Page 215 -

5.7 Approximation Methods in NN Training 203

of determining the conjugate directions, using an iterative algorithm that avoids
having to compute the Hessian. Details on the algorithm implementation can be
found in Haykin (1999).

Figure 5.37. Conjugate-gradient descent in a parabolic surface. The black arrow is
a conjugate direction of the initial white arrow.

The conjugate gradient method is especially suited for difficult problems with
many classes and features. It can then provide a much-improved solution over the
ba~k-~ro~a~ation method.

o A B o C A D - AD + DE 't LD 0 Shift rn FS A Susp
-5
-2.2 -1.2 -0.2 0.8 1.8 2.8 3.8

Figure 5.38. Scatter plot of the 10 classes of CTG data relative to the two main
principal components.