How do natural neuronal networks deal with noise ? – NatNetNoise

The retina processes visual stimuli through a sequence of stacked neuronal layers equipped with horizontal connections, and eventually transmits visual information through the noisy spiking activity of its output layer. The retina can therefore be seen as a concrete biological realisation of a deep neural network with recurrent layers. However, and contrary to artificial networks, neural noise strongly affects information transmission. This proposal aims to show that thanks to the correlation induced by horizontal connections, noise is shared across neurons, and this can strongly improve the system's capacity of transmitting information, even if individual cells' variability is high. With this objective, we will combine deep machine learning models of the retinal stimulus processing inferred from ex-vivo multi-electrode array recordings, with state-of-the-art information theory's tools, to estimate the system's capacity of transmitting stimulus information.