To keep experiments uniform, for all datasets (STL-10, CIFAR-10, and CIFAR-100) we used a train/val/test partitioning. In our experiments we compared FED with four baselines. For all baselines we tried different learning rates [0.1, 0.01, 0.001] and batch sizes [32, 64, 100]. For EnDD and EnDD + AUX, we used the same temperature, temperature annealing, and optimizer that was used in the original paper. For AMT, we tried different alphas [1e1, 1e3, 1e5] and kept the rest as the original paper.

Bayesian models have many desirable properties, most notable is their ability to generalize from limited data and to properly estimate the uncertainty in their predictions. However, these benefits come at a steep computational cost as Bayesian inference, in most cases, is computationally intractable. One popular approach to alleviate this problem is using a Monte-Carlo estimation with an ensemble of models sampled from the posterior. However, this approach still comes at a significant computational cost, as one needs to store and run multiple models at test time. In this work, we investigate how to best distill an ensemble's predictions using an efficient model.

ForDM, due to high memory requirements, we were able to go up to aBatchEnsemble with an ensemble size of 8, while being able to use only batch size of 32. In addition, for this baseline we used a bigger memory GPU, unable tofitthetraining toourstandard 11GBGPU usedfortherestofour experiments. In the procedure of creating a Mixup [8] auxiliary dataset, we used a Beta distribution withα = 0.2. In Mixup augmentation, and valueλ [0,1] is sampled from a Beta distribution. We use batch size of 64.

One popular approach to alleviate this problem is using a Monte-Carlo estimation with an ensemble of models sampled from the posterior. However, this approach still comes at a significant computational cost, as one needs to store and run multiple models at testtime.