アストロサイトによるマルチタイムスケール計算

The basic computational unit of the brain has long been defined as the neuron. However, mounting evidence suggests that other cells, especially astrocytes, also perform computation. Here we demonstrate that cerebellar astrocytes decompose norepinephrine input into slow and fast calcium activities through differential adrenergic receptor engagement. During reward learning in mice, slow and fast activities selectively target distinct synaptic pathways. Causal manipulations reveal that slow α1-adrenergic signals maintain behavioral states and coordinate transitions, while fast α2-adrenergic signals govern event-triggered responses and reinforcement learning. Remarkably, an actor-critic neural network trained on a similar sequence task spontaneously recapitulates these multitimescale dynamics, suggesting astrocytes implement critic-like computations that evaluate states and modulate neuronal learning. These findings establish astrocytes as multilevel processors that transform univariate neuromodulatory inputs into multivariate, pathway-specific circuit control operating in parallel with neuronal processing.