Natural Motion and Optimal Prediction in a Complete Retinal Population

Abstract: Signals in the natural world are often characterized by a mixture of amplitude scales, like the quiet and loud segments of a musical recording. This property is manifested in the form of non-Gaussian, heavy-tailed distributions and nonlinear dependencies, both over time and across signal components. This is in strong contrast to the Gaussian and linear features typically assumed when modeling input signals. In the context of neuroscience, natural signals pose a serious challenge for sensory systems, which must adapt on the fly in order to efficiently encode them. Our recent work has demonstrated that the motion of objects in natural scenes also contains a mixture of scales, with a locally averaged velocity amplitude that fluctuates significantly on sub-second timescales. We have shown that this behavior can be modeled using an autoregressive Gaussian scale-mixture (ARGSM) model, which captures the temporal correlation structure of both the velocity and the fluctuating scale. Retinal responses to object motion have been characterized previously using carefully controlled artificial stimuli with Gaussian and linear statistics, revealing an efficient predictive code through the information bottleneck method. Here, we will extend this analysis to more naturalistic stimuli by incorporating a fluctuating scale variable matched to the statistics of natural scenes. We will bring together new experimental access to full RGC populations and our new theory about predictive coding and natural motion statistics. We will quantify predictive information about 1D and 2D motion trajectories in complete populations and sub-populations of mouse RGCs. Theoretically, this will require new calculations of information bottleneck-optimal representations under the ARGSM model. These will allow us to assess the performance of the retinal code using state-of-the-art recordings of mouse retinal ganglion cells (RGCs). Of particular interest are the contributions of the great diversity of RGC subtypes to the neural coding of these dynamically rich, naturalistic stimuli.