Slow and Fast Neurons Cooperate in Contextual Working Memory through Timescale Diversity

Abstract: Neural systems process information across a broad range of intrinsic timescales, both within and across cortical areas. While such diversity is a hallmark of biological networks, its computational role in nonlinear information processing remains elusive. In this study, we examine how heterogeneity in intrinsic neural timescales within the frontal cortex - a region central to cognitive control - enhances performance in a context-dependent working memory task. We develop a recurrent neural network (RNN) composed of units with distinct time constants to model a delayed match-to-sample task with contextual cues. This task demands nonlinear integration of temporally dispersed inputs and flexible behavioral adaptation. Our analysis shows that task performance is optimized when fast and slow timescales are appropriately balanced. Intriguingly, slow neurons, despite weaker encoding of task-relevant inputs, play a causal role in sustaining memory and improving performance. In contrast, fast neurons exhibit strong but transient encoding of input signals. These results highlight a division of computational roles among neurons with different timescales: slow dynamics support stable internal states, while fast dynamics enable rapid signal encoding. Our findings provide a mechanistic account of how temporal heterogeneity contributes to nonlinear information processing in neural circuits, shedding light on the dynamic architecture underlying cognitive flexibility.