Computational evidence for an inverse relationship between retinal and brain complexity.

Abstract

Visual neuroscientists have long observed an inverse relationship between brain and retinal complexity: As brain complexity increases across species, retinas adapt to simpler visual processing. Lindsey et al. previously provided a computational explanation for this pattern, showing that shallow networks encode complex features in their first stage of processing, whereas deep networks encode simpler features. Here, these findings are extended to a suite of representational analyses and show that shallow networks generate high-dimensional representations with linear decision boundaries and specific visual features that can feed directly into behavioral responses. In contrast, deep networks generate low-dimensional representations with nonlinear decision boundaries and general visual features. These representations require further processing before they can produce the appropriate behavioral response. In summary, the findings extend a longstanding principle linking simpler retinal features to complex brains and offer a computational framework for understanding neural network behavior more generally.