Evaluating pruning methods
Georg Thimm, Emile Fiesler · 1995
: A notorious problem in the application of neural networks is to find a small suitable topology. High order perceptrons already solve a part of this problem as they require no hidden layers. However, the number of connections in a fully interlayer connected high order perceptron grows quickly with their order. Partially connected topologies are therefore highly desirable and can be obtained by applying em connection pruning methods A framework is provided here that allows a practical comparison of pruning methods which is based on final network size and generalization capability, but also considers the total training time. This framework is applied to the comparison of an easy to implement, low complexity method with four other pruning methods of similar complexity. Keywords: pruning, generalization, optimality criteria, high order perceptrons, backpropagation neural networks. Introduction A mayor problem in the application of neural networks is the choice of a topology: a consider...