Spatial computing for the control of working memory

Abstract

Working memory (WM) allows us to selectively remember and flexibly control a limited amount of information. Earlier work has suggested WM control is achieved by interactions between bursts of beta and gamma oscillations. The emerging question is how beta and gamma bursting, reflecting coherent activity of hundreds of thousands of neurons, can underlie selective control of individual items held in WM? Here, we propose a principle for how such selective control might be achieved on the neural network level. It relies on spatial computing, which suggests that beta and gamma interactions cause item-specific activity to flow spatially across the network over the course of a task. This way, control-related information about, for instance, item order can be retrieved from the spatial activity independent of the detailed recurrent connectivity that gives rise to the item-specific activity itself. The spatial flow should in turn be reflected in low-dimensional activity shared by many neurons. We test predictions of the proposed spatial computing paradigm by analysing control-related as well as item-specific activity in local field potentials and neuronal spiking from prefrontal cortex of rhesus macaques performing four WM tasks. As predicted, we find that the low-dimensional activity has a spatial component from which we can read out control-related information. These spatial components were stable over multiple sessions and did not depend on the specific WM items being used. We hypothesize that spatial computing can facilitate generalization and zero-shot learning by utilizing spatial component as an additional information encoding dimension. This offers a new perspective on the functional role of low-dimensional activity that tends to dominate cortical activity.