Implications of single-neuron gain scaling for information transmission in networks

Abstract

AbstractSummary:Many neural systems are equipped with mechanisms to efficiently encode sensory information. To represent natural stimuli with time-varying statistical properties, neural systems should adjust their gain to the inputs' statistical distribution. Such matching of dynamic range to input statistics has been shown to maximize the information transmitted by the output spike trains (Brenner et al., 2000, Fairhall et al., 2001). Gain scaling has not only been observed as a system response property, but also in single neurons in developing somatosensory cortex stimulated with currents of different amplitude (Mease et al., 2010). While gain scaling holds for cortical neurons at the end of the first post-natal week, at birth these neurons lack this property. The observed improvement in gain scaling coincides with the disappearance of spontaneous waves of activity in cortex (Conheim et al., 2010).We studied how single-neuron gain scaling affects the dynamics of signal transmission in networks, using the developing cortex as a model. In a one-layer feedforward network, we showed that the absence of gain control made the network relatively insensitive to uncorrelated local input fluctuations. As a result, these neurons selectively and synchronously responded to large slowly-varying correlated input--the slow build up of synaptic noise generated in pacemaker circuits which most likely triggers waves. Neurons in gain scaling networks were more sensitive to the small-scale input fluctuations, and responded asynchronously to the slow envelope. Thus, gain scaling both increases information in individual neurons about private inputs and allows the population average to encode the slow fluctuations in the input. Paradoxically, the synchronous firing that corresponds to wave propagation is associated with low information transfer. We therefore suggest that the emergence of gain scaling may help the system to increase information transmission on multiple timescales as sensory stimuli become important later in development. Methods: Networks with one and two layers consisting of hundreds of model neurons were constructed. The ability of single neurons to gain scale was controlled by changing the ratio of sodium to potassium conductances in Hodgkin-Huxley neurons (Mainen et al., 1995). The response of single layer networks was studied with ramp-like stimuli with slopes that varied over several hundreds of milliseconds. Fast fluctuations were superimposed on this slowly-varying mean. Then the response to these networks was tested with continuous stimuli. Gain scaling networks captured the slow fluctuations in the inputs, while non-scaling networks simply thresholded the input. Quantifying information transmission confirmed that gain scaling neurons transmit more information about the stimulus. With the two-layer networks we simulated a cortical network where waves could spontaneously emerge, propagate and degrade, based on the gain scaling properties of the neurons in the network.