Information Technology Reference
In-Depth Information
There are two steps that we need to do, (a) create the
log, and then (b)
monitor
the hidden unit activities so
thattheyendupinthelog.
computations like
AVG
or
SUM
over them.
Then press
Ok
.
Although you could just
Run
the testing process, it is
a good idea to
Clear
the log first, because the new data
will have a different format. Now, when you
Run
your
testing process again, you will see that the hidden unit
activations appear on the right hand side. If you didn't
hit clear first, you may need to hit the
Init
button.
Also, you will probably want to resize the window taller
to fit all the events at once.
Do
.logs/New/GridLog
, and set the
updt_proc
(updating process) to
Trial_Tst
. Then,
Run
the test
epoch process, just to see what shows up in the log by
default.
One thing you will notice is that the
sum_se
and
lst_cycles
statistics have overlapping labels in the
header (top row of the display), and are displayed as col-
ored squares instead of as textual values. Furthermore,
their values are out of the default scale range of -1 to 1.
Although one could just press the
Auto
button to cor-
rect the scale to automatically fit the displayed values,
this will make it difficult to also display activity values
in the
0
1
range. Thus, we will switch the display to
a textual output.
,
!
B.4.2
Creating a New Statistic for Testing
In addition to monitoring unit activations, you can cre-
ate a wide range of different statistics. We will create
a closest-event statistic, that compares the output pro-
duced by the network to the target events in the envi-
ronment, and reports on the closest such event. This
statistic was used in the weight-based priming simula-
tion in section 9.2.1.
There are a number of ways to create a new statis-
tic. The most direct way is to select in the project win-
dow the light-blue icon for a group of statistics in one
of the schedule processes (e.g.,
loop_stats
), and
then hit the
New Stat
button. However, you may
not be sure if this is the proper location for the stat, in
which case you could just select the yellow schedule
process icon where you want the stat computed (e.g.,
Trial_Tst
),andthenhit
New Stat
from there. Fi-
nally, you can also use the
New Stat
menu item in
the
.processes
menu, which allows you to select the
statistic type and then prompts you for where to put it.
First, click with the left mouse button and drag the
lst_cycles
header over to the right so it no longer
overlaps with
sum_se
. Then, click with the right mouse
button on each of the header labels in turn, and switch
the
display_style
from
BLOCK
to
TEXT
in the dialogs
that appear. Be sure to press
Update View
in the dia-
log when you make this change (or just hit
Init
in the
log view afterwards). Now, you will see the errors and
cycles as text. (For completeness, you can also use the
middle button to reshape headers that appear as grids
— like the monitor values we'll see in a moment).
Now we are ready to monitor the hidden layer.
Click on the hidden layer in the network window
(make sure that the
Select
tool is active first). Then,
make sure you are viewing the
act
variable. Then,
do
Monitor Values/New
, which is just below the tool
and action buttons on the left side of the network win-
dow.
When we monitor network variables, we are actually
creating a
MonitorStat
object, which is reflected in
the first field of the resulting dialog that comes up. The
second field,
In Process
, is critical.
We'll use the second option. Select the
Trial_Tst
icon as the location for the stat (so it is computed as
each testing event is presented), and hit
New Stat
.
The dialog that comes up contains four impor-
tant parameters. The
Of Type
field specifies what
type of statistic to create, the
In Process
field
specifies in which process to create the statistic, the
Loop/Final
field specifies at which point in the
processing this statistic should be computed, and the
CreateAggregates
toggle determines if higher-
level aggregations of the statistic are created.
Set
In Process
to
Trial_Tst
, so that the mon-
itored values show up in our testing process instead of
in the training one.
You can leave the remaining fields as they are, but
be aware that you have an option here to not just
COPY
the values, but alternatively you could perform simple
Select
ClosestEventStat
for the type of statistic
to create, and
Trial_Tst
is set already for the process.
Search WWH ::
Custom Search