The role of recurrent networks in neural architectures of grounded cognition: learning of control

Recurrent networks have been used as neural models of language processing, with mixed results. Here, we discuss the role of recurrent networks in a neural architecture of grounded cognition. In particular, we discuss how the control of binding in this architecture can be learned. We trained a simple recurrent network (SRN) and a feedforward network (FFN) for this task. The results show that information from the architecture is needed as input for these networks to learn control of binding. Thus, both control systems are recurrent. We found that the recurrent system consisting of the architecture and an SRN or an FFN as a "core" can learn basic (but recursive) sentence structures. Problems with control of binding arise when the system with the SRN is tested on number of new sentence structures. In contrast, control of binding for these structures succeeds with the FFN. Yet, for some structures with (unlimited) embeddings, difficulties arise due to dynamical binding conflicts in the architecture itself. In closing, we discuss potential future developments of the architecture presented here.