You may create the iris example set with the Create iris example... command that you will find under the Feedforward neural networks option in the New menu. Three new objects will appear in the List of Objects: a FFNet, a Categories and a PatternList.
The PatternList contains the iris data set in a matrix of 150 rows by 4 columns. To guarantee that every cell in the PatternList is in the [0,1] interval, all measurement values were divided by 10. In the Categories the three iris species setosa, versicolor, and virginica were categorized with the numbers 1, 2 and 3, respectively. Because there are 4 data columns in the PatternList and 3 different iris species in the Categories, the newly created FFNet has 4 inputs and 3 outputs. If you enter a positive number in one of the fields in the form, the FFNet will have this number of units in a hidden layer. The name of the newly created FFNet will reflect its topology. If you opt for the standard setting, which is 0 hidden units, the FFNet will be named 4-3.
The first thing you probably might want to do is to let the FFNet learn the association in each pattern-category pair. To do this select all three objects together and choose Learn.... A form will appear, asking you to supply some settings for the learning algorithm. Learning starts after you have clicked the OK button. As the example network has only a small number of weights that need to be adjusted, and the learning data set is very small, this will only take a very short time.
Now, if you are curious how well the FFNet has learned the iris data, you may select the FFNet and the PatternList together and choose To Categories.... A new Categories appears in the List of Objects with the name 4-3_iris (if 4-3 was the name of the FFNet and iris the name of the PatternList). We have two different Categories in the list of objects, the topmost one has the original categories, the other the categories as were assigned by the FFNet classifier. The obvious thing to do now is to compare the original categories with the assigned categories by making a confusion table. Select the two Categories and choose To Confusion and a newly created Confusion appears. Pressing the Info button will show you an info window with, among others, the fraction correct.
You may also want to compare the FFNet classifier with a discriminant classifier.
With a PatternList and a Categories selected together, you can for example create a new FFNet of a different topology.
© djmw, April 26, 2004