Back to the abstracts page
Back to the NCCD 2015 home page
Quiescent state activity as a signature of efficient coding
Veronika Koren1 and 2 Sophie Denève3 and David Barrett 4
1Neural Information Processing Group, Technische Universität Berlin, Berlin, Germany 2Bernstein Center for Computational Neuroscience Berlin, Berlin, Germany 3Group for Neural Theory, Departement des Etudes Cognitives, Ecole Normale Supérieure, Paris, France 4Computational and Biological Learning, Information Engineering Division, University of Cambridge, Cambridge, UK

A growing body of experimental evidence suggests that there is a complex interaction between stimulus-driven and spontaneous activity in the cortex [1]. In primary sensory cortices of awake behaving animals, multi-electrode recordings show neural responses during periods when the neural population is driven by a sensory stimulus, as well as during periods of quite-wakefulness, when no sensory stimulus is provided to the animal. During quite-wakefulness, the population activity can take the form of characteristic synchronized bursts of spiking activity, or up-states, interspersed by periods of silence, or down-states. Why does such activity occur and what could be its functional role? Previous mechanistic explanations have proposed that spontaneous activity is due to a dynamic interplay of local excitation, adaptation and inhibition [2]. Another mechanistic explanation is that "up" states could be triggered by modulations from distal areas, either top-down or subcortical, like a "gate" opening and closing [1]. However, the computational role of spontaneous activity remains a mystery.

We propose that up-states appear in E/I networks that are optimised for error correction. In recent work, we showed that E/I balance can be understood as a form of optimal predictive coding: by maintaining a tight balance, a population of neurons monitors its own coding error [4]. We find that when such networks are inactive, noise can induce neural spiking. The error correcting mechanisms of the network then respond to these noise-induced spikes by producing additional spikes in an attempt to correct the error. This error-correcting response appears as up-state of spiking activity. The network remains in an up-state, propagating the error signal throughout the network, until it manages to resolve the noise-induced error with an appropriate spike-train. Once the error has been resolved, the network activity returns to a down-state. According to this computational view, up-states are an attempt to correct errors induced by noise. Finally, we investigate the properties of these up-states, such as the distribution of the duration of up-states and the distribution of time epochs between up-states (inter-burst intervals) and we explore the E/I network parameters that determine the shape of this distribution.

[1] A. Luczak, P. Bartho , K.D. Harris, J.Neurosci., 33(4), (2013).
[2] A. Destexhe, J. Comput Neurosci. 27(3):493-506, (2009).
[3] C. van Vreeswijk, H. Sompolinsky, Science. 274(5293):1724-1726, (1996).
[4] M. Boerlin, C.K. Machens, S. Denève, PLoS Comput Biol. 9(11), (2013).