2a084e55c87b1ebcdaad1f62fdbbac8e-AuthorFeedback.pdf

As shown in Supp Figure 1, the distribution becomes sharper during training. We follow DoReFa and set the value range to[ 1,1]. When the bit-wise is higher than 1, the37 accuracy might be higher because the information loss is small. We will cite [r4] and discuss the differences between them.51 Moreover, [r4] can only work for binary codes (see Eq. (4) in [r4]) and focus on58 hashing comparison experiments.