Proceedings of the Annual Meeting of the Cognitive Science Society

We present several experiments demonstrating the efficiencyand scalability of a biologically inspired spatial representationon navigation tasks using artificial neural networks. Specifi-cally, we demonstrate that encoding coordinates with SpatialSemantic Pointers (SSPs) outperforms six other proposed en-coding methods when training a neural network to navigate toarbitrary goals in a 2D environment. The SSP representationnaturally generalizes to larger spaces, as there is no definitionof a boundary required (unlike most other methods). Addition-ally, we show how this navigational policy can be integratedinto a larger system that combines memory retrieval and self-localization to produce a behavioural agent capable of findingcued goal objects. We further demonstrate that explicitly incor-porating a hexagonal grid cell-like structure in the generationof SSPs can improve performance. This biologically inspiredspatial representation has been shown to be able to producespiking neural models of spatial cognition. The link betweenSSPs and higher level cognition allows models using this rep-resentation to be seamlessly integrated into larger neural mod-els to elicit complex behaviour.