Unseen interactions refer to novel relationships, behaviors, or pairings between entities, such as specific combinations of actions and objects, that a machine learning model did not encounter during its training phase. Because the possible combinations of entities and actions are combinatorially vast, visual recognition systems frequently encounter configurations outside their predefined training sets. Recognizing unseen interactions typically relies on zero-shot learning and compositional generalization, where models utilize semantic knowledge, language embeddings, or shared feature representations to recognize familiar individual components and successfully infer novel, composite relationships without prior direct supervision.
