Compositional learning is a machine learning paradigm where a system learns to recognize and process complex concepts by breaking them down into basic constituent components and learning how those components combine. Instead of treating every unique configuration or interaction as an isolated class, the model learns separate representations for primitive elements, such as objects, actions, or visual attributes. This structural decomposition enables the system to generalize across combinatorial variations and recognize novel, unseen combinations of familiar components during inference. Consequently, compositional learning significantly improves sample efficiency and facilitates zero-shot generalization in computer vision, natural language processing, and multimodal tasks where collecting exhaustive training data for every possible combination is impractical.
