The paper studies how to find useful subnetworks within randomly initialized neural networks without training their weights.