Prototype-Based Learning

Prototype-based learning represents classes or patterns with learned or data-derived examples in a feature space, then predicts by measuring similarity to those prototypes. Common approaches include prototypical networks for few-shot learning, contrastive prototype losses, memory banks, and alignment between visual and text prototypes. Researchers use these methods for settings such as medical segmentation, domain adaptation, anomaly detection, and continual or federated learning, where labeled data or stable representations may be limited. A central challenge is that a small set of prototypes can miss within-class variation or become unreliable under domain shift, so current methods increasingly refine, diversify, or calibrate prototypes while balancing accuracy, interpretability, and adaptation stability.