A More Related Work

There have been some efforts trying to improve contrastive learning from various perspectives, e.g., Our method adopts a similar idea to the CL setting to address one missing piece of the framework . 's are independent with each other given the similarity scores. The max-margin contrastive loss is defined on pair-wise data, thus is less affected by the negative samples. However, the loss is shown to be less effective compared to the fully coupled NT -Xent loss [13]. Please refer to the "Simulation" paragraph in Section 2.2.3 for details Thus, it is feasible to evaluate. For clearness, we redefine some terms used in our proof.