Transient recurrent dynamics shape representations in mice

In the brain, different stimuli are represented by responses of a neural population. How is this representation reshaped over time by the dynamics of local recurrent circuits? We investigate this question in Neuropixels recordings of awake behaving mice and recurrent neural network models. We derive a mean-field theory that reduces the dynamics of complex networks to only three relevant dynamical quantities: the mean population activity and two overlaps that reflect the variability of responses within and across stimulus classes, respectively. This theory enables us to quantitatively explain experimental observables and reveals how the three quantities shape the separability of stimulus representations through a dynamic interplay. We measure the information transmitted from multiple stimuli to the responses with an optimally trained readout on the population signal. This reveals a trade-off between more information conveyed with an increasing number of stimuli, and stimuli becoming less separable due to their larger overlap in the finite-dimensional neuronal space. We find that the experimentally observed small population activity lies within a regime where information increases with the number of stimuli, sharply separated from a second regime in which information asymptotically converges to zero — revealing a crucial advantage of sparse coding.