prototypical contrastive learning

**Prototypical Contrastive Learning (PCL)** is a **self-supervised method that bridges instance-level contrastive learning with semantic-level clustering** — by using cluster prototypes as positive targets, encouraging all instances within a cluster to have similar representations. **How Does PCL Work?** - **Standard Contrastive**: Each image is its own class (instance discrimination). - **PCL Enhancement**: Run clustering (k-means or EM) on the learned features periodically. Use cluster assignments to define additional positive pairs. - **Loss**: Combines instance-level InfoNCE loss with prototype-level contrastive loss. - **Prototypes**: Cluster centroids updated periodically during training. **Why It Matters** - **Semantic Grouping**: Goes beyond instance discrimination to learn category-level similarities. - **Fewer False Negatives**: In standard contrastive learning, two images of the same class are treated as negatives. PCL corrects this. - **Transfer Learning**: Better downstream performance on tasks requiring semantic understanding. **PCL** is **contrastive learning with semantic awareness** — using clustering to teach the model that different instances of the same concept should share similar representations.