MULTI-DAY AND MULTI-MODAL CONCEPT LEARNING IN HUMANS: BEHAVIORAL DYNAMICS AND COMPUTATIONAL MODELING OF EXEMPLAR–PROTOTYPE REPRESENTATIONS

Concept learning requires mapping high-dimensional sensory input onto behaviorally relevant categories. While human learners can generalize beyond experienced examples, the representational transformations that support this ability over extended learning remain poorly characterized. Here, we present behavioral results from two five-day concept learning studies that track how conceptual representations evolve over extended training, complemented by computational modeling.

Across both studies, participants learned to categorize novel, multimodal stimuli ("Ferni" monsters) into two families defined by latent feature regularities. Learning happened over five consecutive days with instant feedback-based training, formation of concept representations was assessed by testing categorization performance for new generalization items. We examined how behavior and concept representations are influenced by training set size and motivational context (study 1), and by the presence or absence of category-irrelevant perceptual features (study 2).

Behavioral performance during training and generalization is analyzed using models of concept representation such as exemplar-based, prototype-based, and mixture models. This modeling approach allows us to examine the representational structure (exemplar vs. prototype-based) of conceptual knowledge. Preliminary analyses indicate robust improvements in generalization across days and systematic shifts in representational structure.

Ongoing work extends these behavioral analyses with biologically grounded modeling to test whether observed learning trajectories can arise from complementary learning mechanisms operating at different timescales. The findings yield detailed insights into the long-term evolution of concept representations. Our work further provides a foundation for future integration with neural data to examine how evolving behavioral representations relate to hippocampal–neocortical learning systems.