Preserving predictive information under biologically plausible compression

Abstract

Retinal ganglion cells (RGCs) show high convergence onto their downstream projections, which poses a problem for information transfer: how can information be preserved through a synaptic layer that has significantly more inputs than outputs? Lossy compression suggests many efficient, yet computation-agnostic, methods for reading out input stimuli or activity patterns. Focusing on prediction as a ubiquitous computation in the brain, we compare compressions that explicitly retain predictive information to common neural compression frameworks that do not. We find evidence that downstream areas may compress their retinal inputs in a way that allows them to perform optimal predictive computations across many natural scenes. Other sensory systems also exhibit compression in their processing hierarchies, such as at the glomeruli stage in the olfactory system, and we hope that our framework will be useful in cases where it is not yet known how information about a specific computation is maintained under compression. SIGNIFICANCE STATEMENTProducing successful behavior, such as escaping predators, requires the visual system to overcome significant sensory processing delays by predicting the future state of the world. Neurons in the eye capture some of the most predictive features of visual information, but this information must be accessible to downstream areas that receive synaptically compressed inputs from the retina. We tested how carefully retinal activity must be compressed to preserve predictive information in natural scenes. Biologically plausible compressions that are agnostic to prediction can preserve substantial future information, but only compression optimized for this task extracts generalizable motifs that allow it to predict in any natural scene. This suggests that downstream circuits may optimally compress their inputs specifically for the task of prediction.