Similar Items: Random Pruning Over-parameterized Neural Networks Can Improve Generalization: A Training Dynamics Analysis