• machine learning surrogate models
• morphologically complex neurons
• optic nerve neuromodulation

We develop fast machine learning surrogate models to approximate the results of biophysically accurate neural simulations, allowing for real-time applications. Precisely, we explore two distinct problems: (1) the regression of firing rates in optic nerve fibers and (2) the classification of the activation states of morphologically complex neurons.

For the first problem, we present an exhaustive characterization of the neuromodulation of optic nerve fibers, and devise a heuristic approach to efficiently optimize stimulation protocols for arrays of electrodes implanted within the optic nerve of a patient, thereby opening up possibilities for vision restoration. We then develop highly accurate machine learning models to derive the induced firing rates and enable the reconstruction of perceived images.

In addressing the second problem, we innovatively employ Graph Neural Networks (GNNs) to classify morphologically complex neurons with a simplistic biophysics as either active or inactive, for both extracellular and intracellular stimulation scenarios, and compare them against a Multilayer Perceptron (MLP) as a baseline. GNNs outperforms MLPs in both cases, achieving promising accuracy rates (90% and 86% respectively).

The findings presented herein hold great promise for advancing our understanding of neural systems, enabling real-time applications in neural research, and potentially contributing to the restoration of vision in patients with optic nerve disorders.