Decodable Spectral Structure of Input Noise in a Layered Cortical Spiking Network

Individual spikes are difficult to track through recurrent cortical circuits, where convergence, divergence, and nonlinear dynamics rapidly transform activity. However, the statistical properties of an input may remain detectable even when individual spike identity does not. We asked whether the spatiotemporal spectral slope of colored noise remained detectable as activity propagated through a simulated cortical network.

We built a biologically inspired, four-layer recurrent spiking network based on mouse primary visual cortex. A drifting grating was embedded in colored noise spanning five spectral conditions (violet, blue, white, pink, and brown), while noise contrast and stimulus drive were varied independently. The network operated in an asynchronous-irregular regime, and population spectra were characterized using their aperiodic component.

Population activity changed systematically with the spectral slope of the noise rather than separating into discrete color representations. A continuous decoder recovered the spectral exponent β from population activity. Decoding generalized to new noise images the decoder had never seen. Removing all synaptic connections left L4 decoding unchanged, showing that the spectral signal there did not require recurrent interaction; recurrence instead determined how much of it reached the other layers. A feedforward model that received the same stimulus but lacked cellular dynamics did not support decoding. Decoding was highest in L4, whereas the univariate effect of β on firing rate peaked one stage downstream, in L2/3, and increasing noise contrast improved spectral discriminability. Recruiting feedforward inhibition in L4 did not change decoding accuracy.

Recurrent cortical dynamics can transform input activity while retaining information about its statistical structure. This information was distributed across multiple dimensions of population activity and remained recoverable across cortical layers. Tracking population statistics rather than individual spikes provides a tractable framework for studying how recurrent circuits transform and retain input activity.