Information Technology Reference
In-Depth Information
The back-propagation algorithm applies a correction
D
w
ji
ð
n
Þ
to synaptic weights
which is proportional to partial derivative
o
eð
n
Þ
o
w
ji
ð
n
Þ
w
ji
ð
n
Þ
which can be written as:
o
eð
n
Þ
Þ
¼
o
eð
n
Þ
Þ
:
o
e
j
ð
n
Þ
Þ
:
o
y
j
ð
n
Þ
Þ
:
o
v
j
ð
n
Þ
ð
6
Þ
w
ji
ð
n
e
j
ð
n
y
j
ð
n
v
j
ð
n
w
ji
ð
n
Þ
o
o
o
o
o
Diffrentiating Eq. (
2
) with respect to e
j
ð
n
Þ
Þ
o
e
j
ð
n
Þ
¼
o
eð
n
e
j
ð
n
Þ
ð
7
Þ
Differentiating Eq. (
1
) w.r.t. y
j
ð
n
Þ
e
j
ð
n
Þ
o
Þ
¼
1
ð
8
Þ
y
j
ð
n
o
Differentiating Eq. (
5
),
y
j
ð
Þ
o
n
Þ
¼ /
j
ð
v
j
ð
n
ÞÞ
ð
9
Þ
v
j
ð
n
o
Differentiating Eq. (
4
) w.r.t. w
ji
ð
n
Þ
o
v
j
ð
n
Þ
o
Þ
¼
y
i
ð
n
Þ
ð
10
Þ
w
ji
ð
n
Using Eqs. (
7
-
10
) in Eq. (
6
),
o
eð
n
Þ
Þ/
j
ð
Þ
¼
e
j
ð
n
v
j
ð
n
ÞÞ
y
i
ð
n
Þ
ð
11
Þ
w
ji
ð
n
o
The correction
w
ji
ð
n
Þ
applied to w
ji
ð
n
Þ
is de
ned by
D
o
eð
Þ
n
D
w
ji
ð
n
Þ¼g
w
ji
ð
n
Þ
o
ð
12
Þ
Þ/
j
ð
¼ g
e
j
ð
n
v
j
ð
n
ÞÞ
y
i
ð
n
Þ
¼ gd
j
ð
n
Þ
y
i
ð
n
Þ
Thus the weight updates for output unit can be represented as
w
ji
ð
n
Þ¼gd
j
ð
n
Þ
y
i
ð
n
Þ
ð
13
Þ
D
where
g
is a constant called learning rate parameter,
d
is local gradient and is a
derivative of error with respect to v
j
and y is the input.
Search WWH ::
Custom Search