Correlated variability in population activity: noise or signature of internal computations?