The coexistence of localized and distributed behavioral information in neural activity

Abstract

The degree to which control of an animals behavior is localized within particular neurons or distributed over large populations is central to understanding mechanisms of decision-making in brains. A first step in answering this question comes from understanding the scales at which neural activity is predictive of behavior. Here, we demonstrate how information measures at the individual, pairwise, and larger group levels characterize the localization of predictive information. We demonstrate these tools using high-dimensional neural data related to nematode and macaque behavioral decisions. Intriguingly, in both examples we find that similar behavioral information coexists across scales: the same information can be extracted from small groups of individually informative neurons or larger groups of randomly chosen neurons that individually have little predictive power. Our results suggest that methods for causal inference may miss potential causal pathways if they are biased toward finding localized control mechanisms.