Page 965 - The Mechatronics Handbook

0066_Frame_C32.fm Page 15 Wednesday, January 9, 2002 7:54 PM

KOHONEN
LAYER WINNER

NORMALIZED INPUTS

(a)

w

(b)

FIGURE 32.14 A winner take all architecture for cluster extracting in the unsupervised training mode: (a) network
connections, (b) single-layer network arranged into a hexagonal shape.

HI AAADDEN NEURONS
+1
OUTPUT
NEURONS
+1

OUTPUTS
+1

INPUTS

WEIGHTS ADJUSTED EVERY STEP +1
ONCE ADJUSTED WEIGHTS AND THEN FROZEN

FIGURE 32.15 The cascade correlation architecture.

Cascade Correlation Architecture
The cascade correlation architecture was proposed by Fahlman and Lebiere (1990). The process of
network building starts with a one-layer neural network and hidden neurons are added as needed. The
network architecture is shown in Fig. 32.15. In each training step, a new hidden neuron is added and its
weights are adjusted to maximize the magnitude of the correlation between the new hidden neuron
output and the residual error signal on the network output to be eliminated. The correlation parameter
S must be maximized:

O P
S = ∑ ∑ ( V p V) E po – E o) (32.38)
(
–
o=1 p=1

©2002 CRC Press LLC