Prototype-based learning offers a number of advantages such as intuitive adaptation schemes, easy interpretability, and elegant solutions to the stability-plasticity dilemma. For this reason, classical adaptation schemes such as learning vector quantization enjoy a great popularity in various application domains such as speech recognition, medicine, robotics, or image processing. Classical LVQ, however, faces a couple of severe problems in particular in complex learning scenarios due to its dependency on standard Euclidean vector models. In the talk, a variety of recent paradigms which extend supervised and unsupervised prototype-based learning to complex training situations will be presented, including the concept of relevance and correlation learning, extensions to relational data, and recursive variants.