466473650870501e3600d9a1b4ee5d44-Supplemental.pdf

It clearly shows thatBEstudents distilled withODSare better calibrated than the baselines. The difference is that the adversarial perturbation is meant to worsen the predictive performance by design because it takes astep toward the directions increasing the classification loss. Weempirically found that the perturbations just increasing diversity without maintaining prediction accuracy can actually harm the performance of student models (this is also related to the performance gain of ConfODS). Table 4: Knowledge distillation fromDE-3 into MIMO-3: ACC, standard metricsand calibrated metrics.