Finding sensors for homeostasis of biological neuronal networks using artificial neural networks

To model a biological system despite a lack of complete information, statistical and machine learning can be used to replace a missing component with a classifier that is trained to give a near-optimal estimation of a target behavior. By filling the information gap in the system, this classifier can improve the analysis of better known components. We applied this approach to study the parameters of a proposed activity sensor of a biological neuronal network model by replacing the unknown sensor readout mechanism with an artificial neural network classifier. The classifier derives an error signal for homeostatic regulation of the pattern-generating neuronal network from the lobster stomatogastric ganglion. Using this approach, we predict optimal biological activity sensor parameters for homeostatic regulation and also provide insights into the biological architecture of the replaced sensor readout mechanism itself.