Tutorial for Constructing Simulations in PDP - Computational Explorations in Cognitive Neuroscience - page 429

parameters that are marked as special to a given child).
First, let's edit the default layer spec.

One optional step in configuring a network that is
sometimes a good idea is to move the bias spec to be a
child of the basic con spec, so that changes in the learn-
ing rate will automatically propagate to the bias spec.

Do .specs/Edit/LeabraLayerSpec_0 .
We can use this layer spec to control the hidden layer,
so let's set the name appropriately. Do this by editing the LeabraConSpec_0 object,

and then selecting Transfer from the menu under the
children field, and select LeabraBiasSpec_0 .
Now when you look in the specs menu, you can see
that the bias spec is under the con spec. Now you are
done configuring the network!

Do Ctrl-u and type in HiddenLayer in the name
field.
Then, we can set the kwta field to a specific number
of active units for the hidden layer.

, !

, !

First, change USE PCT to USE K , and then set the
k value to 4.
The default inhibitory function is KWTA INHIB (be-
cause it is more dependable), but KWTA AVG INHIB
often works better for hidden layers.

B.2.6 Creating an Environment

The next menu down the line in the project window is
the .environments menu.

Set compute i to KWTA AVG INHIB . When
this parameter is changed, you should also change
i kwta pt to .6 instead of .25. Go ahead and Apply
these changes.
Now, we can create two new children of this layer
spec, one for the input layer one for the output layer.

Do .environments/New/Environment , and
press Ok (using the left mouse button).
(notice that there are several other types of environ-
ments that one could create, but we will just make the
generic one.) An environment window will appear.
Note that the New Event button is highlighted.

, !

Use the New Child menu item in the .specs
menu of the project window, and select HiddenLayer
as the parent to create a child from.
You will then get a new object dialog. We will just
create one new spec at a time, so leave the parameters
as they are.

Press New Event , and create 5 events.
These new events appear in the list on the left side
of the window. We could select and extend-select these
events to view them in this window, or we could use the
following shortcut.

Use the View:Actions/SelectEvents menu
(on the right side of the event window, where the view-
specific menu items are), and just press Ok .
This will select all the events. We could manually
click into the events to enter different activation pat-
terns, as we did in some of the explorations. How-
ever, we can also use various automatic event genera-
tion techniques.

Use the right mouse button on the Ok button to edit
the new child layer spec.
Let's make this one for the output layer.

, !

Do Ctrl-u and type in OutputLayer . Then, click
on the button right next to the kwta field label.
This will mark this as unique to this child, and not
inherited from the parent.

Then, change the kwta k parameter to 6, and Ok
the edit dialog.
Next, we can create the input layer spec.

Select Permuted Binary Min Dist in the
Generate menu, which will bring up a dialog window.
This function will create random patterns that all
have the same number of active values, and ensure that
these patterns do not overlap too much (this is the min
dist part). The pat_no field specifies which pattern in
the event to generate patterns for (0 is the input layer,
and 1 is for the output layer). Keep this at 0.

, !

Repeat the above process to create an

InputLayer spec, with a k value of 5.
We're not done yet — we have to apply these new
specs to the appropriate layers (this step is often forgot-
ten!).

Select the Input layer, and then do Selections/
Set Layer Spec , and select InputLayer . Then do
the same for the output layer.

Set the n_on parameter to 5.
This will cause each pattern to have 5 active units
(1's), and the rest inactive (0's).