Clustering and degree homogeneity govern reservoir computing performance

Reservoir computing exploits fixed recurrent networks as computational substrates, and the choice of network architecture shapes performance. Because biological brains are sparse, clustered, and small-world networked, connectome-derived reservoirs are often expected to be computationally favorable. However, its computational capability has not yet been fully explored. Here we examine this expectation by benchmarking reservoirs constructed from a human structural connectome against Watts–Strogatz‘s small-world networks and random networks, matched for size and mean degree. The benchmark tests are two standard tasks: human-activity classification and Mackey–Glass chaotic signal prediction. On both tasks the small-world reservoir attains the highest performance, while the connectome performs comparably to a random network on prediction. Sweeping the small-world rewiring probability, we find that performance on both tasks is associated with high clustering and low degree heterogeneity and is largely insensitive to characteristic path length, with an optimum in the small-world regime. The connectome, characterized by relatively low clustering and high degree heterogeneity, sits away from this optimum, which is occupied by the small-world networks. These findings identify network properties associated with reservoir performance and indicate that connectome-based connectivity is not, in itself, advantageous for the tasks examined.