Abstract
The dynamics of neural populations seem constrained to repeatedly visit a limited sub-set of all possible states. Within sensory populations, and especially in the retina, these repeated states take the form of millisecond-precise activity patterns. Enumerating a population’s activity patterns thus defines a neural dictionary: the set of patterns that could potentially represent different things. Unknown is if such a dictionary is a general principle of cortical populations, and if learning changes the dictionary. To address these questions, we analysed population activity from the medial prefrontal cortex (mPfC) of rats learning new rules in a Y-maze. We found that patterns of co-active neurons on millisecond time-scales occurred far in excess of those predicted by firing rates alone. The set of activity patterns was strongly conserved between waking and sleep. Yet pattern frequencies detectably changed between the sleep epochs before and after a maze session. These changes were greatest for patterns that, during trials, were expressed at the maze’s choice point and predicted the outcome of a trial. Successful learning of a rule system-atically changed the dictionary of patterns, such that the probabilities of patterns after learning where maintained in post-learning sleep. By contrast, during stable behaviour there was no systematic change to the dictionary. Our data show that population activity in the mPfC contains a consistent yet plastic dictionary of task-encoding patterns at millisecond time-scales. We propose that these finding are a signature of the probabilistic representation of behavioural strategies in mPfC.
Significance statement Cortex represents and computes information using the joint activity of many neurons. An open question is what experimentally observed features of this joint population activity are computationally relevant. We show here that the population activity from the prefrontal cortex of rats learning rules in a maze contains a specific dictionary of millisecond-precise activity patterns. This dictionary was altered during training, and encoded key parts of the task. But only during successful learning of a new rule was the dictionary seemingly permanently updated, because the changes could be detected during sleep after training. Our results thus further our understanding of the statistical structure of neural activity in cerebral cortex, and provides clues for the basis of cortical computation.
Conflict of interest The authors declare no conflicts of interest.
Acknowledgments We thank the Humphries lab (Javier Caballero, Mat Evans, Silvia Maggi) for discussions; Rasmus Petersen for comments on the manuscript; and P. Berkes and M. Okun for respectively making their KL divergence and raster model code publicly available. A.S. and M.D.H were supported by a Medical Research Council Senior non-Clinical Fellowship award MR/J008648/1 to M.D.H. A.P. was supported by Human Frontier Science Program Fellowship LT000160/2011-l and National Institute of Health Award K99 NS086915-01.