Appendix-Hard-AttentionforScalableImage Classification

We assume Satt 6=, and Sbbox 6=, meaning that for every image, we attend toatleast1location, and we haveabounding box with area greater than0.