Biology Reference
In-Depth Information
Fig. 2.
is a schematic representation of a feed-forward MLP. Compu-
tational nodes (neurons) are represented as circles with weighted inputs and
output shown as arrows. Also shown, are the three common types of non-
linearity that can be used by the neurons to determine output: hard limiter,
sigmoid and threshold.
architecture, the multiple layer perceptron (MLP) (Fig. 2), neurons
are classified into three types: 1)
input neurons
, responsible for receiv-
ing and passing the signal vector for mathematical processing; 2)
hid-
den neurons
, in which the information received from the input layer
undergoes further processing; and 3)
output neurons
, which produce
the system's output or result. The neurons are arranged into layers
with corresponding names.
Training the ANN.
The network can be trained using either of
two types of learning techniques:
supervised learning
or
unsupervised
learning
. In the supervised learning paradigm, the system is presented
with a set of input vectors that are arbitrarily matched or paired to a
set of corresponding output vectors. Once the learning process is
over, the trained ANN generates outputs to categorize or approxi-
mate, depending on the nature of the task, any given new input vec-
tors. In this type of learning, the ANN is trained with a “teacher.”
Figure 1(B) shows the basic scheme used for this type of learning. In
the unsupervised learning paradigm (training “without a teacher”),
the ANN is presented with only input vectors. The ANN then
becomes tuned to the statistical regularities in the input data and
