A Minimal Computational Astrocyte as a Local Spatiotemporal Operator

Astrocytes integrate neural activity across space and time and modulate nearby synapses, yet it remains unclear which computational primitive is provided by this additional dynamical state. We introduce a minimal computational astrocyte defined by a restricted interface: local event sensing, leaky temporal integration, threshold-triggered astrocytic events, and persistent local modulation of synaptic transmission and plasticity. The resulting mechanism decomposes into spatial pooling, temporal filtering, nonlinear gating, and local feedback modulation. For the deterministic mean-field trajectory under stationary input, we derive an exact threshold-crossing law and a critical input rate; for Poisson spikes, the event-rate law is an approximation. At finite observation times, spatial pooling reduces estimation variance while introducing spatial bias, yielding a predictable operating regime. We embed the mechanism in a canonical unsupervised Diehl–Cook spiking network. Astrocytic integration parameters are calibrated once from input statistics and then frozen across corruption levels. At the nominal $$\begin{gathered} 350 \\ , \\ mathrmms \end{gathered}$$ presentation duration, the astrocytic network achieves $78.1$ mean accuracy over four seeds under $75$ dynamic salt-and-pepper corruption, while the unmodulated baseline reaches $9.8$. The advantage persists across a broad corruption range and then collapses near the analytically predicted loss of mean-field rate separation. Tests on native event streams probe the same local interface in frozen pretrained networks. These results identify a compact computational role for astrocyte-like state: local stateful gating of event-driven information flow.