Image Analysis, Classification and Change Detection in Remote Sensing, 3rd Edition

434 Efficient Neural Network Training Algorithms

and build up H row for row:

H =







v

⊤

1

H

.

.

.

v

⊤

n

w

H







.

The excerpt from the IDL program FFNCG__DEFINE.PRO shown in Listing

B.1 extends the object class FFN introduced in Cha pter 6 and implements a

vectorized version of the above determination of of v

⊤

H (method FFNCG::ROP)

and H (method FFNCG::HESSIAN). The eigenvalues of H are calculated in

metho d FFNCG::EIGENVALUES. The co rresponding Py thon class Ffncg is coded

in the Python module supervisedclass.py.

B.2 Scaled conjugate gradient training

The backpropagation algorithm of Chapter 6 attempts to minimize the cost

function locally, that is, weight up dates