A Supplementary Material

Similar to Figure 1, dashed lines in (a) represent results with full confusion matrices estimated from smoothed counts and solid lines represent using a less accurate confusion matrix estimate where only diagonals are estimated from data. We observe very similar behaviour to that described for ImageNet and ObjectNet in 6.2 and 6.3, respectively. The most interesting difference compared to our experiments on ImageNet and ObjectNet, is that, on CIFAR-10, our methods achieve substantially larger accuracy gains compared with the baselines, even for small budgets.