A method of pruning layered feed-forward neural networks