NEURAL SEQUENCES ARE GENERATED BY A RESTRICTED SET OF CONNECTIVITY MATRICES

Neural sequences, characterized by neurons or groups of neurons that fire one after the other, have been observed in many brain regions and are known to underlie multiple brain functions. To flexibly support computations, sequences show much variability in features like their temporal width and firing rate. Despite this variability and the prominent role of sequences in circuit computation, a framework that explains how a diversity of sequences can be generated and dynamically maintained in a recurrent network is still missing. To fill this gap, we built recurrent neural network models and trained them to generate sequences with different features. We found that for all sequences the connectivity profile showed local excitation and global inhibition, yet its properties depended smoothly on features of the sequences. For wider sequences the connections were weaker, and for sequences with higher baseline level of firing rate the connectivity profile became more asymmetric. Moreover, beyond a critical value of maximum firing rate, oscillations appeared in the connectivity profile. The model predictions were consistent with connectivity inferred from experimental data. Our results therefore suggest that small changes in the connectivity matrix can generate a wide range of neural sequences. We next identified the conditions under which the sequences remain stable and, alternatively, they undergo discontinuities or reset. Finally, we investigated how different features of the sequences could constrain their functional role. Altogether, our findings provide a framework for how network connectivity might support a wide range of sequences across neural circuits and computations.