Biomedical Engineering Reference
In-Depth Information
where
u
p
is the weighted sum input for node
p
. The contribution of
u
i
at the
node
i
to the error of a training pattern can be measured by computing the
error derivative written as
δ
i
=
∂
E
k
∂
u
i
=
∂
E
k
∂
c
i
∂
c
i
∂
u
i
(3.10)
For output nodes this can be written as
c
i
)
f
i
δ
i
=
−
(
y
i
−
(3.11)
where
f
i
is the derivative of the activation function. This will vary depending
on the activation function where, for example, the sigmoid function gives a
particularly nice derivative, that is,
f
i
=
f
(
u
i
)(1
−
f
(
u
i
)). The tanh function
is equally elegant with
f
i
=1
f
2
(
u
i
).
For nodes in the hidden layers, note that the error at the outputs is influ-
enced only by hidden nodes
i
, which connect to the output node
p
. Hence for
a hidden node
i
,wehave
δ
i
=
∂
E
k
∂
u
i
−
=
p
∂
E
k
∂
u
p
∂
u
p
∂
u
i
(3.12)
But the first factor is
δ
p
of the output node
p
and so
δ
i
=
p
δ
p
∂
u
p
∂
u
i
(3.13)
The second factor is obtained by observing that node
i
connects directly to
node
p
and so (∂
u
p
)
/
(∂
u
i
)=
f
i
w
pi
. This gives
δ
i
=
f
i
p
w
pi
δ
p
(3.14)
where
δ
i
is now the weighted sum of
δ
p
values from the nodes
p
, which are con-
nected to it (see Figure 3.8). Since the output values of
δ
p
must be calculated
i
+1
w
i
+1,
i
Slope of
nonlinearity
i
P
∑
w
pi
w
Ni
N
i
=
f
′
(
u
i
)
∑
w
pi
p
p
>
i
FIGURE 3.8
Backpropagation structure. Node deltas are calculated in the backward
direction starting from the output layer.
Search WWH ::
Custom Search