Stability control of metastable states as a unified mechanism for flexible temporal modulation in cognitive processing.

Flexible modulation of temporal dynamics in neural sequences underlies many cognitive processes. For instance, we can adaptively change the speed of motor sequences and speech. While such flexibility is influenced by various factors such as attention and context, the common neural mechanisms responsible for this modulation remain poorly understood. We developed a biologically plausible neural network model that incorporates neurons with multiple timescales and Hebbian learning rules. This model is capable of generating simple sequential patterns as well as performing delayed match-to-sample (DMS) tasks that require the retention of stimulus identity. Fast neural dynamics establish metastable states, while slow neural dynamics maintain task-relevant information and modulate the stability of these states to enable temporal processing. We systematically analyzed how factors such as neuronal gain, external input strength (contextual cues), and task difficulty influence the temporal properties of neural activity sequences-specifically, dwell time within patterns and transition times between successive patterns. We found that these factors flexibly modulate the stability of metastable states. Our findings provide a unified mechanism for understanding various forms of temporal modulation and suggest a novel computational role for neural timescale diversity in dynamically adapting cognitive performance to changing environmental demands. Author Summary: The brain often uses sequences of neural activity to perform complex cognitive tasks such as recognizing speech, making decisions, or holding information in working memory. These sequences can speed up or slow down depending on factors like attention, task difficulty, or expectations-but how the brain controls this timing remains unclear. In this study, we built a biologically plausible model of a neural network that includes both fast and slow neurons and learns tasks through simple, realistic rules. We show that the slow neurons can hold onto past information and control how long the network activity stays in each state of a neural sequence. This control depends on the stability of each state, which is influenced by factors such as the external input strength, task difficulty, and top-down modulation. Our model coherently explains a variety of experimental findings and provides a unified theory for how the brain might flexibly adjust the speed of thought by taking advantage of diverse timescales of neural activity.