Neural-astrocyte interaction enables contextually guided circuit dynamics

Recurrent neural networks, a ubiquitous construct in machine intelligence, have become a powerful hypothesis-generating tool in theoretical neuroscience. Such networks can be used to examine potential circuit mechanisms associated with a variety of cognitive functions. Here, we use recurrent networks to engage a new theoretical question: the potential role of astrocytes in neural computation. Astrocytes are highly abundant cells in the cortex that are capable of modulating many facets of neural dynamics, including excitability, synaptic efficacy and plasticity. However, despite this modulatory capability, astrocytes are generally disregarded in models of neural computation, perhaps due to their much slower time-scale and seeming lack of specificity in their neural targets, and because of a lack of general theories of astrocytes contribution to neural circuit activity. Here, we explore a potential computational function of astrocytes: the contextual guidance of synaptic efficacy for rapid switching between previously learned tasks, inspired by the latest conceptual leap in the field of astrocyte biology. We build a recurrent neuro-astrocyte network in which each astrocyte modulates the efficacy of a subset of synapses, motivated by astrocytic tiling of neural space. Astrocytes in the model are sensitive to the association rule or context and propagate, in essence, a low-rank perturbation to the synaptic weights of the neural network. We show that fast learning can converge in the presence of these perturbations, leading to a single context-dependent network able to quickly switch contexts without relying on re-learning. Critically, astrocytic perturbation need only exist and be context-dependent for this mechanism to succeed; there is no transport of synaptic weights to `design' astrocytic modulation. This work suggests a potential computational role for astrocytic modulation in neural circuits, and new frameworks in multiple time-scale recurrent neural networks.