Neural Network Structure
Although neural networks impose minimal demands on model structure and assumptions, it is useful to understand the general network architecture. The multilayer perceptron (MLP) or radial basis function (RBF) network is a function of predictors (also called inputs or independent variables) that minimize the prediction error of target variables (also called outputs).
Consider the bankloan.sav dataset that ships with the product, in which you want to be able to identify possible defaulters among a pool of loan applicants. An MLP or RBF network applied to this problem is a function of the measurements that minimize the error in predicting default. The following figure is useful for relating the form of this function.
This structure is known as a feedforward architecture because the connections in the network flow forward from the input layer to the output layer without any feedback loops. In this figure:
- The input layer contains the predictors.
- The hidden layer contains unobservable nodes, or units. The value of each hidden unit is some function of the predictors; the exact form of the function depends in part upon the network type and in part upon user-controllable specifications.
- The output layer contains the responses. Since the history of default is a categorical variable with two categories, it is recoded as two indicator variables. Each output unit is some function of the hidden units. Again, the exact form of the function depends in part on the network type and in part on user-controllable specifications.
The MLP network allows a second hidden layer; in that case, each unit of the second hidden layer is a function of the units in the first hidden layer, and each response is a function of the units in the second hidden layer.