Development of a connection matrix for productive grounded cognition

We investigate the development of a productive architecture for grounded cognition. Grounded representations cannot be copied to form compositional structures underlying productive cognition. Yet, they can be embedded in neural blackboard architectures, which provide a basis for productive cognition with grounded representations. In these architectures, compositional structures are formed by temporarily binding grounded representations using selective connection matrices. We present simulations that show that these selective connection matrices can develop on the basis of initially nonselective random connections. Two factors influence this development in particular: the variability of the initially nonselective random connections and the strength of the Hebbian and anti-Hebbian learning that influence the connections. To the best of our knowledge, this paper is the first to demonstrate the possibility of the development of a productive (compositional) architecture based on “in situ” (grounded) representations.