Connectomes inform function: from time-varying dynamics to animal behaviour.

Structure guides computation in biological and artificial neural networks. However, the nature of the relationship between structure and function, in this context, is unclear. For example, there is still debate on whether constraining a network with biological detail confers a non-trivial functional advantage over a network without such constraints. To shine light on this topic, we highlight five experiments which employ biological constraints onto artificial neural networks using empirically-guided wiring diagrams, or connectomes, from an adult fruit fly, a larval zebrafish, and from the Mammalian MRI (MaMI) dataset, and impose these onto reservoir-based recurrent neural networks, while studying changes in performance and prediction dynamics on synthetic and naturalistic time series data. We observe that fly-constrained networks are better at making predictions from chaotic input data, and in executing multiple mutually exclusive tasks simultaneously, all with a robustness to hyperparameter variations, some of which may lead to chaos. Separately, we find that the global clustering coefficient of the fly network improves performance and variance on time-varying predictions. We also report that an empirical functional connectome from the optomotor response circuitry of a larval zebrafish validates its own behaviour, and that this is interrupted by rewiring. Finally, using the MaMI dataset, we determine that rewiring degrades multifunctional capacity, and that more multifunctional networks have a higher mean degree centrality. Collectively, these findings suggest that biological topology constraints confer distinct advantages to arbitrarily-weighted networks.