Presentation Summary

Title : Chaos control: Reliable signal propagation and pattern completion in chaotic neuronal networks
Authors : Pierre Yger, Olivier Marre, Sami El Boustani, Pierre Baudot, Cyril Monier, Manuel Levy, Andrew Davison and Yves Frégnac
Year : 2007


Neural networks spontaneously generate complex activity patterns whose significance is still a matter of debate. By trying to mimic such dynamical behaviour, recent models of spiking neurons can generate asynchronous, irregular, self-sustained activity without any stochastic input. However, the propagation of an external drive within those networks is still a challenge since stimulus-evoked activity is corrupted by the chaotic background, triggering a response neither reliable nor predictable (Vogels and Abbott, 2005). Here we show that reproducible responses can be obtained if the input stimuli are constrained to reproduce the dynamics of the self-sustained activity. We forced a variable proportion of neurons within the network to replay a pattern generated during spontaneous activity, and we measured whether this constrained sub-population reliably drives the remaining neurons, forcing them to converge to the completion of the full pattern. This was implemented using a chaotic neural field model (Sompolinsky et al, 1988), and recurrent neural networks of sparsely connected current- or conductance-based integrate-and-fire neurons. The reproducibility of the responses was compared with that evoked by a Poisson stimulus. Our results show that the more neurons, and particularly inhibitory ones, are clamped in the network, the more the stimulus evoked activity propagates reliably. We thus demonstrate the feasibility of a highly temporally precise encoding of stimuli within a chaotic background activity. Accordingly, the synaptic weight distribution in a neocortical-like neural network can be viewed as the steady-state result of an unseen learning process, suggesting that spontaneous ongoing activity replays learned “songs”. This dynamic convergence towards predictable neural assembly patterns may support pattern completion within associative memory networks, and, more generally, efficient message transmission within an irregular activity regime.