Scaling of temporal correlations in densely connected networks of LIF neurons

Abstract

It has been experimentally reported that neurons in cer-ebral cortex can fire in a decorrelated fashion despite presumably sharing a non-negligible fraction of their inputs[1,2]. Performing a scaling analysis where the synaptic strengths decrease as 1/sqrt(N), being N the number of neurons in the network, it has been analyti-cally shown [1]that randomly connected networks of binary excitatory (E) ad inhibitory (I) neurons show average correlation coefficient of the activity which decreases as 1/N while correlations of the synaptic cur-rent components remain O(1). The above applies to densely connected networks in which the connection-probability does not depend on N[1], in contrast to sparse networks where it is the number of connections of a neuron which is kept constant[3]. Here we present a numerical study of the dynamics of random and den-sely connected EI networks of current-based leaky-inte-grate-and-fire neurons. Preliminary results show the existence of two regimes depending on the magnitude of the external drive onto the network: one in which average correlations do not systematically decrease with N (regime obtained with large external currents) and one in which average correlation of voltages (Fig. 1) and spike counts decrease with N (regime obtained with small external current). Preliminary data indicates that the latter regime occurs in the presence of strong corre-lations of the current components. The lag between excitatory and inhibitory spikes also decays with the net-work size. These results suggest that a mechanism of active decorrelation of synaptic inputs similar to the one previously found for binary networks, also exists in net-works of integrate-and-fire neurons.