Inverse Optimal Transport with Application to Contrastive Learning

Liangliang Shi, Jingtao Fan, Junchi Yan

Keywords: Contrastive Learning, Inverse Optimal Transoprt

Abstract: Previous works in contrastive learning (CL) mainly focus on pairwise views to learn the representations by attracting the positive samples and repelling negative ones. In this work, we understand the CL with a collective point set matching view and solve this problem with the formulation of inverse optimal transport(IOT), which is a min-min optimization to learn the features. By varying the relaxation degree of constraints in inner minimization of IOT, one can naturally get three different contrastive losses and reveal that InfoNCE is a special case of them, which shows a new and more generalized understanding view of CL. Besides, with our soft matching view, a uniformity penalty is also proposed to improve the representation learning. Experimental results show the effectiveness of our methods.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning