06.09.2021 Views

Mind, Body, World- Foundations of Cognitive Science, 2013a

Mind, Body, World- Foundations of Cognitive Science, 2013a

Mind, Body, World- Foundations of Cognitive Science, 2013a

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

anding when a hidden unit’s activities were plotted using a jittered density plot<br />

(Berkeley et al., 1995). A jittered density plot is a two-dimensional scatterplot <strong>of</strong><br />

points; one such plot can be created for each hidden unit in a network. Each plotted<br />

point represents one <strong>of</strong> the patterns presented to the hidden unit during wiretapping.<br />

The x-value <strong>of</strong> the point’s position in the graph is the activity produced in that<br />

hidden unit by the pattern. The y-value <strong>of</strong> the point’s position in the scatterplot is a<br />

random value that is assigned to reduce overlap between points.<br />

An example <strong>of</strong> a jittered density plot for a hidden value unit is provided in<br />

Figure 4-15. Note that the points in this plot are organized into distinct bands,<br />

which is consistent with the geometric analysis. This particular unit belongs to a<br />

network <strong>of</strong> value units trained on a logic problem discussed in slightly more detail<br />

below (Bechtel & Abrahamsen, 1991), and was part <strong>of</strong> a study that examined some <strong>of</strong><br />

the implications <strong>of</strong> activity banding (Dawson & Piercey, 2001).<br />

0.0 0.2 0.4 0.6 0.8 1.0<br />

Hidden Unit Activity<br />

Figure 4-15. An example <strong>of</strong> banding in a jittered density plot <strong>of</strong> a hidden value<br />

unit in a network that was trained on a logic problem.<br />

Bands in jittered density plots <strong>of</strong> hidden value units can be used to reveal the kinds<br />

<strong>of</strong> features that are being detected by these units. For instance, Berkeley et al. (1995)<br />

reported that all <strong>of</strong> the patterns that fell into the same band on a single jittered<br />

density plot in the networks did so because they shared certain local properties or<br />

features, which are called definite features.<br />

There are two types <strong>of</strong> definite features. The first is called a definite unary<br />

feature. When a definite unary feature exists, it means that a single feature has the<br />

same value for every pattern in the band. The second is called a definite binary feature.<br />

With this kind <strong>of</strong> definite feature, an individual feature is not constant within<br />

Elements <strong>of</strong> Connectionist <strong>Cognitive</strong> <strong>Science</strong> 175

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!