Page 246 -

234 5 Neural Networks

very well at the expense of a, a net that recognizes y well at the expense of wl
and finally a net that performs similarly well for all classes.
As shown in Table 5.10, a better solution was achieved with the majority vote
scheme than with the first MLP2:2:3 presented in section 5.6. By performing a
factor analysis in the feature space of the neural modules, we can obtain
representations of the class clusters and draw the boundaries achieved by the neural
s;lutions, as shown in Figure 5.59.

1

-1.5 .0.5 0.5 1.5 2.5
FINN
a

-2.5 . 0 #
-3.5
.1.5 -0.5 0.5 1.5 2.5
FINN
b
Figure 5.59. Neural net solutions for the three-class cork stoppers problem (wl=@,
@=a, @=*) represented in the space of the two main principal components. (a)
MLP2:2:3 with features N, PRT (solid line boundaries) and MLP3:3: with features
N, PRTG, ARTG (dotted line boundaries) ; (b) Majority vote of five neural nets
(solid line boundaries). Notice how these last boundaries retain the best
characteristics of the previous ones.