Architecture (Radial Basis Function)

The Architecture tab is used to specify the structure of the network. The procedure creates a neural network with one hidden "radial basis function" layer; in general, it will not be necessary to change these settings.

Number of Units in Hidden Layer. There are three ways of choosing the number of hidden units.

1. Find the best number of units within an automatically computed range. The procedure automatically computes the minimum and maximum values of the range and finds the best number of hidden units within the range.

   If a testing sample is defined, then the procedure uses the testing data criterion: The best number of hidden units is the one that yields the smallest error in the testing data. If a testing sample is not defined, then the procedure uses the Bayesian information criterion (BIC): The best number of hidden units is the one that yields the smallest BIC based on the training data.

2. Find the best number of units within a specified range. You can provide your own range, and the procedure will find the “best” number of hidden units within that range. As before, the best number of hidden units from the range is determined using the testing data criterion or the BIC.
3. Use a specified number of units. You can override the use of a range and specify a particular number of units directly.

Activation Function for Hidden Layer. The activation function for the hidden layer is the radial basis function, which "links" the units in a layer to the values of units in the succeeding layer. For the output layer, the activation function is the identity function; thus, the output units are simply weighted sums of the hidden units.

• Normalized radial basis function. Uses the softmax activation function so the activations of all hidden units are normalized to sum to 1.
• Ordinary radial basis function. Uses the exponential activation function so the activation of the hidden unit is a Gaussian “bump” as a function of the inputs.

Overlap Among Hidden Units. The overlapping factor is a multiplier applied to the width of the radial basis functions. The automatically computed value of the overlapping factor is 1+0.1d, where d is the number of input units (the sum of the number of categories across all factors and the number of covariates).

How to Specify Architecture for Radial Basis Function

This feature requires the Neural Networks option.

1. From the menus choose:

   Analyze > Neural Networks > Radial Basis Function...

2. In the Radial Basis Function dialog box, click the Architecture tab.