Abstract
Understanding dynamic, directional interactions among large-scale brain networks which support sensory-based cognition remains a major challenge. Focusing on neural dynamics during perceptual encoding in a supraspan immediate free-recall paradigm, we develop a computational framework for estimating effective connectivity from source-localized electroencephalography (EEG). We integrate standardized Low-Resolution Brain Electromagnetic Tomography (sLORETA) with a biologically-informed recurrent neural network. The model incorporates key neurobiological constraints including Dale's Law, separate excitatory and inhibitory populations, and population-specific time constants to provide greater interpretability than conventional black-box machine learning approaches. We apply this framework to the study of the large-scale triple-network framework - Salience Network (SN), Default Mode Network (DMN), and Task-Positive Network (TPN) - comparing bidirectional interactions during rest and the encoding phase of a free-recall task of two distinct stimulus types (geometric shapes and words). The inferred connectivity patterns reveal state-dependent and bidirectional influences, in particular, SN-driven facilitation of both DMN and TPN during recall, stronger TPN-DMN inhibition during rest, and a reversal in DMN-to-SN influence across states. We supplement the computational findings by uncovering the mediating influence of individual differences in mental visual imagery as measured by the Vividness of Visual Imagery Questionnaire (VVIQ). Specifically, we uncover two selective latent associative links between VVIQ and the connectivity strength from SN directly and indirectly (after DMN modulation) to TPN, for geometric shapes and word encoding, respectively. The inferred connectivity patterns challenge switching-based accounts and highlight deep recurrent effective connectivity as a novel way to gain insights into the self-organization dynamics involved in episodic encoding.