Predicting Spike Activity in Neuronal Cultures

Neuronal cultures are small living networks in a closed system. This paper investigates the question whether it is possible to discover the functional connectivity and to model the dynamics of such neuronal cultures. Doing so may contribute to a better understanding of neural information processing. We employ a machine learning approach, which constructs the functional connectivity map of a neuronal culture based on multiple spike trains of its spontaneous activity recorded with Multi-Electrode-Array (MEA) technology. The spike train of an electrode is modeled as a point process, where the firing probability depends on the finite spike history of all electrodes. To capture potential plasticity of the network, we employ a gradient descent method, which naturally allows for online learning. Several experiments with different cultures show that learned models can predict upcoming spike activity quite well.