Item

Abstract

The hippocampus forms concepts by integrating multi-feature relations into a unified representation. A common yet unconfirmed assumption is that such cognitive maps afford interpolations to never-experienced states. We approach this question as a category-learning problem in which prototypes are omitted from training but guide category-based decisions in a subsequent feature-inference task. Consistent with behavior, missing inferred stimulus features were represented at prototypical values in neocortex. This cortical completion effect correlated with hippocampal responses, which in turn reflected the distance between imagined prototypes and experienced exemplars. This was paralleled by a learning-dependent grid-like representation of the underlying concept space in entorhinal cortex. Our results suggest that abstracted prototypes correspond to interpolated central states in a cognitive map that guide cortical pattern completion during category-based decisions.