conceptlearning

Concept learning is the task of inferring a rule or concept that separates positive from negative instances based on observed examples. It is central to both cognitive science and machine learning, informing how people and machines categorize the world and generalize to new data.

In cognitive science, concept learning examines how humans form and refine mental representations of categories. Theories address how features are weighed, how prototypes or exemplars are used, and how new instances are mapped to existing concepts. Challenges include dealing with noise, incomplete information, and variability within categories.

In machine learning, concept learning is typically framed as a supervised learning problem: given a set of labeled examples, the goal is to identify a hypothesis that correctly labels all observed instances. A formal view treats a concept as a subset of an instance space, with the concept class representing possible subsets. Classic approaches include the version-space algorithm and other rule- or tree-based learners that search for hypotheses consistent with the data. The field also encompasses broader frameworks such as Probably Approximately Correct (PAC) learning, which analyzes sample complexity and generalization.

Representations used in concept learning range from simple boolean attribute conjunctions to more complex rules or decision trees, and extend to multi-valued and real-valued features. Although modern methods often operate as general supervised learning or deep learning models, the underlying goal remains the same: to induce a compact, generalizable model of a target concept from examples.

Applications include text and image classification, medical diagnosis, and customer segmentation, while key challenges involve noise, high dimensionality, concept drift, and limited labeled data.