Introduction to the PDP Simulation Environment - Computational Explorations in Cognitive Neuroscience - page 422

The various buttons at the top left of the window
control the display of the events. Evt Label allows
one to select what information is displayed at the top
(header) of each event display. This is normally the
event name, but could be its frequency or other such
information. Pat Label controls what is displayed
as the header for each Pattern within the event (a pat-
tern typically corresponds to one layer's worth of event
information — the pattern of activity that is presented
to a layer). The Pat Block field controls what is dis-
played in the values (blocks) for each pattern. Normally
this is just the activation value for the event, but there
are also special flags that could be displayed.
The configuration and behavior of events is con-
trolled by corresponding event specs — these can be
edited by pressing the Edit Specs button, which
switches the display to spec editing mode. Pressing
Edit Events switches back to editing the events.

els described in section A.6, these control panels just
apply to the individual process in question. The follow-
ing special buttons are available on a process object:

New Init, Re Init initialize the process (cause it to
start over), with the only difference between the two
being that Re Init reuses the previous random
seed, so that the next run will precisely replicate the
prior one, even if random numbers (e.g., noise, ran-
dom event presentation order) are involved, while
New Init gets a new random seed and will do dif-
ferent things if randomness is involved. 1
Run will run the process, starting from wherever it was
last stopped or initialized, and continuing until it is
done — in the case of the epoch process, running
means presenting all the different events in the en-
vironment to the network.
Step will do one step of running at a time, with the
granularity of processing and the number of steps de-
termined by the step parameters in the window.
Stop will stop a running process (and is otherwise un-
available — “ghosted”).
Go To will skip to a given value of the loop counter
(e.g., event number in the EpochProcess).

A.10 Processes

Themainactionofmostprocessesissimplytoloop
over its subordinate processes. Almost all processes
have a loop counter , which reflects the number of times
the process has looped over its subordinate processes
(e.g., the number of epochs that have been processed in
a TrainProcess, or the number of events that have been
processed in an EpochProcess). Processes also act as
placeholders for actions that take place at a given grain
size (epochs, trials, phases, etc.), so that, for example,
the network display can be updated at different grain
sizes by linking its updating to a given process. Simi-
larly, logs are associated with a given process to record
information summarized at that grain size.
As we mentioned, statistics, processes and scripts can
be hung off of any level of process to provide very flex-
ible control mechanisms. There are three levels within
each process where such things can be attached: init,
loop, and final, which correspond to the start of a given
grain size, within each loop, or at the end.

A.10.2 Statistics

As we have mentioned, statistics are attached to pro-
cesses, and compute useful information about the state
of the network so that it can be reported in a log. An
example is the SE (squared error) statistic, which com-
putes the squared error of the network's output relative
to the training target values. This SE statistic must be
computed at the trial level, because that is where the ac-
tual and target activation states are available. However,
we often want to see how the network is performing
at a higher grain size, for example at the epoch level.
To do this, each statistic is capable of aggregating in-
formation from another statistic of the same type at a
lower grain level. These aggregators can perform dif-
ferent operations such as sum, average, and so on. All
1 Note that random numbers on a computer are really pseudo-
random — the sequence can be replicated given the starting seed num-
ber, but there is no systematic relationship between one number and
the next, which makes the sequence effectively random.

A.10.1 Process Control Panels

Somewhat confusingly, processes also have control
panels, but unlike the overall project-wide control pan-