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 compressing retinal inputs to perform optimal predictive computations allows putative downstream neurons to predict the future near-optimally across natural scenes. Other sensory systems also exhibit compression in their processing hierarchies 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.