Retina optimised for prediction across animal species

Authors: Luke Taylor, Andrew J King, Nicol Spencer Harper

Presentation type: Flash talk at SNUFA 2023 online workshop (7-8 Nov 2023)

Abstract

The retina has often been described as operating under efficient coding, acting like a camera, and transforming incoming visual signals into a compressed neural representation. Various experimental observations over the last decades have challenged this view, suggesting the retina to rather extract and signal features within visual stimulus that are predictive of the future. To address this discrepancy, we developed a biologically detailed spiking retinal model, trained on natural movies under metabolic constraints, to either encode the present or predict the future. We found the predictive model to capture a vast number of retinal phenomena, including the major cell types and their mosaic-like organization; retina-like spike codes, such as latency encoding and omission stimulus responses; and motion tuning similar to Y-cells and object-motion-sensitive neurons. We also found the predictive retinal model to better predict neural responses to natural stimuli in marmoset, macaque, mouse, and salamander retina. Our work showcases how a biologically detailed and unsupervised model can capture a large and diverse number of retinal phenomena, suggestive of the retina operating under a predictive rather than compression-like regime.