In supervised learning, it has been shown that label noise in the data can be interpolated without penalties on test accuracy under many circumstances. We show that interpolating →