A common structure in recurrent networks supports neural sequence generation locally and in downstream neurons

Neural sequences, sequences of firing neurons, have been observed across various brain regions in multiple species. These sequences play a crucial role in behavior and cognition, displaying variability in factors like duration, baseline activity, and peak firing rate. However, a comprehensive framework explaining how these flexible sequences are generated and maintained within a neural circuit remains elusive.

In this study, researchers transcended traditional models that focused on the one-to-one connection between network connectivity and specific sequential patterns. Instead, they trained recurrent neural network models to generate a range of sequential dynamics, subsequently analyzing the corresponding connectivity matrices. The findings revealed that despite different connectivity matrices generating the same neural sequence, the matrices sharing a specific sequence exhibit a common connectivity profile.

This profile refers to the average weight between neuron pairs, considering their position in the sequence order.

The researchers argue that the connectivity profile, rather than the connectivity matrix itself, acts as a unique fingerprint of the sequential dynamics. It shapes how the network responds to alterations in neural activity. Their model predictions aligned with experimental data gathered from diverse brain regions across various species, providing further support for their conclusions.

Moreover, the study demonstrated that neural sequences possess the ability to both enable and limit the generation of an extensive array of sequences in downstream brain regions. This capacity suggests that neural sequences could function as scaffolds, facilitating a multitude of computations across different brain circuits. Ultimately, the researchers' results illuminate the role of network connectivity in generating a diverse set of neural sequences across circuits and how these sequences can be adaptively modified.

By identifying sequences as a universal algorithm supporting brain function across various regions and species, the study contributes to a deeper understanding of the common structural principles in recurrent networks that underpin neural sequence generation locally and in downstream neurons.