In our experiments, the fine-grained map, global semantic map, and multi-granularity map are of different sizes (asshowninFigure A)forsaving GPU memory. Object categories predicted by hallucination module. We use an Adam optimizer with a learning rate of 2.5e-4. Specifically,we consider the 10% area with 2 the highest probability in 2D distributionP and ˆP (as described in Section 3.3) as ground-truth andpredicted locations. From Table 1,this variant performs worse than our agent.

We first assign bias classes usingAge and Race attributes. Specifically, for theAge attribute, we divide samples into two groups; bias class 0 for samples withage 20and bias class 1for samples withage 10.

Please refer to the Appendix 5 for details. Besides, in order to balance the classification loss and the transformation loss we set the scalar factor to beλ = 5 throughoutthetrainingphase. Here the first rows are the untransformed and the transformed images, while the second rows are the corresponding interpretations. This is a supplement toFigure 1 in the main body of the paper. And also there are perturbation methods such as randomized input sampling (RISE) [8] and extremal perturbation (EP) [2].