Information Technology Reference
In-Depth Information
cur_task
onedim
twodim
rule_change
batch
batch
env
epc_ctr
Epoch
cur_task
intact_ids
fetles_ids
dimles_ids
intact_idr
fetles_idr
dimles_idr
intact_eds
fetles_eds
dimles_eds
4
2
2
2
2
2
2
2
4
9
17
9
8
5
13
0
0
0
4
4
intact_ids
intact_ids
intact_ids
fetles_ids
fetles_ids
fetles_ids
dimles_ids
dimles_ids
dimles_ids
4
6
4
4
3
4
7
0
0
1
2
6
0
0
2
2
4
8
4
6
8
6
8
12
1
0
0
1
0
1
2
4
2
2
2
2
1
0
2
2
6
2
0
0
2
0
1
2
2
0
2
4
Figure 11.14:
Output of
RepiBatch 0 TextLog
for the
entire set of lesions and categorization rule conditions. The
first line of output discussed in the text is at the top row.
Figure 11.13:
Output of
NEpoch 0 TextLog
for all lesion
conditions of the IDS task.
#
Task
Network
Log Label
also monitor the
env:
counter in the network — stop
just when it gets to 2.
1
IDS
Intact
intact ids
2
Feat lesion
fetles ids
3
Dim lesion
dimles ids
Intradimensional Shift
4
IDR
Intact
intact idr
Now, we are ready to change the categorization rule and
test the network's ability to react appropriately. We be-
gin with the intradimensional shift rule change: the net-
work is presented with two new features within each of
the two previous dimensions, and the new target is one
of the new features from the same dimension previously
rewarded.
5
Feat lesion
fetles idr
6
Dim lesion
dimles idr
7
EDS
Intact
intact eds
8
Feat lesion
fetles eds
9
Dim lesion
dimles eds
Table 11.1:
Schedule of categorization tasks and net-
work lesion conditions that the simulation automatically steps
through, and the label used in the text log
cur task
column.
StepSettle
through the phases for the first few
trials in this sequence.
You should see that the network is still getting this
problem correct — it just happened to learn weights that
work for both problems. Now, let's accelerate the rate
of stepping as the network comes to acquire the new
categorization rule.
,
!
blocks (4 and 2 epochs) and on the rule change block
(2 epochs). Also shown in this log is the
cur_task
column, which reads
intact_ids
indicating that the
network is intact and that the task was IDS.
Next, look at the
RepiBatch_0_TextLog
(fig-
ure 11.14), which summarizes the
epc_ctr
val-
ues across a single line, with the three blocks la-
beled
onedim
,
twodim
,and
rule_change
.The
rule_change
column is the key one for comparison
with the Dias et al. (1997) results.
Next we will test the importance of the PFC areas for
performance on this IDS task. We first test the network
with a lesioned (deactivated)
PFC_Feat
layer (orbital
prefrontal cortex), and then with a lesioned
PFC_Dim
layer (dorsolateral prefrontal cortex). The simulation
will automatically cycle through the appropriate lesions
and task conditions as you step the process, according
to the schedule shown in table 11.1.
First, open one more log that will record the
summary of this sequence of training by doing
View
,
BATCH_LOG
. Then, press
StepEpoch
(which steps over
epochs, not settling phases) until the input patterns
switch back to the single-dimension (
A
) patterns, which
indicates the start of the first training block for the next
sequence of blocks.
To evaluate the overall performance of the in-
tact network on the IDS task, let's examine the
other two logs (you should have noticed that the
Epoch_0_GraphLog
was automatically cleared at
the start of the next block). First, look at the
NEpoch_0_TextLog
(figure 11.13), specifically the
epc_ctr
column, which shows the number of epochs
required to reach criterion on each of the 2 training
,
!
Search WWH ::
Custom Search