Elucidating the selection mechanisms in context-dependent computation through low-rank neural network modeling.

Humans and animals exhibit a remarkable ability to selectively filter out irrelevant information based on context. However, the neural mechanisms underlying this context-dependent selection process remain elusive. Recently, the issue of discriminating between two prevalent selection mechanisms-input modulation versus selection vector modulation-with neural activity data has been highlighted as one of the major challenges in the study of individual variability underlying context-dependent decision-making (CDM). Here, we investigated these selection mechanisms through low-rank neural network modeling of the CDM task. We first showed that only input modulation was allowed in rank-one neural networks and additional dimensions of network connectivity were required to endow neural networks with selection vector modulation. Through rigorous information flow analysis, we gained a mechanistic understanding of why additional dimensions are required for selection vector modulation and how additional dimensions specifically contribute to selection vector modulation. This new understanding then led to the identification of novel neural dynamical signatures for selection vector modulation at both single neuron and population levels. Together, our results provide a rigorous theoretical framework linking network connectivity, neural dynamics, and selection mechanisms, paving the way towards elucidating the circuit mechanisms when studying individual variability in context-dependent computation.