Effects of Input Structure and Topology on Input-Driven Functional Connectivity Stability

The computational properties of neural networks derive largely from their network connectivity. Recurrent connectivity allows for activity to evolve dynamically over time, creating high dimensional representations. This has been exploited in recurrent networks with fixed connections, which avoid the complexity of temporal credit assignment in modifying fixed connections. The processing of such reservoir networks can be further refined and constrained by the network topology. The current research addresses an additional dimension by which network connectivity can be manipulated. Inherent structure in the input can drive the reservoir into different functional connectivity configurations, that are stable over time under certain conditions. We consider that this is of interest, because of the relation between connectivity and computation. Ideally, the network can be reconfigured by structure in the input to perform specific computations. To address this issue, we explore how structure in input provided to the network can impact the functional connectivity. We demonstrate three fundamental properties of input-driven functional connectivity: 1) structured input can produce coherent functional connectivity, 2) the functional connectivity differs depending on the input structure, and 3) coherence of functional connectivity is destroyed in the absence of structure in the input. In the context of input-driven reservoir dynamics, language processing has a particular status because of its structural coherence. We discuss the implications for reservoir computing and for the better understanding of human neurobiology of language.