Structured neural fluctuations can generate noise invariance and inter-areal gating at distinct timescales

The brain processes, computes, and categorizes sensory input. But even in sensory brain areas, the relationship between input signals and neuronal spiking activity is complex and non-linear. Fast subsecond fluctuations in neuronal population responses dominate the temporal dynamics of neural circuits. Traditional approaches have treated this activity as "noise" that can be averaged away by taking the mean spiking rate over wide time bins or over multiple trial repetitions, but this ignores much of the temporal dynamics that naturally occur in neural systems. We find that subsecond flares of increased population activity are layer- and cell-type specific, and large-scale computational modelling suggests they may serve as an inter-areal gating mechanism. Moreover, we find that most of the neural variability is restricted to a population-gain axis. This observation explains why neural systems can function in the presence of excessive variability: population-level spiking dynamics generate invariance to the majority of neural noise.