Modelling Retinal Ganglion Cells Stimulated with Static Natural Images

A standard approach to model retinal ganglion cells uses reverse correlation to construct a linear-nonlinear model using a cascade of a linear filter and a static nonlinearity. A major constraint with this technique is the need to use a radially symmetric stimulus, such as Gaussian white noise. Natural visual stimuli are required to generate a more realistic ganglion-cell model. However, natural visual stimuli significantly differ from white noise stimuli and are not radially symmetric. Therefore a more sophisticated modelling approach than the linear-nonlinear method is required for modelling ganglion cells stimulated with natural images. Machine learning algorithms have proved very capable in modelling complex non-linear systems in other scientific domains. In this paper, we report on the development of computational models, using different machine learning regression algorithms, that model retinal ganglion cells stimulated with natural images in order to predict the number of spikes elicited. Neuronal recordings obtained from electro-physiological experiments in which isolated salamander retinas are stimulated with static natural images are used to develop these models. In order to compare the performance of the machine learning models, a linear-nonlinear model was also developed from separate experiments using Gaussian white noise stimuli. A comparison of the spike prediction using the models developed shows that the machine learning models perform better than the linear-nonlinear approach.