A unified neurocomputational model for spatial and linguistic representations

Abstract

Hippocampus and entorhinal cortex encode spaces by spatially local and hexagonal grid activity patterns (place cells and grid cells), respectively. In addition, the same brain regions also implicate neural representations for non-spatial, semantic concepts (concept cells). However, the relationship between those representations remains to be understood. We propose a unified neurocomputational model for spaces and linguistic concepts, called “disentangled successor information (DSI)”, based on reinforcement learning and word embedding models in natural language processing (NLP). DSI generates spatial representations in a 2-dimensional space that correspond to place cells and grid cells. Furthermore, the same model creates concept-specific linguistic representations which resemble concept cells found in hippocampus and entorhinal cortex. Notably, with DSI representations, we could perform analogical inference of spatial contexts and words by the same computational framework, which can be biologically interpreted as partial modulation of assemblies of concept cells and non-grid cells in EC. Our model suggests the existence of a shared computational mechanism behind spatial and conceptual representations in hippocampus and entorhinal cortex.