Pruning of Neural Networks

Choosing a suitable topology for a neural network, given an application, is a difficult problem. Usually, after a tedious trial-and-error process, an oversized topology is chosen, which is prone to various drawbacks like a high demand on computational resources and a high generalization error. A way to solve this is to trim the network size during the training process. This is done with so-called pruning methods, of which an overview is given. From these methods, those that are potentially suitable for high order perceptrons are selected, and then adapted accordingly. Next, they are tested on a variety of benchmarks by means of a large number of experiments. The conclusions are both of a generic nature, pointing out some pitfalls of neural network pruning in general, and of a more specific nature, identifying the best pruning methods for high order perceptrons.

Related material