Deep neural networks can generalize well even when they're far into the overparametrized regime where classical statistical learning theory predicts overfitting

Deep neural networks can generalize well even when they're far into the overparametrized regime where classical statistical learning theory predicts overfitting. This is usually called "double descent" and there are many papers on it.