Discovering Novel Circuit Mechanisms in Higher Cognition through Factor-Centric Recurrent Neural Network Modeling

Abstract

Recurrent neural networks (RNNs) have transformed how systems neuroscientists generate hypotheses about circuit mechanisms in higher cognition. Yet their promise has been constrained by a fundamental limitation: conventional RNNs are neuron-centric and therefore often difficult to interpret mechanistically. Here we introduce Restricted-RNN, a factor-centric RNN modeling framework developed to uncover interpretable circuit mechanisms. By formalizing factor communication through subpopulations, the dual of neuron communication through subspaces, Restricted-RNN departs fundamentally from standard neuron-centric RNN models and provides a distinct framework for describing circuit mechanisms. Using this approach, we identify novel circuit mechanisms underlying sequence working memory control and the counterintuitive firing-rate reversal observed in perceptual decision-making, with key predictions supported by neurophysiological recordings from monkey frontal and parietal cortex. More importantly, factor-centric RNN modeling reveals a unified low-dimensional neural control state space that links seemingly disparate phenomena across tasks, providing a geometric framework for understanding the pervasive role of control in higher cognition.