Up-Down-Like Background Spiking Can Enhance Neural Information Transmission

More information about the signal can be transmitted by the readout population activity a(t) if background activity undergoes transitions between up and down states. A, B, Firing rate of the background population and raster-plot of the spikes received by an arbitrary readout neuron. C, D, Signal realization (frozen noise). E, F, Normalized population activity and spike raster of the readout population. For the population activity, two trials with the same signal (C, D) and, in the UD case, two-state process (B) are shown. G, Coherence between signal and activity of the readout population. H, Lower bound to the mutual information rate and , the mean firing rate of the readout population, as a function of the mean background rate . The dashed line marks the value used in A–G. Parameters, if not indicated otherwise: Hz, ms, ms.

Effects of up/down duration and temporal structure of the signal. A, Coherence for AI background (blue) and UD background at three different values for the mean duration . Here, Hz. B, Mutual information rates as a function of the background rate for the cases shown in A. The inset shows the theoretical curves over a wider range; the gray line marks what could optimally be reached by infinitely slow switching. C, D, Signal power spectra and coherence for a lower cutoff f₀ of 0 Hz and 25 Hz, respectively. E, Mutual information rates as a function of f₀. The dashed line marks cutoff frequency used in A–C. In A–E, ms, ms, Hz.

The beneficial effect of a UD background persists in a more realistic setup. A, Sketch of the modified setup. B, Mutual information rate for different mean background rates . The dashed line marks the value used in C–F. C, Sample traces of the readout population activity for irregular (gray, ) and rather regular (red, ) UD switching. D, Histogram for the duration of up states in the readout population. To obtain this, population activity is binned (width ΔT = 10 ms); n consecutive bins in each of which at least one readout neuron spikes then define a readout up state of length T_U = nΔT. E, Effect of changing the regularity of UD switching on the mutual information rate. F, Effect of changing the recurrent synaptic weights. The dotted line marks the value used in B–E. In all panels, ms.

A simple model for the effect of the signal on the UD switching We assume that the transition rates between UD states in the background are modulated by the signal. Shown is a realization of the signal and the corresponding background rate for three values of the modulation strength , as well as the mutual information rate over the mean background rate (obtained in simulations) for these . A, Parameters as in Figure 2. B, A much slower signal with f_C = 1 Hz. Note the different y-axis scaling in the mutual information plots.

Traveling up states allow continuous signal transmission A, Sketch of the setup. B, Activity and spike raster of the readout population (neurons ordered by the position of their background population). C, Mutual information rates for different background rates. Red solid and dashed lines correspond to the theoretical limits and , respectively; blue line is theory for AI background. The black dashed line marks the background rate used in the other panels. D, Mutual information rates for varying wave speed c. The dotted line marks the wave speed used in B, C. E, Silence density, i.e., fraction of time bins (length ΔT = 4 ms) in which the population activity is zero. Where nothing else is indicated, the wave speed is c = 10 mm/s and the mean background rate is Hz. In C–E, symbols represent simulation results, while lines are theory.