How practical are VC dimension bounds? [learning from examples]

In recent years computational learning theory has provided significant insights into the process of learning from examples, and in particular the property of generalization. However, little attempt has been made to assess the available theoretical results from a practical point of view. This article addresses the question of whether some recent bounds based on the Vapnik-Chervonenkis (VC) dimension can in any way be applied as a practical design tool. It is found that although these bounds offer significant improvements when compared with earlier results, and can offer a limited guide to the design of pattern classifiers in certain circumstances, they still have important practical shortcomings that must be addressed before a truly useful theory is obtained