High-dimensional hippocampus representation facilitates learning of low-dimensional abstract representation and supports generalization

Abstract

Semantic knowledge encompasses facts, concepts, and their relationships. Recent work has proposed that some forms of semantic knowledge could be encoded in low-dimensional neural representations, such as the abstract, disentangled representational geometries observed in many cognitive brain areas. Experimental work has shown that hippocampal (HPC) lesions impair the formation of flexible and generalizable semantic knowledge. However, HPC representations are often high-dimensional and sparse. How can these representations contribute to the formation of low-dimensional abstract representation across these cognitive cortical areas? Here, we reconcile these observations using multiregional neural network models. Modeling HPC as a sparse, decorrelating autoencoder, we show that its high-dimensional representations promote the robust and efficient formation of low-dimensional abstract representations in downstream regions such as prefrontal cortex. This effect is most pronounced when the geometry of sensory inputs is misaligned with the geometry of desired outputs, requiring a more complex input-output transformation. Mathematical analysis shows that hippocampal expansion recoding suppresses misleading low-dimensional input structure and reshapes the loss landscape of downstream networks, allowing abstract semantic geometry to emerge reliably. The resulting abstract representations provide an inductive bias for sample-efficient learning of novel semantic-to-action mappings. Finally, we show that consolidating hippocampus-dependent representations into a hippocampus-independent pathway improves out-of-distribution generalization based on input structure. Together, these results suggest a circuit mechanism by which high-dimensional hippocampal representations can support, rather than oppose, the emergence of low-dimensional abstract representations.