks are completed,
the categorization rule changes so that the target: (1)
shifts to be a newly presented feature within the di-
mension (IDS), (2) reverses to be the other (previously
seen) feature within the dimension (IDR), or (3) shifts
to a newly presented feature within the other dimension
(EDS), thereby reversing at the dimensional level.

11.4.2 Exploring the Model

Open the project id_ed.proj.gz in chapter_11
to begin.
The network is as pictured in figure 11.12.

, !

Begin by exploring the network connectivity using
r.wt and clicking on units in each of the different layers.
You should notice that the hidden units receive di-
rectly from their corresponding input unit, and from the
appropriate units in both the PFC layers. Note also that
the PFC units do not encode the location information,
only the object properties. The PFC units each receive
a single self-connection, which provides their mainte-
nance ability (using isolated representations). The out-
put units have random initial weights from the hidden
layer units — it is these weights that are adapted during
learning to solve the task.

, !

Training

Now, let's step through the categorization learning.
As described earlier, the first block of training trials
presents features from only one dimension.

First, make sure that you are viewing act (activa-
tions). Then, open up two logs to monitor training by do-
ing View , EPOCH_LOG , and View , NEPOCH_LOG . Then,
do ReInit and StepSettle on the overall id_ed con-
trol panel.
You will see the minus phase of the first categoriza-
tion trial, which has the first 2 units within dimension

, !

representing one feature on the left, and the second 2
units representing the other feature on the right. The lo-
cation of these features (which is on the left and which
is on the right) varies randomly. In this simulation, the
target feature is represented by the activation of the first
2 units in dimension A . The hidden layer cannot ini-
tially decide which of the two features to represent, so
it has weak activation for both of them.

Press StepSettle again to get to the plus phase.
You should now see that the output unit correspond-
ing to the location of the first feature is activated, pro-
viding the correct response. Because the network did
not get this answer correct, the AC unit is provided
with the absence of reward. (For purely technical im-
plementation reasons, this is actually encoded by a very
small positive value, 1.0e-12, which you should see as

, !