Note that only the samples that the predictions of the network were correct are counted for a fair evaluation. Since segmentation labels are provided as 0's and 1's, it is possible to evaluate the quality of attribution maps as abinary classification task. This process can be repeated with different thresholds to produce a ROC curve. These examples also predict the correct label (person, horse, cow,train, bus, cat). Finally, a separate classifier is retrained on this perturbed dataset.

This phenomenon arises due to the fact that conventional network compression methods only preserve the predictions of the network while ignoring the quality of the attributions.

One of the principal assumptions of the GP-LVM is that the prior p(f) regularises the smoothness of all mappings equally, so we only consider onekernel.

Unlike the uniform application of asingle penalty,CPR enforces an upper bound on astatistical measure, suchas theL2-norm, ofindividual parameter matrices.

The first method, explained in Section A1.4.1, consists of directly calibrating a sequence of nested two-sided intervals, as outlined in Section 3.3. The second method, explained in Section A1.4.2, consists of separately calibrating two sequences of lower and upper one-sided confidence intervals, each adopting the significance level α/2 instead of α. Pu j=l ˆϕj(x)amongthefeasible ones with minimal |u l|, whenever the optimization problem does not have a unique solution. Therefore, we can assume without loss of generality that (1) has a unique solution; if that is not the case, we can break the ties at random by adding a little noise to ˆϕ. For any integer T 1, consider an increasing sequence tτ [0,1], for τ {0,...,T}. A nested sequenceofT intervalsindexedbyτ {0,...,T},whichmaybewrittenintheformof St = ˆLm,α(Xm+1;tτ), ˆUm,α(Xm+1;tτ), for appropriate lower and upper endpoints ˆLm,α(Xm+1;tτ) and ˆUm,α(Xm+1;tτ), respectively, is then constructed from (1) as follows.

A flexible conformal inference method is developed to construct confidence intervals for the frequencies of queried objects in very large data sets, based on a much smaller sketch of those data.

This approachyields biased estimates ofthe expected risk, and therefore leads to poor decisions for two reasons.