Spatial computing enables flexible cognitive control in neural networks

Flexible and generalizable control over sensory representations is needed to achieve human-level cognition. Here we implement a novel principle, Spatial computing, in a cortical neural network model. By explicitly utilizing the cortical topography as an additional coding dimension, it is possible to flexibly update the cognitive status of sensory representations. We first demonstrate that Spatial computing requires distance-dependent like-to-like connections and local winner-takes-all-dynamics resulting from non-specific feedback inhibition to be stable. Both motifs are consistent with cortical connectivity. By making use of topography, it is trivial to perform several cognitive tasks such as prioritizing, deleting or re-ordering the rank of items held in working memory, as needed during mental arithmetic. Importantly, spatial computing dissociates the substrate of sensory representations (cortical connectivity) from that of cognitive control (topographical dynamics), thereby allowing learned operations to automatically generalize across representations. Our results suggest that biology is likely to utilize the physical dimensions of the cortex to perform computations and that taking space into account in artificial networks may significantly improve their computational capabilities.