In response to the significant challenges faced by artificial intelligence techniques, particularly deep learning, in terms of interpretability, heavy data dependence, and limited robustness, the theoretical methodologies and recent advances of PL (prototypical learning) were systematically reviewed. By clarifying the fundamental concepts and mathematical formulations of prototypical learning, a prototype generation framework was established that encompasses statistical machine learning, deep feature-driven representations, and semantic representation perspectives. Prototype-based mechanisms for single-modal and multimodal data augmentation and fusion were analyzed, and the underlying rationale for overcoming data quality bottlenecks was elucidated. Particular emphasis was placed on the effectiveness of prototypical learning in interpretable deep neural networks, fuzzy rule-based reasoning, causal abduction, and time-series analysis. Furthermore, the evolutionary dynamics of prototypical learning across interdisciplinary domains, including generative learning, capability enhancement of large-scale models, and graph learning, were explored. By synthesizing the distinctive advantages of prototypical learning in representation efficiency and logical transparency, its critical role as a key enabling technology for constructing trustworthy and efficient artificial intelligence systems was highlighted. Finally, future development trends of prototypical learning were discussed, including directions related to generative artificial intelligence, collaboration with large models, and sustainable learning.