The correlation structure of spontaneous activity in populations of cortical neurons

Abstract:

Cortical neurons in sensory areas respond to external stimuli unreliably, a feature which could severely constrain information encoding. This trial-to-trial variability seems to be, in most cases, partly shared between nearby neurons. The standard mechanistic model explaining this variability and co-variability is built upon two principles: (1) cortical circuits operate in a balanced state, and (2) neighbour neurons in densely connected cortical circuits share a measurable fraction of their inputs. In the balanced state of large recurrent networks, neurons counterbalance the large supra-threshold excitation with large negative inhibition coming from I cells rendering a subthreshold highly-fluctuating net current. As a consequence of the fraction of anatomically shared inputs, it had been assumed that an irreducible fraction of their spiking variability would also be shared. We will challenge this view by showing that, in a standard randomly-connected network model in the balanced state, the strong and fast inhibitory feedback de-correlates neural firing in the presence of large fractions of shared inputs yielding marginal average correlations. Moreover, in this state individual pairs can show significant non-zero correlations both of positive and negative sign. We analyse spontaneous population activity from anaesthetised rats and obtain pair-wise correlation matrices from the spike count. During Activated periods cortical populations exhibit near-zero average correlations in the presence of two non-symmetric subcategories of neurons showing positive correlations within each category and negative correlations across categories. We investigated the time scale of correlations by computing pair Cross-Correlograms (CCGs) and found both peaks and troughs with a variety of time scales. We extended the randomly connected network model to capture these finer structure characteristics and consider a minimal model defined by two symmetric neural populations with strong excitatory feedback within each group and strong reciprocal inhibition. We will show that this simple architecture can generate correlation matrices showing some of the qualitative features observed in the data. We will finally discuss the implications of these results on interpreting the mechanisms which generate neuronal variability.