Understanding Neural Networks

An Artifical Neural Network (ANN) is an information processing paradigm that is inspired by biological nervous systems. It is composed of a large number of highly interconnected processing elements called neurons. An ANN is configured for a specific application, such as pattern recognition or data classification.

Neural Networks have the ability to derive meaning from complicated or imprecise data, extract patterns and detect trends. It is used for Adaptive Learning and Real Time Operation. Conventional Computers use an algorithmic approach, but neural networks work similar to the human brain and learns by example.

Neural Networks(NN) are successfully being used in many areas in connection with Artificial Intelligence.

A classic application for NN is image recognition. A network that can classify different standard images can be used in several areas such as Medical diagnostics, by classifying x – ray pictures for tumor diagnosis, Detective tools, by classifying fingerprints to a database of suspects, etc.

A well known application using image recognition is the Optical Character Recognition (OCR) tools that we find available with the standard scanning software the home computer.

Another popular application for NN is Customer Relationship Management (CRM). Many companys have at the same rate as electronic data storage has become commonplace built up large customer databases. By using Neural Networks for data mining in these databases, patterns however complex can be identified for the different types of customers, thus giving valuable customer information to the company.

One example is the airline reservation system which could predict sales of tickets in relation to destination, time of year and ticket price. The NN strategy was well suited for the purpose because the system could be updated continuously with the actual sales.

To evaluate whether a problem is suitable for a Neural Network implementation:

• There must be a large example dataset of the problem in order to be able to train the network.
• The data relationships in the problem are complex and difficult or impossible to program using conventional techniques.
• The output does not need to be exact or numeric.
• The desired output from the system changes over time, so a high flexibility is needed.

NN uses backward propagation to adjust the weights between the neurons. The actual technique is a black box as we don’t know how the exact processing takes place but it is learning from the output to the input. The output should be equal to the estimation and if not then that is a possible degree of error. For example classifying a picture as an apple.

R has a few packages for creating  neural network models (neuralnet, nnet, RSNNS). The functions in this package allow you to develop and validate the most common type of neural network model, i.e, the feed- forward multi- layer perceptron. he functions have enough flexibility to allow the user to develop the best or most optimal models by varying parameters during the training process. One major disadvantage is an inability to visualize the models. In fact, neural networks are commonly criticized as ‘black-boxes’ that offer little insight into causative relationships among variables. In recent research several approaches have since been developed to ‘illuminate the black-box’.

Image: courtesy Wikipedia