Exploring Negatives in Contrastive Learning for Unpaired Image-to-Image Translation
Yupei Lin, Sen Zhang, Tianshui Chen, Yongyi Lu, Guangping Li, Yukai Shi
Proceedings of the 30th ACM International Conference on Multimedia (ACM MM 2022)
Figure 1: Left: The PatchNCE loss [27]. Right: the proposed RankNCE loss. A typical patch-wise contrastive paradigms usually maximize the mutual information on all the samples, regardless of positive or negative. In comparison, RankNCE selects negative samples with good qualities based on their contribution to the mutual information (MI).
Overview
Studies whether all negative patches are necessary in contrastive learning and introduces negative pruning for unpaired image-to-image translation.
