Rich single neuron computation implies a rich structure in noise correlation and population coding

Pairwise correlation in a population activity is a widelyobserved neural phenomenon. In particular, even withthe same mean stimulus, noisy fluctuations in the popu-lation firings are often correlated, and this so-called noisecorrelation has attracted a lot of attention in regard towhether it might transfer independent informationbeyond a mean population response [1]. However, in thecontext of the common input model where a commoninput noise drives the noise correlation, a recent influen-tial study suggested that the noise correlation must have asimple relationship with the average firing rate, or moreprecisely the average gain, and therefore claimed that thenoise correlation might not carry any independent infor-mation [2].In this work, we carried out a model study to probe thecorrelation-gain/rate relationship with biophysicallydefined single neuron models and found out that the rela-tionship with gain actually fails to capture large noise cor-relations in some models. We suggest that this is closelyrelated to the type 3 excitability of these neuron models.Type 3 excitability has been seen recently in model studies[3] and in some cortical neurons in the