Representational geometry reveals how neuronal diversity supports perceptual performance

Abstract

A complete understanding of population coding requires connecting multiple levels of neural processing: individual responses, population representations, and behavior. We link these by relating the distribution of neuronal tuning properties to a populations representational geometry and its efficiency for perceptual tasks. We use theory, analysis of recordings from macaque primary visual cortex (V1), and simulations to reveal how diversity of tuning amplitude and bandwidth enhances the population code for visual discrimination and identification. Both types of diversity drive different, but complementary changes to the representational geometry. Amplitude diversity increases the Euclidean distance between the responses to different stimuli, while bandwidth diversity creates a larger angular distance between them. The first utilizes the range of firing rates available to neurons, and the second exploits the high-dimensional nature of population responses. Population codes can be improved using these two different geometric changes, and amplitude and bandwidth diversity provide biological mechanisms for doing so. HighlightsO_LI- Perceptual performance is improved both by increased diversity of response amplitude and increased diversity of tuning bandwidth. C_LIO_LI- Both kinds of diversity improve visual discrimination and identification. C_LIO_LI- Amplitude diversity improves discrimination more, and bandwidth diversity improves identification more. C_LIO_LI- Representational geometry reveals the mechanisms of these effects. C_LI