Скачать с ютуб Paper review:X-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs в хорошем качестве

Paper review:X-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs

The paper proposes X-Sample Contrastive Loss (XSCL), an enhancement to the standard contrastive learning paradigm. The main motivation behind this work is to improve representation learning by incorporating cross-sample relations into the learning process. Conventional contrastive learning methods, like SimCLR and MoCo, typically rely on instance discrimination, which focuses on pulling positive pairs together (augmentations of the same sample) and pushing negative pairs apart (augmentations from different samples). However, these methods often overlook the relationships between different negative samples, which might result in suboptimal representations, especially when some negative pairs share similarities in their underlying data structure.