Four knowledgeable referees support acceptance for the contribution; they like the authors' novel idea of Glance and Focus net and ImageNet scale evaluations showing superior accuracy-efficiency tradeoff against state-of-the-art-baselines. Please make it sure to properly include and discuss missing references and experimental comparisons, as promised in the rebuttal. [1] (and related works) is also related from another perspective and should be properly discussed in related work - it is using early-exist inference methods via RL but for missing data (if image patches in different resolutions are understood as 'missing' features, then they share the same spirit). [1] Shim et al. Joint Active Feature Acquisition and Classification with Variable-Size Set Encoding. NeurIPS 2018