Exploring Ensemble Coding in Retina

Computational models founded on the prevailing paradigm of retinal processing, while able to replicate the coarse structure of responses to white noise stimulus, fail to replicate responses to natural stimuli. The textbook view of retina, which posits independent filters that decorrelate stimulus features, reduce representational redundancy \cite{barlow1961} and encode local features in retinal ganglion cell (RGC) spike rates, leaves severe puzzles, unexplained about observed retinal anatomy and activity. We present, here, an addendum to the prevailing paradigm hypothesizing that \emph{perhaps}, the retina reduces uninformative correlations in stimulus with outer layers, as claimed, in order to reintroduce informative correlations, observed in RGC responses to ethologically relevant stimuli, with the circuitry in the inner retinal network (bipolar, amacrine, ganglions). Rather than strict independent coding and redundancy reduction, a notion which Barlow himself amended \cite{barlow2001}, we explore ensemble coding in retina and what information beyond the traditional view might exist in the retinal code. This work is in two related, yet independent parts. First, we develop a proof-of-concept abstract computational model of image segmentation using phase coding in the retina, hypothesizing that fine-time correlations in spike trains are induced by phase interactions influenced by the visual stimulus and that these fine-time correlations, informative about segments in an image, are multiplexed into spike-trains along with rate-coded local stimulus features. Following, we present a latent variable statistical model that aims to detect cell assemblies, or groups of cells that fire are often co-active, possessing fine-time correlations irrespective of their source and apply it to retinal spike-trains responding to both white noise and natural movie stimulus. The work presented here is novel and controversial and therefore, worth a read