Efficient Working Memory Maintenance via High-Dimensional Rotational Dynamics
Working memory (WM) is fundamental to higher-order cognition, yet the circuit mechanisms through which memoranda are maintained in neural activity after removal of sensory input remain subject to vigorous debate. Prominent theories propose that stimuli are encoded in either stable and persistent activity patterns configured through attractor mechanisms or dynamic and time-varying activity patterns brought about through functionally-feedforward network architectures. However, cortical circuits exhibit heterogeneous responses during WM tasks that are challenging to reconcile with either hypothesis. We hypothesised that these complex response dynamics could emerge from an optimally noise-robust and energetically efficient solution to WM tasks. We show that, in contrast to previous theories, networks optimised for efficient WM encoding exhibit high-dimensional rotational dynamics. We find direct evidence for these rotational dynamics in large-scale recordings from monkey prefrontal cortex. Our findings suggest that the complex and dynamic response properties of WM circuits emerge from efficient coding principles.