Human latent-state generalization through prototype learning with discriminative attention

Latent causes that give rise to experience are encountered in complex, high-dimensional feature spaces. How then do people approximate the external world with lower-dimensional internal representations that generalize to novel examples or contexts? Theories suggest internal representations could be determined by discriminative boundaries, or based on the distance from prototypes/exemplars. We developed theoretical models that use both discriminative and prototype/exemplar components to form internal representations via action-reward feedback. We then developed three new latent-state learning tasks to test human use of discrimination attention and prototypes/exemplars. The majority of subjects attended to discriminative features, as well as the covariance of features within a prototype. A minority of subjects relied on a single discriminative feature. Behavior of all subjects was captured by a model that forms prototype representations and deploys context-specific discriminative attention. These results provide insights into the human ability to generalize across causal latent states learned in high-dimensional environments.