Explicit regularization and implicit bias in deep network classifiers trained with the square loss.