Skip to Book Content
Book cover image

Chapter 16 - Heuristics for Improving Generalization

Neural Smithing: Supervised Learning in Feedforward Artificial Neural Networks
Russell D. Reed and Robert J. Marks II
Copyright © 1999 Massachusetts Institute of Technology
 

16.4 Constructive Methods

Pruning methods train a larger-than-necessary network and then remove unneeded elements. The opposite approach is to build a network incrementally, adding elements until a suitable configuration is found. The two approaches are complementary and often used together. Like pruning, constructive techniques are a means of adjusting the size of a network rather than a method for deciding what size is appropriate. Other criteria are still necessary to decide when to stop adding elements. A number of constructive methods are discussed in chapter 12. Cascade-correlation [120] is often cited as an example.