Abstract

Learning from experience and adapting to new situations is fundamental to human development and wellbeing. Research on human learning has primarily focused on domain-specific skills (e.g. learning language, sports). However, a wide range of everyday behaviours—from detecting targets in clutter (e.g. finding a friend in the crowd), to navigating a new city, and co-ordinating our social interactions—critically depend on extracting meaningful structure from an initially incomprehensible stream of events. Extracting meaningful structures is critical for optimising not only our ability to interpret incoming information but also to predict upcoming events and react successfully to novel situations. We refer to this domain-general skill of extracting the principles of organisation that determine the structure of the environment, as generative learning. To understand individual ability for generative learning, we combine behavioral and brain imaging measurements with computational modeling. Our work determines prototypical strategies for generative learning and links individual learning strategies to brain computations. Our behavioural findings suggest that successful learning relies on extracting behaviorally relevant structures that are predictive of upcoming events rather than simply memorizing all possible statistical dependencies. Further, we provide evidence that generative learning is implemented by a cortico-striatal circuit that may select, monitor and adjust individual learning strategies over time. We propose that these brain computations may determine domain-general strategies for generative learning that can become specialized to support higher cognitive functions, such as navigating a new city, understanding music and language.