Networks of Neurons - Computational Explorations in Cognitive Neuroscience - page 108

tions about what cats and dogs in general have in com-
mon, and what is unique about them. It can also tell you
about the consistency of a given feature with either cats
or dogs — this is where the harmony function can be
useful in assessing the total constraint satisfaction level
of the network with a particular configuration of fea-
ture inputs. The network can also be used to perform
pattern completion as a way of retrieving information
about a particular individual or individuals. Thus, this
simple network summarizes many of the topics covered
in this chapter.
The knowledge embedded in the network is summa-
rized in table 3.1. This knowledge is encoded by sim-
ply setting a weight of 1 between an instance node rep-
resenting an individual cat or dog and the correspond-
ing feature value that this individual possesses (c.f., the
Jets and Sharks model from McClelland & Rumelhart,
1988). Each of the groups of features (i.e., values within
one column of the table) are represented within distinct
layers that have their own within-layer inhibition. In
addition, all of the identity units and the name units are
within their own separate layers as well. We use the
KWTA_AVG inhibitory function here, because it will be
important for the network to have some flexibility in the
actual number of active units per layer. The k parameter
is set to 3 for most layers, except Species where it is
1, and Size where it is 2.

Press Run in the control panel.
You should see that the network activates the appro-
priate features for Morris. You can think about this
process as finding the most harmonious activation state
given the input constraint of Morris, and the constraints
in the network's weights. Equivalently, you can think
about it as settling into the Morris attractor.

, !

Go ahead and try a few other name activations (be
sure to press the highlighted Apply button in the En-
viroView to make your changes stick). Also be sure to
click the previous one off by just clicking on it again (this
toggles the activity of that input).
Now, let's see how this network can give us gen-
eral information about cats versus dogs, even though at
some level it just has information about a set of individ-
uals.

, !

Select only the Cat Species input in the Envi-
roView, and press Apply and then Run .
You should find that the network activates all those
features that are typical of cats. You can repeat this for
dogs.

Question 3.15 (a) Explain the reason for the differ-
ent levels of activation for the different features of cats
when just Cat was activated. (b) How might this be

useful information?

Now let's make use of some of the constraint satis-
faction ideas. We can view the harmony of the network
over cycles of settling using a graph log.

Open project cats_and_dogs.proj.gz in (in
chapter_3 ) to begin.
As usual, take some time to examine the weights in
the network, and verify that the weights implement the
knowledge shown in the table.

, !

Press View and select GRAPH_LOG . Clear the log,
and run the network again with just the Cat input se-
lected.
Notice that, as we expected, this value appears to
monotonically increase over settling, indicating that the
network is increasingly satisfying the constraints as the
activations are updated.
Now, let's make a more specific query for the net-
work.

Now, press View in the cat_dog_ctrl overall con-
trol panel and select EVENTS to open the EnviroView
window.
You will see a replica of the network displayed in the
environment window — these are the inputs that will be
soft clamped into the corresponding network units when
the Run button is pressed.
Let's first verify that when we present an individual's
name as input, it will recall all of the information about
that individual. This is a form of pattern completion
with a single unique input cue. You should see that
Morris in the Name layer units of the EnviroView is
already on by default — we will use this.

Activate the Orange color input in addition to Cat ,
and press Run .
You should see that although the initial harmony
value was slightly larger (reflecting the greater excita-
tion present from the input), the final harmony value
was significantly lower that that for just Cat alone.