Towards an Understanding of Benign Overfitting in Neural Networks