Tutorial for Constructing Simulations in PDP - Computational Explorations in Cognitive Neuroscience - page 431

Finally, we need to specify that this pattern will be a
TARGET , providing a plus-phase training signal to this
output layer.

in the environment, calling the TrialProcess on
each one. Each event is processed by one or more
SettleProcess iterations, one for each phase of set-
tling (minus then plus, or just minus during testing), and
finally each settling phase consists of multiple iterations
over the CycleProcess , which performs one cycle
of activation updating in the network. Fortunately, we
can usually just create the top-level process, and all the
others will be created automatically.
Creating processes happens in the .processes
menu in the project window (the next one down the
line).

Do this by editing the type field for this pattern
(which shows an initial value of INACTIVE ), using the
right mouse button clicking trick.
Another way of editing these values is to select
the field using the left button, and then click the
Edit Val(s) button. Once the specs are configured,
we can go back to view the events.

, !

Press Edit Events to see the events.
You will see that the events now have the new output
layer. We can set some activation values for this new
pattern.

Do New/TrainProcess (note that there are many
different types of processes shown here that one could
create).
Note that the new object dialog that comes up has a
Create Sub Procs option — if this is clicked (as it
is by default), then appropriate subprocesses under this
one will be automatically created (which you want to
do).

Click on one of the Output2 units for each event to
provide some activation target for this layer (don't forget
to Apply ).
We will return to the environment in a later sec-
tion when we take up the issue of programming in the
PDP++ script language, which can enable you to create
more sophisticated environments in a relatively efficient
manner.
You can iconify the window now — the project win-
dow can also be used to iconify and view windows.
Clicking twice with the left button on an object will tog-
gle between iconifying or viewing that object's window.
Note that iconifying a process icon collapses the display
of information under it, as they don't have real display
windows. You will probably want to use this view/icon
control to manage the windows through the rest of this
tutorial.

Press Ok on the dialog, and then look at the project
window, which contains a display of all the new pro-
cesses.
In the project window, you should see icons
representing the network and environment (pink
and green, respectively), and then a hierarchy
of yellow icons representing the processes just
created ( EpochProcess , TrialProcess ,
SettleProcess ,and CycleProcess as de-
scribed above). The buttons on the left of the project
window provide many different ways of manipulating
and viewing these processes and their interactions with
the other objects in the project.

Click twice on the environment icon.

We will explore a

couple of them here.
Within most of the main scheduling processes, you
will also see icons (in blue) for two commonly used
statistics that are automatically created — the sum-
squared error statistic and a cycles-to-settle statistic.
These statistics are computed originally in the trial pro-
cess ( Trial_0 ),andthenare aggregated up through
the higher levels of the processing hierarchy to compute
epoch-wise sums, etc.

B.2.7 Creating Training Process

Having created a network and an environment to train
it on, we just need to create a set of processes which
will systematically control the presentation of the events
to the network, and generally orchestrate the process-
ing and learning of the network. As you should re-
call from the previous appendix, processes are orga-
nized into a hierarchy, with each process responsible
for a different temporal scale of processing. Thus, the
TrainProcess simply calls the EpochProcess a
number of times, which in turn iterates over the events

Select the sum_Trial_0_LeabraSE_Stat icon
(representing the sum-squared-error statistic) at the
Trial_0 level, and then hit the Show Aggs button at