How Minds Take Shape: Graph-ESN Reveals How Neural Ensembles Engineer Stable Representations

Abstract

Deciphering how neural populations encode, integrate, and dynamically transform information to generate predictive neural dynamics remains a fundamental pursuit in systems neuroscience. Although information is processed and represented across distributed neural ensembles, a critical question remains: How do these populations accumulate information and interact over time to form stable, coherent representations? The cortex comprises ensembles of interacting populations that cooperate to enable thought. However, an open challenge is to map how each population evolves its internal context in response to incoming information and crucially, how this evolving context shapes what each population communicates to other populations. To address this, we employ a customized variant of the Graph Echo State Network (Graph ESN) architecture that involves specialized populations. This formulation enables the model to disentangle and represent multi-scale oscillatory patterns in neural data, offering a more biologically plausible and task-relevant alternative to traditional ESNs. By leveraging the rich hidden-state dynamics of this architecture, we illustrate how neural ensembles iteratively interact and converge toward stable, temporally evolving representations of information. We further investigate the distinct contributions of individual populations in this collaborative process of representational stabilization. Mapping this temporally unfolding, cooperative structure sheds light on the neural mechanisms underlying distributed representation and inter-population coordination--processes that may ultimately support organized cognition.