A vector navigation and inference architecture can construct universal cognitive maps for abstract reasoning

The idea of cognitive maps in the hippocampal formation has its origins in decades of spatial cognition research. However, increasing evidence points to cognitive maps also supporting non-spatial tasks. It appears that the hippocampal-entorhinal system can map any combination of systematically varying stimulus dimensions onto the same neural manifolds that support spatial representations-grid cells. Here, I propose a model that shows how extant spatial navigation architectures can iteratively construct universal cognitive maps via non-local inference. The model exhibits noise tolerance congruent with behavior, and the mapping process accommodates discontinuous stimulus spaces. It is also shown that the model can support several types of abstract reasoning, including analogy-making, subspace construction, and perspective taking. Importantly, these capabilities are supported by the same computational modules that are used to build a map. Neural-level implementations for several model components are proposed, and it is shown how the (abstract) 1D attributes that are to be combined on a map can be isolated upstream of entorhinal cortex. The combination of mechanisms at play thus suggests how spatial navigation architectures can function as domain-general substrates of cognition.