An Artificial Neural Network (ANN) is an adaptive computing paradigm that aims to solve problems by emulating how a biological nervous system processes information.  Just like its biological counterpart, it consists of a large and highly-interconnected group of processing elements called neurons that can change in structure as it encounters new information while learning to solve a problem. An artificial neuron has many inputs but just a single output whose state depends on the inputs to the neuron.

An ANN has to be configured to solve a certain problem, which means that it has to learn what to do first before it can be used in information processing. It is not programmed algorithmically on what to do. ANNs learn by example like living organisms. And in a similar manner too - by adjusting interconnections among its neurons whenever called for by new external and/or internal information.

The structure of the most common type of ANN has three layers: 1) the input layer; 2) the hidden layer; and 3) the output layer. The input layer consists of nodes or neurons that receive inputs or raw information from outside.  These raw information are passed on to the neurons of the hidden layer for processing in a manner dictated by the existing interconnections among the neurons. Finally, the output layer receives the hidden layer's information, finalizes the processing, and displays the appropriate output.

The processing of information from the input layer to the output layer consists of many occurrences of an event called 'neuron firing'.  'Firing' a neuron simply means triggering it to change its state based on changes in the states of the neurons attached to its inputs.  A change in the state of this neuron may then affect the state of the neurons whose inputs are designed to be triggered by it.  The 'intelligence' in the firing of a neuron is provided by the firing rule for each neuron, which determines whether the neuron should fire or not for any given input pattern. This sequential firing of neurons sweeps across the ANN from the input to the output in many parallel paths in the hidden layer.

Since uncertainty factors can be incorporated in firing rules, neural networks can handle imprecise data.  As such, it can be used to detect patterns or forecast trends that humans or other computing techniques may not notice.  ANN's can be a very powerful tool for 'what-if' questions that involve the analysis of many data that have complex or fuzzy relationships with each other.

Advantages of ANNs over other computing models include the following: 1) ability to exhibit adaptive learning, i.e., learning to do a task based on training and experience; 2) ability to represent and organize information by itself;  3) ability to generate outputs in real time as new inputs are being delivered; 4) ability to tolerate faults or partial damage to the network.

See Also:   More Industry Articles