Item

Abstract

The hippocampal formation is critical for spatial and relational inference in navigation problems. The neural code underlying such inference is factorized in the entorhinal cortex (EC) and conjunctive in the hippocampus (HC). A factorized code implies a separate encoding of sensory and relational knowledge, which can be flexibly conjoined to an object representation that reflects both sensory and relational properties. We hypothesize that the same neural mechanisms are employed in complex decision-making and compositional planning, which requires the flexible generalization of knowledge to novel instances. We tested this hypothesis in a task where subjects had to construct novel visual objects based on a set of basic visual building blocks and relations. We found behavioral evidence that subjects form a hierarchical representation of this task that allows them to flexibly apply compositional knowledge to novel stimuli. Using fMRI adaption, we found evidence that the construction of novel objects depends on compositional neural representations in HC-EC and medial prefrontal cortex (mPFC). Further, we found that these structures also encoded purely relational information, indicative of a factorized representation. These results suggest that compositional neural representations in the hippocampal formation and prefrontal cortex enable the generalization of abstract knowledge to novel stimuli during visual construction.