ZPBerg

In the paper, you state "We estimate the variance by averaging over N = 3 such random splits (i.e., we train 6 = 3 × 2 copies of each model)."...