Abstract
Finding elements of a complex network which contribute most to the network's overall behavior is an open problem in various fields. This challenge is particularly difficult in neuroscience as it requires identifying which of a mammalian brain's many millions of neurons inform specific behavioral choices. Using methods inspired by compressed sensing, we identified subsets of CA1 neuronal ensembles recorded while only male rats performed spatial memory and cue-approach tasks in a plus maze. These subsets consisted of the units with firing rates which co-varied most closely with overall ensemble activity. Unit activity from these predictive subsets asymmetrically predicted the activity of other units in the ensemble. Excluding the predictive subset had no effect on ensemble decoding of the rat's current location but reduced decoding of past and future locations, suggesting that the predictive subset encodes nonlocal information. Predictive subsets likely represent a hierarchical and sparse coding scheme used by CA1, and further investigation of the properties of these sub-populations may lead to additional insights into the basic computational processes of the brain.