Connecting Multi-modal Contrastive Representations

1.Motivation:

Current Multi-modal Contrastive Representation (MCR) learning relies on massive high-quality data pairs, which limits its further development on more modalities.
当前的多模态对比表示（MCR）学习依赖于大量高质量的数据对，这限制了其在更多模态上的进一步发展。

2.Challenges:

1. Embeddings in MCR spaces are incapable of comprehensively reflecting all the semantic information of the input.
   MCR空间中的嵌入无法全面反映输入的所有语义信息。

2. MCR spaces exhibit a modality gap phenomenon, i.e., the embeddings of different modalities are located in two completely separate regions in each MCR space.
   MCR空间表现出模态间隙现象，即不同模态的嵌入位于每个MCR空间中两个完全独立的区域。

3.Contribution:

1. Propose C-MCR, a novel paired-data free and training-efficient method for MCR learning.
   提出了一种新的无数据配对、高效训练的MCR学习方法C-MCR。
2. Pr