Proceedings of the Annual Meeting of the Cognitive Science Society

We present a neural dynamic architecture that grounds sen-tences in perception which combine multiple concepts throughnested spatial relations. Grounding entails that the model getsfeatures and relations as categorical inputs and matches themto objects in space-continuous neural maps which represent vi-sual input. The architecture is based on the neural principlesof dynamic field theory. It autonomously generates sequencesof processing steps in continuous time, based solely on highlyrecurrent connectivity. Simulations of the architecture showthat it can ground sentences of varying complexity. We thusaddress two major challenges in dealing with nested relations:how concepts may appear in multiple different relational roleswithin the same sentence, and how in such a scenario variousgrounding outcomes may be “tried out” in a form of hypothesistesting. We close by discussing empirical evidence for crucialassumptions and choices made when developing the architec-ture.