A modified version of a formal pruning algorithm based on local relative variance analysis
Nader Fnaiech, S. Abid, Farhat Fnaiech, Mohamed Cheriet · 2004
A modified version of a formal pruning algorithm initially proposed by Englebercht [November, 2001] using variance analysis of sensitivity is presented. We propose a new modification of the algorithm by applying the pruning procedure on each layer starting from the output layer to the input layer. Contrarily, to the work of Englebercht where the pruning is performed on the entire net that we denote in this paper global pruning, we shall prune layer by layer with the use of a pruning decision based on a local parameter variance nullity coefficient (LPVN). These coefficients are then classified in an ordered list which allows the decision making examples showing that in some cases we can reach about 30% improvement in terms of coefficients and neurons removal in order to get the best neural network pruned. A comparison study is given on some real world learning and generalization.