Leveraging logit uncertainty for better knowledge distillation.

The accuracy gain of 3.57% (from 75.11% to 78.68%) achieved by our LUD method over the best-performing existing method (CRD) is indeed practically significant.

also in 1,558 other papers