Using Transfer Entropy to Understand Information Flow and Connectivity in Neural Networks

One central goal of psychology is to link human behaviors and cognitive processes with concrete neural network. In social science, there is growing interest in understanding social structures through the lens of social network. Though definition of networks and their dynamics are context dependent, the link between network connectivity and dynamics plays a key role in all fields. In this thesis, we study how transfer entropy can elucidate neural network structures. We use transfer entropy to examine two types of model, a network with firing-rate-based nodes with threshold linear transfer function, and a balanced network consisting of leaky integrate-and-fire type neurons. With a firing rate neural model, the effective connectivity inferred by transfer entropy matches the underlying directional network. For balanced networks, presence of connection between two neurons is not predictive of existence of information flow. However, for a balanced network with cluster or layer structures, the multivariate transfer entropy calculation using samples of each neural cluster accurately captures the direction of the cluster-wise projection. We focused on networks in neuroscience, but the framework can be generalized to understand how social network propagates information.