Task--Based Pruning

A pruning method is presented that is applicable to a one hidden layer classification network, where each unit of that layer can be conceived as performing one discrimination task. The method, termed "task--based pruning", depends on selecting a subset of units that will satisfactorily perform the necessary tasks. 1 Introduction In general, a multi--layer perceptron is started off with an arbitrarily chosen architecture and random weights. However in the case of a classification network with one hidden layer, it is possible to select, in an obvious way, an architecture that is designed specifically for the problem. This can be done by ensuring that for each pair of classes there is at least one separating hyperplane. 1.1 Single layer learning Knerr et al [1] present such a network using subnets of different conceptual interpretation. Their network relies on each unit in a layer starting off with random weights but, unlike a standard multi--layer perceptron, each unit responds only to..