Below is a pictorial representation of a 3-layered neural network.
"Neural networks were initially studied by computer and cognitive scientists
in the late 1950s and early 1960s in an attempt to model sensory perception
in biological organisms. Neural networks have been applied to many problems
since they were first introduced, including pattern recognition, handwritten
character recognition, speech recognition, financial and economic modeling,
and next-generation computing models." [encarta 99]
A neural network can be
either feed-forward only or have the ability to backtrack. Feed-forward
networks are relatively simple to understand. Information is passed
in a straight-forth manner from one node to its connecting node.
At each juncture a threshold is either reached and the data continues on
to its connecting node, or the threshold is not reached; thus the connecting
node receives a null value which prevents that node from being activated.
However, in a network that is able to backtrack, receiving nodes are able
to send back to its preceding node the result it has been sent. This
allows a hidden-node (hidden-cell) the ability to return a value that does
not met its threshold back to its connecting input-node. The input-node
can then try sending the data along its connection paths to an alternative
node. [Graci]
Nodes may be either discrete or continuous. "Discrete nodes send an output signal of 1 if the sum of received signals is above a certain critical value called a threshold value, otherwise they send an output signal of 0. Continuous nodes are not restricted to sending output values of only 1s and 0s; instead they send an output value between 1 and 0 depending on the total amount of input that they receive—the stronger the received signal, the stronger the signal sent out from the node and vice-versa. Continuous nodes are the most commonly used in actual artificial neural networks." [Encarta 99]
Neural
networks first go through a training phase where the network establishes
the associated weight between each connection. This is done by first
inputting a data set with known output results. After each trial,
the weights are slightly adjusted until the network produces 100% accuracy.
After the network has been trained it is then ready to be placed into production
to be used in application procedures with unknown answers. [Graci]