Wethankallthereviewersfortheireffortsinreviewingourpaper. Wefirstaddressthecommonconcernofallreviewers.1 Common2

We use the classification loss (categorical cross entropy) as an evaluation25 measure (L in Equation 3) for each candidate policy. The FAA is able to select "Cutout", since "Cutout" can26 (probabilistically) eliminate irrelevant backgrounds and improve the classification accuracy when the inference is27 performed ona(well-) trained network. Therefore,we44 can consider the number of sub-policies as ahyperparameter to tune, since the training time overhead by increased45 number ofsub-policies isalso limited asshowninthebelowexplanation. Having this inmind, weperformed FAA46 with different numbers of sub-policies and determined the number of sub-policies that produces the best average47 performances across different datasets and networks.