On the presence of high-order interactions among somatosensory neurons and their effect on information transmission

Abstract

In order to understand how populations of neurons encode information about external correlates, it is important to develop minimal models of the probability of neural population responses which capture all the salient changes of neural responses with stimuli. In this context, it is particularly useful to determine whether interactions among neurons responding to stimuli can be described by a pairwise interaction model, or whether a higher order interaction model is needed. To address this question, we compared real neural population activity obtained from the rat somatosensory cortex to maximum-entropy models which take into account only interaction of up any given order. By performing these comparisons, we found that interactions of order two were sufficient to explain a large amount of observed stimulus-response distributions, but not all of them. Triple-wise interactions were necessary to fully explain the data. We then used Shannon information to compute the impact of high order correlations on the amount of somatosensory information transmitted by the neural population. We found that correlations of order two gave a good approximation of information carried by the neural population, within 4% of the true value. Third order correlations gave an even better approximation, within 2% of the true value. Taken together, these results suggest that higher order interactions exist and shape the dynamics of cortical networks, but play a quantitatively minor role in determining the information capacity of neural populations.