Adaptive computation in neuron–astrocyte networks

Context-dependent network dynamics

In our PLOS Computational Biology study, I developed and analyzed a dynamical model of neuron–synapse–astrocyte interactions. Astrocytic activity modulates synaptic adaptation and responds to neural activity and contextual inputs. This creates nested feedback loops whose components evolve at different rates. Mathematical analysis shows how slowly varying astrocytic signals can alter the attractor structure of faster neural and synaptic dynamics.

We then trained these networks on bandit tasks with changing contexts. The model links astrocytic modulation to learning in nonstationary environments, providing a mechanistic example of how slower biological processes can support flexible behavior. My contributions included model formulation, mathematical analysis, software, and computational experiments.

Connecting models with experimental observations

In a subsequent collaboration, I contributed computational modeling and analysis to a study of multi-timescale astrocytic computation during reward-guided behavior. Experimental collaborators characterized fast and slow calcium signals in cerebellar astrocytes and their distinct relationships to behavior. A neuron–astrocyte actor–critic network trained on a related sequence task developed heterogeneous temporal activity resembling these observed patterns.

The collaborative study remains a preprint. Its actor–critic interpretation provides a computational hypothesis for the observed division of function, linking state evaluation and the modulation of neuronal learning.

Together, these projects connect mechanistic modeling with reinforcement learning and experimental neuroscience. They use explicit interactions among neurons, synapses, and astrocytes to test how biological architecture can support contextual adaptation.

My complementary work on learning dynamical components from neural population data and inferring their timescales asks how recordings can reveal dynamical structure when the underlying mechanisms are not directly observed.