Separable neural population representations are constructed from mixed single neuron selectivity in the mouse early visual system

Abstract

Both sensory and non-sensory brain regions receive mixed inputs from single neurons which require decomposition and integration before proceeding through a processing hierarchy. Whether mixed input signals are used in biological neural networks to derive pure single neuron representations, or distributed as new population representations from mixed single neurons, is not clear. In this study, we measured the distribution of single neuron hue and luminance tuning in the dorsolateral geniculate nucleus (dLGN) and primary visual cortex (V1) of mice, as well as the information about and structure of hue and luminance representations in populations of hundred of simultaneously sampled neurons. We compare single neuron and population encoding to null models expected for random integration and extraction of pure categorical single neuron representation. Using both univariate and multivariate regression techniques, we consistently noted that tuning for hue and luminance, rather than clustering into categorical response structures, formed uniform distributions. While the distribution of single neuron selectivity varied across the thalamocortical circuit, we found no evidence of categorical tuning organization emerging in the hierarchy. Nevertheless, populations contained complete information, in either high-dimensional linear representations or low-dimensional non-linear representations. In summary, we find that as early as primary sensory cortex and thalamus single neurons that have mixed selectivity for hue and luminance form a high dimensional representation of those variables, which can be non-linearly embedded in multiple separable representations.