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

444 Efficient Neural Network Training Algorithms

B.3 Kalman filter training

In this section the recursive linear regression method described in Appendix

A will be applied to train the feed-forward neura l network of Figure 6.10.

The appropriate cost function in this case is the quadra tic function, Equation

(6.30), or more specifically, its local version

E(ν) =

1

2

kℓ(ν) − m(ν)k

2

, (B.34)

however our algorithm will also minimize the cross-e ntropy cost function, as

will be mentioned later.

✖✕

✗✔

✖✕

✗✔

✖✕

✗✔

✖✕

✗✔

✖✕

✗✔

1

j

L

k

.

.

.

.

.

.

⑦

q

✿

❃

✲

0

m

k

(ν)

w

o

k

1

n

1

(ν)

n

j

(ν)

n

L

(ν)

FIGURE B.2

An isolated output neuron.

We begin with consideration of the

training process of an isolated neu-

ron. ...