Finding Predictive Collective Variables in a Large Population of Retinal Neurons

The vertebrate retina has been shown to perform predictive computation on incoming visual signals, and it has been hypothesized that further prediction occurs at each successive layer of the visual stream. Using response data collected from 93 salamander retinal ganglion cells under naturalistic stimulus, we take the viewpoint of hypothetical downstream predictor neurons and search for features which optimally encode information about future responses. Across stimuli and prediction intervals, we find that all such information is compressible into a few (less than 10) linear collective variables of the present retinal output state. By leveraging variational inference and repeated stimulus trials in our dataset, we find that this predictive information is collectively encoded; it is mostly carried by correlations between neurons. At short timescales, individual effects matter more and noise autocorrelations contribute significantly, while at later timescales predictive features are highly collective and stimulus-induced correlations dominate. Our analysis demonstrates the feasibility of uncovering biologically relevant correlation structure in high-dimensional data using variational inference and basic machine learning tools.