A NEURAL MODEL TO TRANSFER GRID-LIKE SPATIAL REPRESENTATIONS TO CONCEPTUAL DOMAINS

The hippocampal formation has long been viewed as a neural substrate for spatial navigation, yet accumulating evidence shows that it also encodes abstract, conceptual, and relational information. In particular, human fMRI studies have revealed grid-like representations in conceptual spaces in the entorhinal cortex, paralleling classical spatial grid cells. How such “conceptual grid cells” emerge remains a fundamental question. A parsimonious hypothesis is that preexisting spatial representations are repurposed for conceptual domains. However, it is computationally unclear how neural systems achieve such cross-domain transfer without explicit structural cues. Here, we propose a neuro-computational model that reuses grid cells to transfer spatial representations from physical to conceptual spaces. Our key assumption is that grid cells encode the intrinsic geometry of 2D space through their state-transition dynamics. Our model exploits this geometric knowledge to search for a mapping of conceptual entities on to the low-dimensional manifold formed by valid grid-cell population activities. This “manifold-constrained optimization” allows the same population of grid cells to acquire grid-like response properties in conceptual domains when geometric structures match. Simulation shows that the model rapidly learns novel conceptual domains without specialized pretrained representation, accommodates both discrete and continuous conceptual tasks, and reproduces neural signature of conceptual grid cells similar to the human entorhinal cortex. Moreover, the model reveals a functional link between the modular structure of grid cells and the formation of conceptual grid-like representations. These findings provide a mechanistic explanation of conceptual grid cells as a byproduct of geometric reuse rather than de novo organization.