Category Learning

Category learning is the process by which cognitive agents — humans, animals, or artificial systems — acquire and organize concepts into categories that support inference, prediction, and action. It is one of the foundational problems of cognitive science, connecting psychology, neuroscience, machine learning, and philosophy of concepts.

Unlike supervised learning in machine learning, where categories are given by an external teacher, category learning in natural systems is often unsupervised or weakly supervised. The learner must discover the relevant dimensions of variation, the appropriate level of abstraction, and the boundaries between categories from sparse and noisy data.

Major theoretical frameworks include:

Prototype theory: categories are represented by central tendencies, and classification is based on similarity to the prototype.
Exemplar theory: categories are represented by stored instances, and classification is based on similarity to individual exemplars.
Rule-based learning: categories are defined by explicit rules or decision boundaries.
Statistical learning: categories emerge from the statistical structure of the input, often without conscious awareness.

The adaptive resonance framework provides a neural mechanism for category learning that balances stability and plasticity. The deep learning approach uses hierarchical representations learned through gradient descent. Each framework makes different commitments about the nature of concepts, the role of similarity, and the mechanisms of generalization.

Category learning is not merely a cognitive convenience. It is a compression mechanism that allows agents to generalize from limited experience. The categories an agent learns determine what it can perceive, what it can predict, and what it can do. In this sense, category learning is a form of world-model construction — the agent is not merely labeling objects but building a representation of the structure of its environment.