Comparison of experimental resting state data with large scale neural network simulations

The analysis of massively parallel spiking activity during behavior from monkey motor cortex reveals interesting patterns (e.g. Torre et al. 2016). We aim for a better understanding of the network mechanisms leading to such features based on simulated neural networks. Existing large scale models, however, do not account for behavior, but represent the ‘resting state’. Therefore we recorded spiking activity with a Utah array (100 electrodes, 4x4 mm2, Blackrock Microsystems) when the monkey was not subject to any task or stimulus. It forms the basis for the comparison between experiment and model aiming to improve the simulation until the experimental statistics can be quantitatively reproduced. We started out with a layered model of a cortical microcircuit (Potjans & Diesmann 2014), subsequently extended to 4x4 mm2 corresponding to the cortical patch we record from. For the comparison, model data are subsampled to 140 neurons (layer 5), and the experimental data are separated into inhibitory (INH) and excitatory (EXC) neurons based on their spike width (Dehghani et al. 2016). Applying the same analysis workflow and software (Elephant, http://neuralensemble.org/elephant/), we derive distributions of firing rates (FR), regularity, pairwise fine temporal correlations (CC), and rate covariances. In both data sets the average covariance and its standard deviation are significantly larger for INH than EXC population. The same holds for the experimental CC. In the model, however, the average CC is higher for EXC neurons. Also the mean FR differs between populations in experiment, but not in the simulation. Since our model is currently based on connectivity parameters from various cortices and species, the deviations between the data sets were to be expected. Next, we aim to adjust the model towards motor cortex