• It is set of connected i/o units in which each connection has a weight associated with it.
  • During the learning phase the network learns by adjusting the weights so as to be able to predict the correct class label of i/p labels.
  • It also referred as connectionist learning due to connection between units.
  • It has long training time and poor interpretability but has tolerance to noisy data. 
  • It can classify pattern on which they have not been trained.
  • Well suited for continuous valued i/ps.
  • It has parallel topology and processing.
  • Before training the network topology must be designed by:
  1. Specifying number of i/p nodes/units: Depends upon number of independent variable in data set.
  2. Number of hidden layers: Generally only layer is considered in most of the Two layers can be designed for complex problem. Number of nodes in the hidden layer can be adjusted iteratively.
  3. Number of output nodes/units: Depends upon number of class labels of the data
  4. Learning rate: Can be adjusted iteratively.
  5. Learning algorithm: Any appropriate learning algorithm can be selected during training phase.
  6. Bias value: Can be adjusted iteratively.
  • During training the connection weights must be adjusted to fit i/p values with the o/p values.