Abstract
To study sensory representations, neuroscientists record neural activities while presenting different stimuli to the animal. From these data, we identify neurons whose activities depend systematically on each aspect of the stimulus. These neurons are said to be "tuned" to that stimulus feature. It is typically assumed that these tuned neurons represent the stimulus feature in their firing, whereas any "untuned" neurons do not contribute to its representation. Recent experimental work questioned this assumption, showing that in some circumstances, neurons that are untuned to a particular stimulus feature can contribute to its representation. These findings suggest that, by ignoring untuned neurons, our understanding of population coding might be incomplete. At the same time, several key questions remain unanswered: Are the impacts of untuned neurons on population coding due to weak tuning that is nevertheless below the threshold the experimenters set for calling neurons tuned (vs untuned)? Do these effects hold for different population sizes and/or correlation structures? And could neural circuit function ever benefit from having some untuned neurons vs having all neurons be tuned to the stimulus? Using theoretical calculations and analyses of in vivo neural data, I answer those questions by: a) showing how, in the presence of correlated variability, untuned neurons can enhance sensory information coding, for a variety of population sizes and correlation structures; b) demonstrating that this effect does not rely on weak tuning; and c) identifying conditions under which the neural code can be made more informative by replacing some of the tuned neurons with untuned ones. These conditions specify when there is a functional benefit to having untuned neurons.
