Dissecting Pruned Neural Networks