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.

The multilayer network illustrated in Figure 4-4 is atypical because it directly<br />

connects input and output units. Most modern networks eliminate such direct connections<br />

by using at least one layer <strong>of</strong> hidden units to isolate the input units from<br />

the output units, as shown in Figure 4-5. In such a network, the hidden units can<br />

still be described as carving a pattern space, with point coordinates provided by the<br />

input units, into a decision region. However, because the output units do not have<br />

direct access to input signals, they do not carve the pattern space. Instead, they<br />

divide an alternate space, the hidden unit space, into decision regions. The hidden<br />

unit space is similar to the pattern space, with the exception that the coordinates <strong>of</strong><br />

the points that are placed within it are provided by hidden unit activities.<br />

Output<br />

Unit<br />

Modifiable<br />

Connections<br />

Hidden<br />

Units<br />

Modifiable<br />

Connections<br />

Input<br />

Units<br />

Figure 4-5. A typical multilayer perceptron has no direct connections between<br />

input and output units.<br />

When there are no direct connections between input and output units, the hidden<br />

units provide output units with an internal representation <strong>of</strong> input unit activity.<br />

Thus it is proper to describe a network like the one illustrated in Figure 4-5 as being<br />

just as representational (Horgan & Tienson, 1996) as a classical model. That connectionist<br />

representations can be described as a nonlinear transformation <strong>of</strong> the input<br />

unit representation, permitting higher-order nonlinear features to be detected, is<br />

why a network like the one in Figure 4-5 is far more powerful than one in which no<br />

hidden units appear (e.g., Figure 4-3).<br />

When there are no direct connections between input and output units, the representations<br />

held by hidden units conform to the classical sandwich that characterized<br />

Elements <strong>of</strong> Connectionist <strong>Cognitive</strong> <strong>Science</strong> 147

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

Saved successfully!

Ooh no, something went wrong!