Information Technology Reference
In-Depth Information
Note that values of units are arbitrary but they should range from 0 to 1
(sometimes -1 to 1 range). In general, every unit has following aspects:
−
A set of inputs connects to it. Each connection is defined by a weight
−
Its weighted sum is computed by summing up all the inputs modified by their
respective weights
−
A bias value is added to weighted sum. This weighted sum is also called
activation value.
−
Its output is the outcome of activation function on weighted sum. Activation
function is crucial factor in neural network.
Activation function is the squashing function which “squashes” a large weighted
sum into possible smaller values ranging from 0 to 1 (sometimes -1 to 1 range).
There are three types of activation function:
−
Threshold function
takes on value 0 if weighted sum is less than 0 and
⎧
0
if
x
<
0
otherwise. So
μ
(
x
)
=
⎩
⎨
1
if
x
>
0
−
Piecewise-linear function
takes on values according to amplification factor in a
1
⎧
0
if
x
≤
−
⎪
⎪
⎪
⎨
2
1
1
certain region of linear operation. So
μ
(
x
)
=
x
if
x
=
−
<
x
<
2
2
⎪
⎪
⎪
1
1
if
x
≥
2
⎩
−
Sigmoid function
takes on values in range [0, 1] or [-1, 1]. The formula of
1
sigmoid function which is
μ
(
x
)
=
1
.
+
e
−
x
Fig. 8
Sigmoid function
There are two topologies of neural network:
−
Feed-forward neural network
. It is directed acyclic graphic in which flow of
signal from input units to output units is one-way flow so-called feed-forward.
There are no feedback connections