dc49dfebb0b00fd44aeff5c60cc1f825-Supplemental.pdf

AAblationStudy1 In this ablation study, we further investigate the power of the policies searched by our approach2 andtheclosely related method AutoAug [1]. Thenwegradually removethemostimportant operations fromthe5 searched policy one by one and investigate the change of the Top-1 test error rates, as reported in6 Tab.1. Figure 1: We investigate the key hyper-parameterNlate by visualizing the difference it brings to thesearchdynamics. We investigate different numbers of epochs in the late training stage(Nlate). By adjusting Nlate20 we can still maintain the reliability ofpolicyevaluation toalarge extent.